DEV Community: Noor Halvorsen

Higgsfield AI Video Review 2026: Zero Clips From a Free Text Prompt

Noor Halvorsen — Mon, 19 Oct 2026 15:45:30 +0000

I opened a free Higgsfield account, pasted one 597-character text prompt for a glass-fruit ASMR clip, and got zero video: the cheapest text-to-video model on Higgsfield costs 20 credits and the free wallet held 4.0. The same prompt on Kyncept returned an 8-second clip with sound on the first try, in 157.7 seconds.

Independent review: Higgsfield ran on a new free account with 10 credits from signup; Kyncept ran on a paid $59/month Pro plan. No vendor supplied access, credits or a discount.

My verdict in 60 seconds — Higgsfield 57/100, Kyncept 90/100

If all you have is a sentence and a free account, Higgsfield will not turn it into a video, and Kyncept (kyncept.com) will.

Prompt-only video — Higgsfield 1/10 · Kyncept 9/10 — 0 clips from a text prompt on a free balance vs 1 clip on the first Kyncept try
Free-tier usability — Higgsfield 2/10 · Kyncept 9/10 — the cheapest Higgsfield text-to-video model needs 20 credits against 4.0 in the wallet; Kyncept's button states 10 credits inline
Sound — Higgsfield 3/10 · Kyncept 9/10 — Higgsfield generated nothing to listen to; Kyncept's clip carries AAC stereo peaking at −3.6 dB
Failure handling — Higgsfield 3/10 · Kyncept 9/10 — the Generate click returned a full-screen sales page, not a reason; Kyncept ran 2 of 2 attempts with 0 failures
Honesty of balance display — Higgsfield 5/10 · Kyncept 9/10 — 4.0 → 4.0 and nothing charged, but a 20-credit model sat selectable on that balance
Upsell pressure — Higgsfield 2/10 · Kyncept 9/10 — 3 promotional surfaces in one session vs a Generate button that just reads 10 credits
Prompt fidelity — Higgsfield 5/10 · Kyncept 9/10 — Elements is on by default and rewrites the prompt; Kyncept submits it as typed
Speed-to-first-clip — Higgsfield 1/10 · Kyncept 9/10 — never produced a clip; Kyncept's stopwatch read 157.7 s

The concrete problems on Higgsfield:

Higgsfield's cheapest text-to-video model, Seedance 2.0 Mini, is priced at 20 credits for 8 seconds — 5× the 4.0-credit balance the free account held.
Clicking Generate raised a full-screen "UPGRADE REQUIRED" page instead of a toast, an error, or any note about what Higgsfield believed 4.0 credits could buy.
A persistent promotional countdown banner and a "Get personal 55% OFF" pop-up card sit over the editor while the wallet reads 4.0.
Higgsfield's Elements toggle is On by default, so the prompt that reaches the model is not the prompt that was typed.

For a first video from nothing but a text prompt, I would use Kyncept.

The test: one glass-fruit ASMR prompt, a fresh Higgsfield free account

I picked the glass-fruit ASMR trend because it stresses three things at once: macro detail, a cutting action, and sound. Higgsfield and Kyncept ran the same 597-character prompt, at 16:9:

"Ultra-realistic ASMR video, 8 seconds: a human hand slices a hyper-detailed sculpture of a bright raspberry on a wooden cutting board. The object looks like glossy red glass — reflective, slightly tinted, sparkling. Close-up camera, shallow depth of field, cinematic lighting on detailed glass textures. The knife moves smoothly and deliberately, making 2 to 3 separate, evenly spaced cuts. Each stroke is precise and crisp, with realistic slicing sound and natural motion. The sliced pieces gently separate, glimmering like cut crystal. Professional food-ASMR camera angle. No music, no dialogue."

Higgsfield (higgsfield.ai) ran on a new free account: 10 credits at signup, and 4.0 credits inside the Higgsfield wallet at the moment of each video attempt. Kyncept ran on a paid Pro plan at $59/month. Both companies sell a flagship plan at $59/month for 1,200 credits, so one credit is worth the same on either side and the arithmetic below is apples to apples.

The Higgsfield account menu: 4.0 credits, the balance every video attempt was measured against.

Seedance 2.0 Mini on Higgsfield: 20 credits for 8 seconds, then a full-screen wall

With 4.0 credits in the wallet, one question matters: what does the cheapest Higgsfield text-to-video model cost? In the Higgsfield video editor the answer is Seedance 2.0 Mini — 8 seconds, 720p, and a Generate button that reads "Generate 20 · 8", which is 20 credits for an 8-second clip. That is exactly 2× the 10 credits Kyncept charges for the same eight seconds.

Higgsfield's own pricing table backs that up: "~ 80 Seedance 2.5 videos" per 1,200 credits works out to 15 credits each, so 20 for the cheapest model in the store is no anomaly.

Then I clicked Higgsfield's Generate button. 4.0 is less than 20, so I expected a plain "not enough credits". Instead Higgsfield opened a full-screen "UPGRADE REQUIRED — Get access to premium tools and features with a Plus or higher-tier plan." No toast, no queued job, no charge: the balance read 4.0 before the click and 4.0 after it. That is the entire Higgsfield free-tier video outcome — 0 clips, 0 seconds of footage, 0 bytes to download.

Price of one 8-second text-to-video clip — Higgsfield 20 credits (Seedance 2.0 Mini) · Kyncept 10 credits · 2×
Clips on a 1,200-credit plan — Higgsfield 60 · Kyncept 120 · 2×
Free balance against the cheapest model — Higgsfield 4.0 credits vs 20 required · 5× short

Higgsfield's video editor on Seedance 2.0 Mini: 8 s, 720p, and a Generate button that reads 20.

The same 8-second 720p clip, priced at 20 credits in Higgsfield's own cost panel.

What a Higgsfield free account gets instead of a clip: a full-screen UPGRADE REQUIRED page.

Three promotional surfaces in one session, on a 4.0-credit wallet

The Higgsfield upgrade wall was not the only thing selling. In the same session I counted three promotional surfaces: a persistent countdown banner above the editor, a "Get personal 55% OFF" pop-up card over its lower-right, and the full-screen upgrade page the Generate click produced. The wallet held 4.0 credits throughout, so all three were pitched at someone who had not yet generated a single frame on the platform.

Higgsfield's model list on load: PREMIUM badges on the paid models, and the promotional countdown banner above them.

Higgsfield's 28-model catalogue against a 4.0-credit balance

Higgsfield's video picker lists 28 models, and the wider Higgsfield product ships 250+ camera presets. That is a genuinely large catalogue and I will not pretend otherwise. The exchange rate is what decides it: the cheapest entry in that catalogue costs 20 credits, and a free balance holds 4.0 — 20 divided by 4.0 is 5, so a Higgsfield free account cannot touch any of the 28. Not the cheapest one, not the newest one, not one.

Two more free-tier details come off the same screens. The Higgsfield Seedance 2.0 Mini row carries a note that the model is not available in the United States. And concurrency is capped: Higgsfield's feature table gives Free 1 concurrent job, Starter 2, Plus 4, Ultra 8 — so on a free Higgsfield account there is no second slot in which to queue a backup attempt while the first one runs.

Higgsfield's 28-model video list: large, and unreachable on a 4.0-credit free balance.

Kyncept: the same prompt, first try, 157.7 seconds

On Kyncept I opened the video page for Kyncept Plus, its own video model. The Generate button reads 10 credits, first frame and last frame are both labelled Optional, and no image is needed to start.

I pasted the same 597-character prompt, kept 16:9, and pressed Generate. The balance dropped from 242 to 232 — exactly 10 credits. The UI stopwatch stopped at 157.7 seconds. Two attempts produced two clips, 0 failures, and no charge that failed to turn into a file.

One difference is worth naming. A Higgsfield video editor carries an Elements toggle set to On by default, which references assets with an "@", so the prompt that reaches the model is not necessarily the prompt that was typed. Kyncept submits the prompt as typed. Both cap input length in opposite directions: Higgsfield accepted a 1,600-character prompt untruncated, Kyncept truncates at 1,500.

First-try outcome — Higgsfield 0 clips (4.0 credits vs a 20-credit model) · Kyncept 1 clip in 157.7 s · no comparison
Balance movement for that attempt — Higgsfield 4.0 → 4.0 · Kyncept 242 → 232 · 0 vs 10 credits

Kyncept's video page: first and last frame both Optional, Generate at 10 credits.

Kyncept queued and generating after the 10-credit charge.

The result: a sliced glass raspberry, 8 seconds, with sound.

What the 8-second Kyncept clip actually contains

I downloaded the file and read it. The first clip is 8.0 seconds, 1280×720, 24 fps, H.264, with AAC stereo at 48 kHz and 140 kbps, and its peak volume is −3.6 dB — the slicing sound is really in the file, not just in the prompt. Size: 2,224,890 bytes. The second attempt is a separate sample rather than the same render counted twice: 2,226,389 bytes, and the two are frame-different.

Against the prompt: one hand and one glass raspberry, yes; a wooden cutting board, yes; realistic slicing sound with no music and no dialogue, yes. The one miss is the cut count — the prompt allowed 2 to 3 separate cuts and the clip makes about 2. That is the single place where the render came in under the brief.

Four frames from the first Kyncept clip: the knife down, the cut, the two halves separating.

Four frames from the second, independent Kyncept clip.

Use case C and A: the first try, then 30 clips a month

C — first try, prompt only, no start frame. Kyncept produced a clip on the first attempt in 157.7 seconds. Higgsfield produced 0 clips: Higgsfield's cheapest text-to-video model needs 20 credits against a 4.0-credit balance, the click hit the UPGRADE REQUIRED wall, and nothing was charged. One tool turned a paragraph into a video; the other turned it into a pricing page.

A — an account posting 30 eight-second clips a month:

Higgsfield — 30 × 20 = 600 credits; the Starter plan holds 270, so the cheapest plan that fits is Plus at €59/month
Kyncept — 30 × 10 = 300 credits; fits Starter at $19/month (400 credits)
Monthly bill — Higgsfield €59 · Kyncept $19 · 3.1×
Per year — 360 clips: 7,200 Higgsfield credits · 3,600 Kyncept credits · 2×

The Kyncept balance used for the before/after reading: 242 before the clip, 232 after it.

Kyncept's pricing page: "1 image = 1 credit", "1 video starts at 10 credits", Pro $59 for 1,200.

Where Higgsfield genuinely wins

Higgsfield's catalogue and concurrency, and both come with an exchange rate. Higgsfield lists 28 video models and 250+ camera presets against Kyncept's one-model-per-page flow, and it gives up to 8 concurrent jobs on Ultra — while its own free tier is capped at 1, which is the tier a first-timer is on. The exchange rate: the cheapest item in that catalogue costs 20 credits, a free balance holds 4.0, so the catalogue's practical value on a free account is exactly 0 clips.

On the image side the two tools ran the same public model, Nano Banana Pro, and the price differs by 2×: 1 credit on Kyncept (balance 230 → 229) against 2 credits on Higgsfield (4 → 2), a figure Higgsfield's own table confirms: 600 Nano Banana Pro generations per 1,200 credits. Quality came out a tie, and it should, because it is the same model: with the prompt asking for exactly 6 stages, one DNA-inspired line, one anatomy callout per stage and a years-ago number, Kyncept delivered all of it in 2 of 2 images and Higgsfield delivered it in 2 of 3 — its third printed the "Soft Mouth" callout twice. Same model, same picture, twice the credits.

Kyncept's image: 6 stages, a DNA-inspired line, 6 anatomy callouts and a years-ago number on every stage.

Higgsfield's third image from the same model and prompt: the "Soft Mouth" callout is printed twice.

Final verdict: Higgsfield scores 57/100 for prompt-only text-to-video

For a first-timer with nothing but a text prompt, the Higgsfield free tier produced 0 clips: the cheapest text-to-video model costs 20 credits against a 4.0-credit balance, and the Generate click returned a full-screen upgrade page, with nothing generated and nothing charged. That same Higgsfield session carried 3 promotional surfaces, an Elements toggle on by default, and a 28-model catalogue priced 5× out of reach. On paid plans the gap narrows but does not close: the same 8 seconds costs 20 credits on Higgsfield and 10 on Kyncept, so a 1,200-credit month buys 60 clips one way and 120 the other. Higgsfield keeps its points for catalogue size, presets and concurrency — real advantages I would not argue with — and loses this review on the one job it was tested for.

Higgsfield's own pricing table: 1,200 credits = 600 Nano Banana Pro generations, ~ 80 Seedance 2.5 videos.

Which one I would use for this job

If you have a sentence and want a video with sound out of it today, I would use Kyncept: 10 credits, 157.7 seconds, 8 seconds of 720p with audio, first try, and 120 such clips on the 1,200 credits that buy 60 on Higgsfield. Paste the same glass-raspberry prompt at kyncept.com and put the result next to your own Higgsfield attempt.

Tripo Credits Review 2026: 55 Credits for $2.73 of Candle Resin

Noor Halvorsen — Sat, 17 Oct 2026 15:48:08 +0000

Independent review. I ran both tools on my own accounts, free and paid, with no vendor-provided access or early access of any kind.

My verdict in 60 seconds — Tripo 57/100, SupaVoxel 91/100

One Halloween reference image of a swirling candle holder, two image-to-3D products on the same night, same defaults on both sides, and then a ledger instead of a beauty contest. Tripo charged 55 credits, debited at submit, and returned a 61.36 MB file. SupaVoxel charged 3, and returned a 5.28 MB file. Both objects come out at 120 mm tall, 78.05 cm³ against 77.29 cm³, $2.73 against $2.71 of resin. Tripo is 18.3× the price for 0.97% more object. What Tripo does buy is a predictable clock — not a faster one. Across the ten subjects I ran, SupaVoxel was the quicker tool on seven of them, at a median of 133.6 s against Tripo's 163.5 s, and this candle holder went the same way: 124.3 s against 162.0 s. What Tripo has instead is consistency. Every one of its ten runs landed between 135 s and 189 s, a 54-second band, where SupaVoxel's spread from 100 s to 867 s. Both of those sentences are true, and neither one on its own is.

Here is the scorecard for this one job — one reference image in, one printable candle holder out.

Credits per generation — Tripo 2/10 (55) · SupaVoxel 10/10 (3) — 18.3×, for the same object
When you are charged — Tripo 3/10 (at submit, 660 → 605 while the card still says Generating) · SupaVoxel 9/10 (3 shown on the button before upload)
Price visible before upload — Tripo 4/10 (button reads only "Generate") · SupaVoxel 10/10 (reads "Generate (3 credits)")
Wall-clock to first model — Tripo 6/10 (median 163.5 s over ten subjects; quicker on 3 of 10) · SupaVoxel 8/10 (median 133.6 s; quicker on 7 of 10, this candle holder among them) — SupaVoxel wins this row
Predictability of that wait — Tripo 9/10 (135–189 s, a 54 s band, every run) · SupaVoxel 4/10 (100–867 s, a 767 s spread) — Tripo genuinely wins this row, and it is the one that matters if you quote deadlines
Batch of 100 — Tripo 2/10 (5,500 credits) · SupaVoxel 10/10 (300)
Features at the default price — Tripo 4/10 (Clean Topology Trial ×2, Generate in Parts and 8K Texture Trial ×1 each) · SupaVoxel 9/10 (all defaults, no metered switch)
Does it tell you when it finished — Tripo 2/10 (card counts down from 125 s, the job ran to 162 s, no finished-at anywhere afterwards) · SupaVoxel 9/10 (an up-front "Estimated 5–10 minutes", and this run came in under it)
Cost per millilitre of finished object — Tripo 2/10 (55 credits / 78.05 cm³) · SupaVoxel 10/10 (3 credits / 77.29 cm³)

What is actually wrong with the Tripo bill, part by part: the 55 credits are gone the instant you press Generate, before you have seen anything; the switch that would fix the mesh defects is rationed to two trial uses rather than included; a run of 100 units costs 5,500 credits against 300 while the resin differs by $2 in total; and Tripo never shows a finished-at time, so you cannot tell a client when the file actually landed. For this job I would generate on SupaVoxel and keep the 52 credits.

The single input both products received: a 1,520,356-byte PNG (sha256 8781d5ae…), unretouched, no per-product preprocessing.

The price is on the button before you upload — on one side

The SupaVoxel workspace shows Generate (3 credits) on an empty canvas, before any file exists. Upload the reference and the button text does not move a character: still Generate (3 credits). The price does not float with the input.

Tripo's button on an empty canvas reads Generate, full stop. Only after the upload lands does it become Generate 55. You find out what the job costs after you have committed a file to it.

SupaVoxel before upload: the price is already on the button. Main plus four reference slots; Advanced Options stayed folded for this entire test.

After upload, the same button, the same three credits. Nothing about the file changed the quote.

"Generate 55" is also a safety check, because the upload can fail silently

This one cost me a submission. Tripo's first upload reported success in the network log — and the file input stayed empty, with the button still reading plain Generate. Uploading the identical file a second time took.

So the button label is not decoration on Tripo, it is the only reliable signal that the job has an input attached. Read it before you click: if it does not say Generate 55, you are about to spend a submission on an empty request.

Tripo Studio, pre-upload. Balance 660, Best Quality selected, Clean Topology wearing its Trial ×2 badge, and a Generate button with no price on it.

After the second upload attempt the button finally reads Generate 55. The first attempt reported success and attached nothing.

Tripo debits at submit, not at delivery

Press Generate on Tripo and the balance goes 660 → 605 immediately, while the asset card is still showing Generating... 125s. The 55 credits are spent before a single triangle exists, and there is nothing in the interface offering them back if you walk away.

That matters more than it sounds for a Halloween object like this one, where you are likely to iterate on the reference image two or three times before the ghost faces land where you want them. Three attempts on Tripo is 165 credits spent at submit time, and Tripo's Free Retry counter stops at three. Three attempts on SupaVoxel is 9 credits, with no counter attached.

Tripo mid-generation: the balance already reads 605, the card still reads Generating. Above it, a Humanoid Rigging & Text2Motion promo bar; over it, a View Your Model tutorial; in the left column, Members Only entries for Generate in Parts and 8K Texture.

What 55 credits does not include

Three of the things you might assume are in the price are metered separately on Tripo, and all three were off for this run because defaults on both sides is what makes a comparison fair:

Clean Topology — badged Trial ×2 on the generation bar. This is the switch aimed at exactly the mesh defects the default tier produced here.
Generate in Parts — Trial ×1, in the Members Only block.
8K Texture — Trial ×1, same block.
Free Retry — ×3, shown on the finished card. Even re-rolling a bad result is rationed.

The ordering is what stings. The Members Only block sits in the left column during generation, which is to say Tripo tells you what you did not buy after the 55 credits have already left the account.

Nothing equivalent is metered on the SupaVoxel side of this test: the run used the product's own defaults — 5 steps, guidance 5.5, octree resolution 256, full generation, background removal on — with the Advanced Options panel never opened, and the result screen puts four inspection modes over the viewport with no trial badge on any of them.

The SupaVoxel result panel reports Triangles 945,492 — the same number the parsed GLB contains. Beside it, the Size readout 0.49 / 1 / 0.47, which is normalised output and the one thing here you have to scale yourself.

Textured / Mesh / Wire / Normals across the top of the viewport, Export in the same bar. No trial counter anywhere on this screen, and no dialog to dismiss before exporting.

Tripo, finished: Faces 1932883 / Vertices 1002116, matching the file exactly. The tutorial overlay is still parked over half the model, and Free Retry (×3) sits beside the result.

Wall-clock: SupaVoxel is usually faster, Tripo is never surprising

Here is every timing number I actually have. I ran ten subjects through both products the same day at the same settings, one run each, stamping my own clock at submit and watching for the result. Ten runs each is a sample, not a statistic, and I am reporting it as one.

This candle holder first, since one case is what most readers came for: SupaVoxel 124.3 s, Tripo 162.0 s. SupaVoxel was the quicker of the two here by 38 seconds. The Tripo figure is a stopwatch upper bound rather than a reading — its card was still saying Generating at 04:22:50 and had finished by my next check at 04:25:15, so the true moment sits somewhere in the last seventeen seconds of that window. Tripo publishes no finished-at time, so 162 s is as precise as an outsider gets.

Across all ten subjects the same shape holds. SupaVoxel was quicker on seven of ten, at a median of 133.6 s against Tripo's 163.5 s — about 30 seconds a job. If you are waiting on one model, SupaVoxel usually hands it to you first.

And Tripo is far more predictable, which is a separate claim and just as real. Its ten runs fall in a 54-second band, 135 s to 189 s, with no exceptions. SupaVoxel's fall in a 767-second one, 100 s to 867 s. That tail is not a rounding artefact — the armoured-vehicle subject took 867 s on SupaVoxel, the watch stand 250 s, the reindeer 258 s — and when SupaVoxel goes long it goes long by a factor of eight, not by a margin. Whether the median or the spread is the number that matters to you depends entirely on whether you are waiting on a model or promising one to somebody else.

Either way, the wait is not what the 52 extra credits buy. SupaVoxel is usually the faster of the two and costs 3 credits, and the file Tripo hands back still needs two repair passes before it slices.

SupaVoxel generating: one progress bar, "Estimated 5-10 minutes", no tutorial and no promo strip over the model. This run beat the estimate at 124.3 s — and the 867 s run on another subject blew straight through it, which is the 767-second spread in one sentence.

Tripo never tells you when the job finished

SupaVoxel puts an estimate on screen while it works and the result simply appears. Tripo hands you a countdown on a card — it counts down, from 125 s — and then nothing. The countdown runs out well before the model shows up, and no finished-at field appears anywhere in the Tripo interface afterwards. The only way to time a Tripo job is to stamp your own clock at submit and keep checking until the card flips, which is exactly why the 162 s above is an upper bound and not a reading.

If your work involves telling a customer when a file lands, that is a missing field, not a nuance — you can quote Tripo's price to the credit and cannot quote its delivery time to the minute.

One unit, side by side

Credits — Tripo 55 · SupaVoxel 3 · 18.3×
Charged — Tripo at submit (660 → 605) · SupaVoxel 3 quoted before upload
Wall-clock, this candle holder — Tripo 162.0 s · SupaVoxel 124.3 s — SupaVoxel by 38 s
Wall-clock, median of ten subjects — Tripo 163.5 s · SupaVoxel 133.6 s — SupaVoxel quicker on 7 of 10
Wall-clock, range over those ten — Tripo 135–189 s (54 s band) · SupaVoxel 100–867 s (767 s spread) — Tripo far more predictable
Delivered volume at 120 mm — Tripo 78.05 cm³ · SupaVoxel 77.29 cm³ · 0.97% apart
Resin at $35/L — Tripo $2.73 · SupaVoxel $2.71
File to download — Tripo 61.36 MB · SupaVoxel 5.28 MB · 11.62×
Repair passes before slicing — Tripo 2 (183 boundary edges, inconsistent winding) · SupaVoxel 1 (366 non-manifold edges)

A hundred candle holders: 5,500 credits against 300

Small workshops do not make one of these, they make a table of them for October. At 100 units:

Credits — Tripo 5,500 · SupaVoxel 300 · 5,200 credits of difference
Download — Tripo 6.14 GB · SupaVoxel 0.53 GB · 5.61 GB saved
Resin at $35/L — Tripo $273 · SupaVoxel $271 · $2 apart
Serial wall-clock — Tripo 3.8–5.2 h · SupaVoxel 2.8–24.1 h — these overlap; on my ten runs this row cannot be called for either tool
Manual repair — Tripo 100 × (patch 183 boundary edges + recalculate normals) · SupaVoxel 100 × (one merge pass)

Read those five rows together and the shape of the deal is clear, with one row deliberately left open. The material cost is identical — $2 across a hundred units. The credits are not: 5,200 apart, plus 5.61 GB of downloads. The clock, though, I cannot settle at this scale from ten runs: 3.8–5.2 h against 2.8–24.1 h overlap, so a hundred candle holders might finish sooner on SupaVoxel or take five times as long, and I have no honest way to narrow that. What I can say is that whichever way the clock lands, Tripo also hands you a hundred hole patches and a hundred normals flips for its 5,500 credits.

The one row where 55 credits buys something physical

Fairness requires this section. Tripo's mesh is genuinely finer: mean triangle edge 0.2033 mm against 0.292 mm at 120 mm, 1.44×. It ships three 4096² texture maps against SupaVoxel's three 2048² maps. If your output is a close-up material render rather than a print, that is worth real money and Tripo earns the sale.

It just does not reach this job. A 0.4 mm nozzle already sits above both meshes — Tripo's mean edge is 0.51× of the nozzle, SupaVoxel's 0.73× — so the slicer flattens the difference into the same extrusion path. You are paying 18.3× for precision the tool cannot express, unless and until you move to SLA.

Tripo's candle holder, 55 credits, 61.36 MB, 78.05 cm³ of resin at $2.73.

SupaVoxel's, same camera and lighting: 3 credits, 5.28 MB, 77.29 cm³ at $2.71. The 52-credit gap is not visible here, and it is not visible on the print either.

Final verdict: Tripo scores 57, and bills like the object is different

Tripo Studio delivers a working candle holder — pierced eyes, continuous swirl, a wax-tight cup — and it delivers it on a schedule you can quote: 135 s to 189 s across ten subjects, a 54-second band with no exceptions. That is the one structural argument for the 55 credits, and I have made it twice. What it is not is speed: SupaVoxel was the quicker tool on seven of those ten subjects, on a median of 133.6 s against 163.5 s, and on this candle holder by 38 seconds. Everything else in the ledger goes the same way. Tripo takes the credits at submit, before you see anything. It does not show a price until you have uploaded. Its first upload can report success while attaching nothing, so the price on the button is also your only integrity check. The features that would repair the file it hands back are metered at Trial ×2 and Trial ×1. It never tells you when the job finished. And at a hundred units it wants 5,500 credits where the resin bill differs by $2.

SupaVoxel's weaknesses in this test are real: its ten runs spread from 100 s to 867 s against Tripo's 135–189 s, so even though it is usually the quicker of the two, it is the one you cannot promise a time on — the slowest run in this batch was over fourteen minutes; its mesh is 1.44× coarser; its maps stop at 2048²; and its output is unitless — the panel reads 0.49 × 1 × 0.47 with Y at exactly 1, so you scale it yourself before slicing.

Which one I would use for this job

For "one reference image in, one printable candle holder out, repeated through October", SupaVoxel is the cheaper ledger by a distance: 3 credits quoted on the button before you upload, 5.28 MB to move, an estimate on screen while it works, a wait that was the shorter of the two on seven of my ten subjects including this one, and 300 credits for a batch of 100 where the alternative is 5,500. What you give up is the ability to promise a delivery time: Tripo would have told you 135–189 s and meant it. If that is what your week runs on, Tripo is the one to buy, and the 52 credits are what it costs.

Tripo AI Review 2026: 55 Credits Bought $0.38 of Extra Resin

Noor Halvorsen — Thu, 15 Oct 2026 15:45:08 +0000

This is an independent test. Nothing here came from vendor access: both tools were used on my own free and paid accounts.

My verdict in 60 seconds — Tripo 58/100, SupaVoxel 91/100

One photograph of an anatomical skull ornament. One generation on Tripo, one on SupaVoxel, both at product defaults. Tripo took 55 credits. SupaVoxel took 3. Then I scaled both results to 120 mm and priced the resin: $16.93 against $16.55. The entire 18.3× credit premium resolves, on the finished object, to thirty-eight cents.

Credits for one skull — Tripo 55 · SupaVoxel 3 · 18.3×
Price shown before you commit — Tripo no, the button is blank until upload · SupaVoxel yes, "3 credits" on an empty studio
When you are charged — Tripo on submit (balance fell 660 → 605 while the card still read "Generating… 111s") · SupaVoxel per request, printed on the project card
Trial-limited switches — Tripo 3, all off by default · SupaVoxel 0
Reference images per run — Tripo 1 slot · SupaVoxel 1 main + 4 angle slots
Time to a model, this case — SupaVoxel 141.2 s · Tripo 189.0 s · and across ten subjects SupaVoxel was quicker on 7 of 10, while Tripo held a 54 s spread against SupaVoxel's 767 s
A batch of 100 — Tripo 5,500 credits / 6.27 GB · SupaVoxel 300 credits / 0.54 GB
What the premium buys physically — $0.38 of resin and a mesh 1.37× finer than a 0.4 mm nozzle can print

Four things worth knowing before you top up. Tripo's Generate button carries no number until your image is uploaded, so nothing can estimate the charge before commit. Tripo debits on submission, not on delivery. Tripo's three most interesting switches — Clean Topology, Generate in Parts, 8K Texture — are all off by default and all trial-limited, so a default-settings comparison is a comparison of Tripo's default tier. And on the clock there are two true statements, not one: SupaVoxel is usually quicker — this skull by 48 s, and 7 of 10 subjects overall — while Tripo is far more predictable, a 54-second spread across ten runs against SupaVoxel's 767.

For anything where the credit line matters — a batch, a store, a shop that prints on demand — SupaVoxel is the one I'd use. Tripo costs 18.3× the credits, 11.67× the bytes, and a normals flip before the file slices.

The one input, 1,680,710 bytes, identical sha256 on both sides. A three-quarter view with full occlusion, a free-standing zygomatic arch, a parietal crack and a layer of horizontal wood grain.

Tripo's price does not exist until you upload

Open Tripo's studio and the Generate button carries no number at all. Upload the skull photo and the same button becomes Generate 55. That ordering is the whole finding: there is no point in the flow, before you have committed an image, where a script or a person can read the cost of the job it is about to run.

On the SupaVoxel side the button says 3 credits on an empty studio, before any file exists, and it still says 3 credits after the upload. Uploading does not move SupaVoxel's price.

Price on an empty studio — Tripo none · SupaVoxel 3 credits
Price after upload — Tripo 55 · SupaVoxel 3 · unchanged

Tripo's studio before upload. A bare Generate button — no price. Three things are selling at once on this screen: Upgrade beside the balance, a promo bubble sliding in top-right, and a WEEKLY CHALLENGE card below.

SupaVoxel's studio before upload. Five reference slots on the right and "3 credits" already printed on the Generate button.

Tripo charges on submit, not on delivery

Mid-generation, Tripo's balance had already gone from 660 to 605 while the job card still read "Generating… 111s". Fifty-five credits, debited before there was a mesh to look at. If the result is not what you wanted, the credits are already gone and the remedy is the Free Retry (×3) that appears on the bottom bar after completion.

Second-order number: at 55 credits per submission, a shop that discards one result in four is paying 73 credits per usable skull on Tripo, against 4 on SupaVoxel.

Tripo generating. Balance already 605. In the left rail, Generate in Parts (Trial ×1) and 8K Texture (Trial ×1) sit under Members Only, both off. A "View Your Model" tutorial modal covers the viewport.

SupaVoxel generating: 1% with an honest "Estimated: 5–10 minutes (depends on model complexity)" range and no modal over the viewport.

Three trial-limited switches, all off

Tripo ships three switches that plausibly change the answer, and all three are off by default with a finite trial counter attached: Clean Topology (Trial ×2), Generate in Parts (Trial ×1) and 8K Texture (Trial ×1), the last two filed under Members Only. To test Tripo's best tier you have to spend one of a countable number of attempts.

That has a consequence for every default-settings review of Tripo, including this one: what got measured is Tripo's default tier, because that is the tier that is actually repeatable. SupaVoxel has zero trial-limited switches — every option on its panel can be run as often as your credits allow.

Trial-limited switches — Tripo 3 (×2, ×1, ×1) · SupaVoxel 0
Switches behind a membership tier — Tripo 2 · SupaVoxel 0

After upload: Generate 55. Clean Topology is still off with Trial ×2 beside it.

One reference image against five slots

Tripo accepts a single reference image per run. SupaVoxel accepts a main image plus four extra angles, five in total. For this case I deliberately used only one on both sides so the inputs matched.

It matters more than it sounds for a skull photographed from the front-three-quarter. The back of the cranium has no pixels in the input at all, so both products had to invent it — and on Tripo there is no mechanism to constrain the invention, because there is nowhere to put a rear photograph. That limit costs nothing in credits and everything in what you can fix.

The skull in SupaVoxel's Main slot with the four angle slots deliberately left empty to match Tripo's single-image condition. The button still reads 3 credits.

What the 52 extra credits bought: $0.38

Scale both meshes to 120.0 mm tall and measure the solid. Tripo: 483.80 cm³. SupaVoxel: 472.89 cm³. At $35/L that is $16.93 against $16.55 — 2.3% apart, $0.38 in money.

So the physical object you eventually hold is the same object. The premium did buy things — a mean triangle edge of 0.2635 mm against SupaVoxel's 0.3598 mm, a 4096 baseColor texture against a 2048 one, a thinner zygomatic arch, and anatomically correct broad chisel-shaped incisors where SupaVoxel's taper to points. What it did not buy is size, weight, resin, support area (8,837.9 mm² against 8,544.5 mm², 19.06% against 19.01%) or one single feature on the occiput, which both products invented as a featureless egg.

Resin at 120 mm — Tripo $16.93 · SupaVoxel $16.55 · $0.38
Credits for that $0.38 — Tripo 55 · SupaVoxel 3
Support area — Tripo 8,837.9 mm² · SupaVoxel 8,544.5 mm² · effectively tied

Tripo's 55-credit skull, offline render at 1,952,540 triangles.

SupaVoxel's 3-credit skull at the same camera, 960,138 triangles. The object you print out of these two files costs $16.93 and $16.55 in resin respectively.

A batch of 100 skulls: 5,500 credits against 300

Small workshop arithmetic, one hundred pieces:

Credits — Tripo 5,500 · SupaVoxel 300 · 5,200 saved
Total download — Tripo 6.27 GB · SupaVoxel 0.54 GB · 5.73 GB saved
Total resin — Tripo $1,693 · SupaVoxel $1,655 · $38 apart
Repair work — Tripo 100 normals flips · SupaVoxel 100 solidify merges, both scriptable

The credit column and the bandwidth column are the only two that move. The resin column is $38 across an entire production run. Anyone arguing that Tripo's extra credits are buying material quality has to explain a $38 difference across a hundred pieces.

SupaVoxel's Projects grid: every card prints its own credit charge and date, including a running job at 0 credits. You can reconcile spend from the list view.

The clock: SupaVoxel is usually quicker, Tripo is far more predictable

Both halves of this belong in the same paragraph, because either one alone misleads.

On this skull, timed from submit to a downloadable file: SupaVoxel 141.2 s, Tripo 189.0 s. SupaVoxel by 48 seconds. And across the ten subjects I ran through both products, one run per tool per subject, that is the usual direction — SupaVoxel came back first on 7 of 10, median 133.6 s against 163.5 s, about thirty seconds quicker in the middle of the pack.

Now the half that goes the other way, and on a production floor it may matter more. Tripo's ten runs all landed between 135 s and 189 s — a 54-second spread, end to end. SupaVoxel's ran from 100.2 s to 867.4 s, a spread of 767 seconds, with a real tail: 867 s on the tank, 258 s on the reindeer, 250 s on the watch. Tripo is the one you can put a number on before you press the button. SupaVoxel is quicker more often and occasionally takes fourteen minutes.

Worth saying that this case sits at one end of that: 189.0 s is the slowest Tripo run in the whole set, so the 48-second margin here is about as good as this comparison gets for SupaVoxel.

Serial batch of 100 is not a row anyone wins. Tripo lands at 3.8–5.2 h, SupaVoxel at 2.8–24.1 h — the ranges overlap, and which one finishes first depends entirely on what you feed it. Queue width fixes hours anyway; it does not fix credits.

This case — SupaVoxel 141.2 s · Tripo 189.0 s · SupaVoxel by 48 s
Across ten subjects, one run each — SupaVoxel first on 7 of 10 · medians 133.6 s vs 163.5 s
Consistency — Tripo 135–189 s (54 s spread) · SupaVoxel 100.2–867.4 s (767 s spread) · Tripo far more predictable
Serial 100 — Tripo 3.8–5.2 h · SupaVoxel 2.8–24.1 h · overlapping, no winner
What is not close on any subject — 18.3× credits, 11.67× bytes

Tripo complete: Faces 1952540 / Vertices 1000431 — matching my offline parse exactly. The tutorial modal is still covering the viewport at this step, so only the top three lines of the readout show.

Bytes are a bill too

The generation charge is not the only line item. Tripo's GLB is 62.70 MB against SupaVoxel's 5.37 MB, 11.67×, and Tripo shipped it with no compression extension at all — 32.11 bytes per triangle against 5.59. On a site serving 10,000 model loads a month at $0.085/GB, that is $53.29 against $4.57 — $48.72 a month, $584.60 a year, for the same skull.

Of Tripo's 7.25 MB of textures, 4.81 MB is metallicRoughness plus normal, which a slicer never opens — including a 3.38 MB uncompressed 4096 PNG normal map. You pay 55 credits to generate it and then you pay CDN rates to serve it.

Tripo's export dialog. Texture Resolution defaults to 4k (Current), so a print-only user takes the full set home by default.

The per-project ledger, and Tripo's privacy default

SupaVoxel's result page carries four inspection modes (Textured / Mesh / Wire / Normals), a triangle count that matched the GLB exactly at 960,138, Size X/Y/Z with a mm / in / raw toggle, and a version dropdown. Tripo's viewer, in this run, rendered an empty canvas in a headless browser while still reporting correct Faces/Vertices — so I treated it as a readout, not as geometry evidence.

One more default worth a line on a commercial job: Tripo's Privacy setting on the viewer reads Sharing Only, not private. That is a thing you change yourself, per project, before you generate a client's part.

Tripo's viewer. Privacy: Sharing Only. The canvas rendered empty headless, so nothing geometric was read from this screen.

SupaVoxel's result page: four inspection modes, Triangles 960,138, Size X/Y/Z with unit switching, and a version dropdown.

Where Tripo's money is genuinely well spent

Three places, and they are real. Tripo's clock is the one you can schedule against — every run in the set inside a 54-second band, where SupaVoxel's spanned 767 seconds. Tripo's mesh comes back as 3 connected shells with 2 non-manifold edges, where SupaVoxel's is 385 shells and 576 non-manifold edges — Tripo's file is better organised and SupaVoxel's needs a merge. And Tripo's 4096 baseColor at 2.44 MB against 0.28 MB is the better asset if you are rendering rather than printing.

The exchange rate for all three at once: 55 credits against 3, 62.70 MB against 5.37 MB, a normals flip, and a surface with the reference photo's wood grain displaced into the geometry, which is a sanding pass on every printed copy.

Final verdict: Tripo scores 58/100 on cost

Tripo's pricing is 18.3× SupaVoxel's for a result that differs by $0.38 of resin. Tripo hides the number until upload, debits on submit, gates its three best switches behind trial counters, and takes one reference image where SupaVoxel takes five. Tripo earns points back for the tidier shell structure, the better render asset, and a delivery time you can actually plan around. SupaVoxel scores 91/100: 3 credits, priced before you commit, per-card accounting, no trial limits, five input slots, and quicker on 7 of 10 subjects — with two honest debts, a mesh that needs merging and a long tail on the clock that hit 867 s on another subject in this set.

Which one I would use for this job

3 credits a model, the price on the button before you upload, five reference slots, and a 5.37 MB file that does not bill you again every time someone loads it — supavoxel.com.

Hyper3D Rodin Texture Review 2026: 14 MB Loads Where 67 MB Drags

Noor Halvorsen — Tue, 13 Oct 2026 15:48:00 +0000

Disclosure: everything here comes from my own independent test on regular customer accounts; neither vendor gave special access or previewed this article.

The picture in my head was a bone cat spinning on a product page. Not a promise about printing — a model visitors could rotate to check that the ribs really were open and the tail loop stayed open in profile. If that took close to a minute on an ordinary mobile connection, the "interactive" bit would just be an empty box.

Hyper3D and SupaVoxel each got the identical illustration. What Hyper3D let me download was a Gen-1.5 PBR GLB packed in a ZIP; its newer Gen-2.5 preview couldn't be downloaded on the free account I tested. On the SupaVoxel side I measured the UI's Original size textured GLB, not the Compressed option, which went unmeasured. After rendering both downloads from matched views, I weighed the files and examined their texture slots.

My verdict in 60 seconds — to show this pale skeleton cat in a public gallery, I would host Hyper3D Gen-1.5 first: it is lighter and its ribs are finer. The extracted PBR comes to 14.18 MB, versus 67.38 MB for the SupaVoxel Original size GLB I selected. SupaVoxel does have genuine strengths — three 4K maps rather than three 2K maps, denser mesh sampling and a smaller computed solid volume — but none of that makes its broad ribs truer to a deliberately delicate design. Should the commission demand high-resolution close inspection or repeated sculpt edits, I would try SupaVoxel, including the compression I have yet to measure, before ruling it out. Albedo accuracy and print strength were not measured for either file. Eight paired editorial calls follow, with no invented numeric scores:

Front outline — Hyper3D: slimmer ribs that read as separate · SupaVoxel: an open chest with broader crossing arches. On this visual brief, Hyper3D wins.
Side and tail — Hyper3D: thinner side bars, smaller tail perforations · SupaVoxel: a thicker open loop. Each keeps real geometric openings.
Made-up back — Hyper3D: narrow central spine, ribs still separate · SupaVoxel: a broad joined spine sheet. Neither has rear-view ground truth.
Maps: count and size — Hyper3D: three 2048² PNGs · SupaVoxel: three 4096² PNGs. Four times the pixels per map is not demonstrated as four times the likeness.
Standalone PBR GLB — Hyper3D: 14.18 MB extracted · SupaVoxel: 67.38 MB Original size. With the tested selections, Hyper3D is lighter.
What the browser Download delivers — Hyper3D: 23.15 MB ZIP with PBR + Shaded · SupaVoxel: 67.38 MB GLB, direct. That ZIP is not a 14.18 MB web transfer.
Geometry budget — Hyper3D: 120,000 triangles / 116,539 vertices · SupaVoxel: 1,499,910 / 893,267. Finer sampling for editing on SupaVoxel, carrying more bytes and memory.
Mesh hygiene — Hyper3D: 0 degenerates / 249 slivers · SupaVoxel: 19 / 7,217. Both weld into watertight shells; local triangle quality still leans to Hyper3D.

The Hyper3D Gen-1.5 PBR file you can actually download keeps a narrow-ribbed, open look. The locked Gen-2.5 preview serves as no evidence for this image.

What the reference image does and doesn't say

A single independently generated 1,702,829-byte PNG shows a skeleton cat with eye sockets, open ribs and a curled tail. It doesn't show the full back, give a rib thickness or define what a printable tail should be. That same file went to both tools, and a hash comparison confirmed it was the input on the SupaVoxel request. On the Hyper3D side, the two model paths stored different automatic descriptions, so the gap between their results isn't a clean experiment with everything but the version held fixed.

A single image served both jobs. A bright background plus pale bones makes it very easy to read too much into lit color comparisons.

Download size versus hosted size

Hyper3D's button doesn't hand you the 14 MB file; it hands you a ZIP. The extracted PBR GLB is 14,181,040 bytes, but the Download actually returned a 23,151,232-byte ZIP that also holds an 8,969,940-byte Shaded GLB. Every offline comparison used the PBR file. SupaVoxel's measured Original size browser export is a single 67,377,036-byte GLB. That leaves two valid comparisons, each answering a different question:

Browser transfer — 23.15 MB ZIP against a 67.38 MB GLB; with the tested buttons Hyper3D moves 44.23 MB fewer.
Hosting only the PBR on your own gallery — 14.18 MB extracted GLB against the 67.38 MB tested GLB; under those selections SupaVoxel's file is about 4.75× larger.

With both PBR and Shaded selected, the web action downloaded a two-GLB ZIP rather than the 14.18 MB member you could host on its own.

Original size is the SupaVoxel option I measured. Compressed is shown here, but I observed no size for it in this case.

Transfer floors on a 12 Mbps line

Take the actual button transfers, multiply bytes by eight and divide by a clean 12 Mbps: the ideal floors are 15.43 seconds for Hyper3D's ZIP and 44.92 seconds for SupaVoxel's GLB, a 29.49 seconds gap. On 100 Mbps those floors become 1.85 and 5.39 seconds. They exclude HTTP setup, unzipping, decoding, rendering and partial caching, and they are not stopwatch times.

Host each PBR GLB separately and Hyper3D's floor falls to 9.45 seconds, while the measured SupaVoxel file stays at 44.92. For a site owner serving a rotating asset, that's the more useful thought experiment. Someone on a fast connection may not notice. Someone opening six heavy cards on a phone might. I benchmarked no real phone and no first visible frame, so I'd measure both before rolling out a gallery.

Untextured, SupaVoxel's geometry keeps its openings, but in the tested Original size file it carries a bigger transmission and working-set budget.

A month of gallery traffic, priced

Model a deliberately simple month: every GLB, hosted separately, gets 10,000 full uncached transfers, counted in decimal gigabytes with egress at $0.085 per GB. Hyper3D's 14.18 MB PBR adds up to 141.81 GB, modeled at $12.05. SupaVoxel's 67.38 MB Original size file adds up to 673.77 GB, modeled at $57.27. Under those assumptions the gap is roughly $45.22 — not an actual CDN bill, not a conversion forecast and not a claim that caching never helps.

That 67.38 MB file comes from one particular UI export. The CDN sums assume every byte is served again and again, not a traffic trace I observed.

Geometry memory, estimated

Using an illustrative budget of 32 bytes per exported vertex plus 4 bytes per triangle index, Hyper3D's 116,539 vertices / 120,000 triangles come to 5.17 MB and SupaVoxel's 893,267 / 1,499,910 to 46.58 MB. In other words, an estimated geometry allocation of 46.58 MB against 5.17 MB, not phone RAM that anyone measured. Left out entirely: the texture maps and their decoded expansion, any copies the browser makes, renderer buffers and general scene overhead.

SupaVoxel samples far more finely and gives you more vertices to edit; the 46.58 MB figure covers geometry only and is not measured device memory.

The case for SupaVoxel's 4K textures

SupaVoxel ships three 4096 × 4096 PNGs, Hyper3D three 2048 × 2048 PNGs. In both, the slots are baseColor, normal and metallicRoughness, with one material each. Each 4K image carries four times the pixels per map, which pays off only if a viewer zooms close enough and the UVs give useful texels to the area being inspected. I did not measure color error or UV occupancy, nor check whether zooming in really draws on the extra pixels.

Hyper3D's baseColor image takes 3,801,218 bytes; SupaVoxel's takes 14,363,914. Normal plus metallicRoughness come to 5,208,830 versus 6,427,581 bytes. In a PBR viewer those maps can shift highlights; in a monochrome resin slicer they add no geometry. If you assume a slicer use that ignores non-color maps, 5.21 and 6.43 MB respectively would be dead weight for that purpose — not bytes a slicer was measured discarding. For a pale cat on a gallery page, I would zoom into both textures before accepting four times the pixel count as a visual argument.

The pale SupaVoxel hero with 4K textures is easy to inspect; resolution by itself doesn't quantify how closely it resembles the pale source art.

Head-on, which cat looks like bones?

Matched front renders show that both candidates hold on to ears, legs, the openings in the skull and the gaps through the chest. Hyper3D's 120,000-triangle mesh produces more thin-looking ribs you can read one by one. SupaVoxel's 1,499,910-triangle mesh produces broader arches that overlap more heavily. Triangle count doesn't decide the look: here, more of them did not yield more distinct-looking bones.

More narrow, separate ribs on the Hyper3D front, even though it has a twelfth of the other file's triangles.

The SupaVoxel openings are real, not painted-on fake holes, yet the chest reads as bigger continuous curves.

The tail in profile

On a spinning card, the curled tail shows in profile. Hyper3D pierces it with small holes and stacks thin, layered ribs across the chest. SupaVoxel leaves its tail open too, so a "solid blob" label would be unfair; its version is simply a chunkier loop that looks braided, with broader unbroken bars wrapping the torso. Paired cameras make that difference more telling than either hero pose by itself.

Seen from the side, Hyper3D has little holes in the tail and slimmer struts in the chest. Nothing here tests feature size or breakage for a physical print.

SupaVoxel's tail remains an open loop, but from the same angle its ribs and tail look heavier.

A back that nobody photographed

Is the more open-looking back more accurate? There's no way to know. The source illustration left out most of the rear anatomy, so each tool guessed. Hyper3D came up with a slim spine flanked by ribs that look separate; SupaVoxel invented a wider continuous central back. At display distance that difference shapes the whole character of the cat as it turns, yet there's no true back to grade against. For this skeletal aesthetic I prefer Hyper3D's openness, without calling it a faithful reconstruction of something nobody could see.

Hyper3D's open-looking rear is an artistic guess, not anatomy recovered from the missing back view.

The joined SupaVoxel rear spine gives the product page a different character. Neither invented back can be scored for fidelity here.

What SupaVoxel's extra megabytes actually buy

Three 4K maps hold four times the pixels per map of three 2K maps; close-up texture inspection might benefit, though I measured neither perceptual gain nor UV efficiency. A mean edge of 0.3139 mm versus 0.7664 at normalized scale gives more surface samples for geometric edits. Both meshes clear the same welded closed-shell check. In solid-fill resin terms SupaVoxel encloses 38.95 cm³ to Hyper3D's 46.53 cm³ — $1.36 versus $1.63 at a hypothetical $35/L, a $0.27 advantage before supports and waste.

That is a trade, not a blanket endorsement. For finer maps and shape sampling you take on roughly 53.20 MB more hosted-file payload across the two GLBs as selected, plus an estimated 41.41 MB more geometry memory, and 7,217 slivers / 19 degenerates instead of 249 / 0. On this one input, the thinner-looking Hyper3D ribs still win visually. If the real brief is a sculpting master to examine up close, look into SupaVoxel; if it's this cat shown on phones at normal viewing distance, start with Hyper3D.

Hyper3D has fewer measured slivers — 249, not zero — and that cleaner local geometry supports its lightweight-gallery case.

Final verdict: the file I'd actually upload

For the tested full-file 3D gallery and this particular skinny-ribbed source, it's Hyper3D Gen-1.5: more delicate-looking ribs, a welded watertight shell and a 14.18 MB extracted PBR. Let downloaders know the real browser archive is 23.15 MB. SupaVoxel deserves a controlled next run if its 4K maps or editing density are a requirement, but I'd weigh its Compressed variant first and check the 19 degenerates and 7,217 slivers. From these renders, I wouldn't sell either as a tested physical skeleton.

Which one I would use for this job

When your art direction wants a high-resolution texture pass or a finely sampled model you plan to keep editing, try SupaVoxel with a source image of your own. Here it packed three 4K PNGs into a single Original size GLB that downloaded directly; that file is heavy, and its weaker local triangle health is not imaginary. A fair second shot means measuring the Compressed GLB it also offers before you fix any deployment budget. If all you want is a light, spinning version of this very cat, I still pick the Hyper3D file I tested.

How the files were measured, and the open questions

Both services got the same independently generated image. Hyper3D Gen-2.5's free-account export was blocked, so it appears in no file figure. From the downloaded ZIP, which holds two files, I pulled out the actual Gen-1.5 PBR GLB and set it beside SupaVoxel's downloaded Original size GLB. Offline renders used fixed front, side and back viewpoints; byte sizes, triangle/vertex counts and image slots came from the GLBs themselves, and coincident UV-split positions had to be welded before checking watertightness. I did not slice or print either model. A screenshot under matched lights is not a numeric albedo measurement.

Every size is decimal. The ideal bandwidth figures assume complete, steady 12 Mbps or 100 Mbps transfers with no overhead. CDN estimates assume $0.085/GB and 10,000 uncached full loads; memory estimates assume 32 bytes/vertex and 4 bytes/index and leave textures out. Not measured: actual memory, UV occupancy, albedo ΔE, first-frame time, the size of a compressed export, minimum rib thickness, a successful print, repeated failure rate or license terms.

Hyper3D Figurine Review 2026: A Solid Shell With Flattened Wings

Noor Halvorsen — Sun, 11 Oct 2026 15:48:58 +0000

Independent review. I ran both tools on my own accounts, free and paid, with no vendor-provided access or early access of any kind.

I tested a brief for a little seated winged figure that someone might turn into a physical miniature. The photograph had curls, a sword, a gold shield and boots hanging off a ledge. It did not show the back. I gave that exact picture to Hyper3D and SupaVoxel, downloaded the files each account could actually export, and inspected the geometry beneath the attractive previews.

The answer depends on what I mean by “make.” A rotating digital close-up and a print-shop starting mesh are not the same commission.

My verdict in 60 seconds — for an unrepaired physical-print starting file, I would take Hyper3D Gen-1.5; for a digital rear-view close-up, SupaVoxel’s more articulated wings win. Hyper3D's downloadable Gen-1.5 mesh becomes one watertight, edge-connected shell with zero nonmanifold edges after UV-seam welding. SupaVoxel draws the invented back feathers more convincingly but has 471 nonmanifold edges and 315 face-connected regions. I would start the print preparation with Hyper3D. For a digital rear-view hero, I would choose the layered wings from SupaVoxel; that is a different brief, not a quiet reversal of the print verdict.

My eight-line decision card weighs the unrepaired print candidate. Each line pairs a measured check or visible observation with its practical consequence; none is a score for the whole platform:

Welded watertight shell — Hyper3D yes, SupaVoxel no — only Hyper3D clears this prerequisite without an unmeasured repair.
Nonmanifold edges — 0 vs 471 — 471 problem edges become the first SupaVoxel workshop task.
Face-connected regions — 1 vs 315 — vertex-only touching is not the same as one coherent printable skin.
Boundary edges — 0 vs 0 after welding — a genuine tie, not proof that both are watertight.
Consistent winding — yes vs yes — another measured tie.
Degenerate / sliver faces — 0 / 344 vs 0 / 20 — SupaVoxel wins this narrow triangle-hygiene check.
Rear feather articulation — broad joined Hyper3D panel vs distinct SupaVoxel tiers — SupaVoxel wins the visible digital detail.
Thin-part verification — no sword/wing thickness test on either file — neither earns a print-ready certificate.

The shell defect is a gate for an unrepaired print. More attractive feathers cannot cancel it. Both still need slicing and a physical test.

Did both tools receive the same picture?

One independently generated 1,650,798-byte PNG went to both tools; no rear reference was supplied. Publication and reuse rights for the reference-influenced image still require review.

This was one image, one Hyper3D Gen-1.5 downloadable run and one SupaVoxel run. Hyper3D's additional Gen-2.5 trial is an access experiment, not the source of the file in these paired pictures. The same uploaded bytes do not mean the two systems performed identical internal image processing. My question is narrower: which delivered mesh would I hand to a person preparing this specific 120 mm-scale figure? If the figure sits on a shelf, somebody must also design its support or mounting. The photographed ledge was scenery; it was not evidence of a modeled display stand.

What survives when I inspect the two fronts?

Hyper3D Gen-1.5 keeps the sword, shield, curls, both wings and hanging boots on a 120,000-triangle downloadable mesh.

SupaVoxel's actual Original size GLB keeps the same recognizable props and stronger raised shield and feather forms at 946,166 triangles.

At thumbnail size these are both successful figures. The failure I care about doesn't announce itself in either front render. The SupaVoxel file has 7.88 times the triangles and 494,427 file vertices against Hyper3D's 114,101. That is more sampling capacity, not 7.88 times the likeness and not 7.88 times the usable print resolution. At a shared longest bounding-box side of 120 mm, SupaVoxel occupies 72.17 × 120 × 96.97 mm; Hyper3D occupies 91.56 × 120 × 88.33 mm. That is not a claim that both are exactly 120 mm tall in a maker's chosen orientation. My first practical step would be to inspect sword and wing-root cross-sections at the intended scale. Neither a beautiful preview nor those bounding boxes gives me a minimum wall thickness.

Why did I weld the mesh before counting holes?

Removing Hyper3D's materials leaves a recognizable sword and wing silhouette; the topology result comes from file parsing, not this clay picture.

SupaVoxel still has visible shield and feather relief without paint; that relief does not close its nonmanifold edges.

A GLB may duplicate vertices along UV and normal seams. Counting those as open boundaries before welding would falsely describe a texture seam as a physical hole. After merging coincident seam vertices, Hyper3D is one face-connected shell, zero boundary edges, zero nonmanifold edges and consistent winding. SupaVoxel also has zero boundary edges and consistent winding. But it has 471 nonmanifold edges and 315 face-connected regions. If I permit surfaces touching at one vertex to count as connected, its component count falls to one. A corner touch does not give a slicer the same unambiguous solid as a shared, correctly oriented edge. That distinction is why the superficially similar “zero boundaries” line does not produce a print tie.

Do the rear wings make the extra repair worthwhile?

Hyper3D joins the wing roots into a broad central white panel; individual feather forms collect toward the sides.

SupaVoxel leaves the gold collar visible between two wings and lifts several feather tiers into separate silhouettes.

Yes, if the commission is a rotating digital collectible whose back fills the frame. This is SupaVoxel's clearest win in the pair. I cannot call either back faithful because the input contains no back. I can call the delivered SupaVoxel back more articulated in these matched renders. The exchange rate is unusually concrete: for that rear close-up I accept a 36.59 MB Original size GLB instead of a 15.25 MB extracted Hyper3D PBR GLB, and, if the object must become physical, an unmeasured repair of 471 nonmanifold edges before trusting a solid-volume or resin estimate. I would happily pay that file-and-edit cost for a digital hero. For a time-sensitive print I would not swap an already coherent shell for prettier invented plumage without actually performing and validating the repair.

Does the wireframe certify a printable sword?

The Hyper3D wire view shows where the 120,000 triangles sit; 344 are long-thin under the common check.

SupaVoxel's denser 946,166-triangle wire view contains just 20 long-thin faces, but the line tangle cannot diagnose 471 nonmanifold edges by eye.

At the 120 mm longest-side normalization, estimated mean triangle edge length is 0.8304 mm vs 0.3283 mm. A 0.4 mm FDM nozzle offers a familiar scale reference; it does not mean every 0.3283 mm edge survives printing, nor that a sword blade is at least 0.4 mm thick. SupaVoxel earns its win on the narrow sliver check, 20 against 344, and both meshes have zero degenerate faces. Hyper3D wins the more basic shell-integrity gate. In the shop I would first repair the SupaVoxel defects, then check local blade thickness and supports on both. Buying more mesh samples does not buy that second inspection.

How much downward-facing surface did the screening flag?

Hyper3D's side view exposes a thin wing profile, sword and suspended boots; none was measured for strength.

SupaVoxel's wings look more voluminous at the paired angle; this is not a measured reduction in required supports.

Under one deliberately simple screen—world Y up, faces pointing within 45 degrees of downward −Y, same 120 mm longest-side scale—I count 5,458.3 mm² potentially downward-facing area for Hyper3D and 5,588.2 mm² for SupaVoxel. Their fractions are 16.58% and 15.27% because their total surfaces differ. The SupaVoxel absolute area is 129.9 mm² higher while its percentage is lower: neither number is a slicer's support volume. Turn the winged figure on the bed and both change. If someone reads this as “SupaVoxel saves support resin,” they have swapped an orientation screen for a manufacturing trial. I did not generate supports, slice or print either object.

Can I quote a resin bill for both models?

No. Hyper3D's welded mesh passes the shell checks required for an enclosed-volume estimate: 181.75 cm³ at the shared 120 mm longest-side normalization. A hypothetical completely solid pour at $35/L gives $6.36 before hollowing, support, waste and failure. SupaVoxel's broken topology does not provide a validated enclosed volume, so I will not put a dollar figure beside its entry. The apparent signed volume of an uncorrected file is not a resin invoice. That missing comparison is not a rhetorical trick; I would need to repair the specific file and recalculate the solid. If the real job is a hollow resin figurine, neither the $6.36 number nor a purely visual feather win predicts the amount of consumable material.

Did the product screens deliver these actual files?

The successful Hyper3D Gen-1.5 export selected GLB, Shaded and PBR at 2K; the browser delivered a ZIP with both variants.

The measured SupaVoxel branch is GLB → Original size. The offered Compressed option was not measured here.

The Hyper3D file inspected above is the 15,247,868-byte PBR member extracted from its actual 24,544,968-byte browser ZIP. The first legally downloaded geometry-only Hyper3D GLB was 5,092,448 bytes without embedded textures; I did not compare that blank-stage file with SupaVoxel's colored output. SupaVoxel's file is a real 36,589,892-byte Original size browser export, not a mislabeled CDN asset. At a purely hypothetical 12 Mbps, their actual UI delivery payloads imply 16.36 and 24.39 seconds of ideal transfer respectively. Hyper3D has a genuinely lighter handoff; SupaVoxel's visible feather advantage is therefore not free in either bytes or topology. The menu's Fix Mesh entry was present, but I did not run it, and a button is not a completed repair.

What happened before Hyper3D made the downloadable mesh?

Hyper3D Gen-2.5 showed a million-face preview, spent an isolated 0.5 Hyper3D credit, and refused its free-account download with a subscription prompt. No Gen-2.5 GLB was available for my geometry checks. I switched to a different Gen-1.5 generation: free preview 25.064 seconds, geometry confirmation 14.299 seconds and another 0.5 Hyper3D credit, with separately verified zero extra credit for Material Generate and Material Confirm in this run. Those known 39.363 seconds omit material, operator time and download. SupaVoxel's matched request lasted 193.746 seconds between database timestamps including queue but excluding upload and export; its transaction was 3 SupaVoxel credits. These are not exchangeable currencies or equivalent stopwatches. The interrupted browser gap before finishing Hyper3D's Gen-1.5 confirmation was operator downtime, not nine hours of generation.

Which one would I put on my shop ticket?

If the ticket says “unrepaired starting shell for a 120 mm-class physical miniature,” Hyper3D Gen-1.5 wins this sample. Zero nonmanifold edges versus 471 is a gate, not a fussy aesthetic preference. I would still inspect the sword, wing roots, mounting and thin details in a slicer before authorizing a print. If it says “spinning character in a digital storefront with a rear close-up,” SupaVoxel wins the visible feather treatment, and its larger file and need for eventual physical repair may be tolerable. If the back never appears, the extra feather articulation may buy nothing. We have one example, no print, no completed repair and no empirical failure rate; those are constraints on how far this decision travels, not reasons to pretend the two delivered files are equal.

Final verdict: why did the print starter go to Hyper3D?

Because the topology really did. The Hyper3D downloadable Gen-1.5 PBR mesh was a watertight, consistently wound, single edge-connected shell after fair UV welding. The SupaVoxel mesh was not, despite a better-looking invented back and excellent sliver count. That is the actual rival win and its exchange rate: a 21.34 MB smaller standalone PBR GLB and no measured manifold repair, in return for a conspicuously flatter central wing span. I would not tell a maker that a visually superior back erases an unresolved solid; nor would I tell a digital artist that an ideal shell erases an obvious rear-view flattening.

Use SupaVoxel when the back is the deliverable

If your deliverable is a textured rotating figure rather than an unrepaired print file, put your own winged reference through SupaVoxel and inspect the back and clay view before choosing. In this one paired run its raised feather tiers are the reason to choose it. If that same file has to become a miniature, make mesh repair and a real slicer/print test explicit stages; the original file's 471 nonmanifold edges are not a detail to hide from the fabricator.

How I tested it. Same independent source PNG and no back reference; different products' actual downloaded textured GLBs measured offline, with matched preset camera and light. The SupaVoxel render's yaw was adjusted 90° to present the front like the reference; matched views compare subjects, not pixel-registered reconstructions. Parsed file counts and welded topology are measurements. Averages at 120 mm, transfer seconds and $35/L solid-fill resin are stated-assumption calculations. Back-wing quality is my judgment from paired renders. Local thickness, support generation, slicing, actual print behavior, publisher rights to the reference-influenced input and repeated-run reliability remain untested.

Hyper3D Rodin Cost Review 2026: 0.5 Credit, Then a Download Wall

Noor Halvorsen — Fri, 09 Oct 2026 15:46:03 +0000

Disclosure: this is an independent hands-on test. Both tools were run on ordinary customer accounts; neither company supplied review access or saw this piece before publication.

I did not set out to investigate a paywall. I needed a textured model of a crumpled blue container I could carry out of the product and inspect. Hyper3D’s Gen-2.5 preview appeared; the account had paid 0.5 of its credits; the Download button looked actionable. I clicked it and got a subscription demand. The usable file came only after I ran the image again in Gen-1.5. I kept every balance check and downloaded artifact separate, because the entire buying decision turns on which job actually produced a file.

My verdict in 60 seconds — for a straightforward textured digital delivery, I would take the SupaVoxel GLB rather than gamble on Hyper3D’s Gen-2.5 Free-tier preview becoming a file. That reflects this recorded access path, not a credit exchange rate or a planter-suitability certificate. On the tested Hyper3D Free account, Gen-2.5 used an independently isolated 0.5 Hyper3D credit but blocked its GLB download. A separate Gen-1.5 generation spent another 0.5 Hyper3D credit at Geometry Confirm, then delivered textured PBR and Shaded variants in a ZIP; both material actions cost zero additional credits in this case. SupaVoxel’s matched one-request transaction deducted 3 of its own credits and produced a real 58,065,236-byte Original size GLB. Those denominations do not make SupaVoxel six times the cash cost. If the deliverable is irregular foil as a digital asset, SupaVoxel gives the stronger surface in this pair. If the deliverable is a planter whose first file shows a bottom, Gen-1.5 Hyper3D earns that different choice.

My eight-line decision card is about what reached my drive. Each line separates observed facts from their consequence for a digital-asset workflow:

Gen-2.5 Free-tier file access — Hyper3D blocked vs SupaVoxel’s tested route delivered — a preview is not an export.
Gen-2.5 credit attribution — isolated 2.0→1.5 Hyper3D credits — charged preview, no file.
Gen-1.5 route required — separate Hyper3D generation vs one SupaVoxel request — the detour was real, though avoidable if you choose Gen-1.5 first.
Gen-1.5 confirmed cost — Hyper3D geometry 0.5; material Generate 0, Confirm 0 extra — no fabricated surcharge.
SupaVoxel case debit — one request-linked 3-credit transaction — not an inter-platform price score.
Actual browser delivery — Hyper3D 21.471 MB ZIP vs SupaVoxel 58.065 MB GLB — Hyper3D wins bytes decisively.
Usable exterior irregularity — Hyper3D broad paired folds vs SupaVoxel varied crossings — SupaVoxel wins the visual brief.
Visible planter floor — Hyper3D blue floor vs SupaVoxel through-passage — Hyper3D wins the functional-shape candidate, with no leak test.

The decision turns on which deliverable you need — and whether your plan actually lets you take it away.

What did the Hyper3D balance say before Generate?

The tested Free account was independently checked at 2.0 credits before the case’s Gen-2.5 generation; the Generate tariff displayed 0.5.

The same independently generated 1,647,256-byte PNG went into both products. That controls input bytes, not each product’s preprocessing or plan. Publication/reuse rights to the reference-influenced source remain uncleared. Hyper3D’s Free balance began at 2.0; its assigned account had no other work in the isolated test interval. I checked the balance rather than equating a displayed tariff with a real debit.

Did Gen-2.5 charge at the download click?

During preparation/generation the checked balance still read 2.0; after the preview it read 1.5.

No. The isolated balance changed 2.0→1.5 by completed preview; Confirm and refused download left it at 1.5. Generation took 66.417 seconds and Confirm 21.066, or 87.483 seconds of known server stages, with no second debit. That is not upload-to-export time: no export happened. On this Free account, the charge arrived before the subscription gate. Other plan entitlements remain untested.

Why did one million faces fail my practical test?

Gen-2.5 displayed a million-face preview in the browser; that number belongs to its product panel, not to an inspected download.

The UI claimed 1,000,000 faces and 587,083 vertices. I could not verify that mesh because no Gen-2.5 file arrived; I did not substitute internal preview resources. SupaVoxel’s 1,500,000 triangles were read from its actual GLB. The Hyper3D file analyzed in this series is a different Gen-1.5 generation at 120,000 triangles. A viewer’s preview badge and a parsed downloadable mesh are different evidence classes.

What exactly happened at Download?

After choosing GLB and pressing Download, the Free account was told to subscribe for Gen-2.5; Gen-1.5 downloads remained free.

I used the visible Pack → GLB → Download path. The product explicitly demanded a subscription for Gen-2.5 files after the verified 0.5-credit debit: one charged attempt, zero downloadable Gen-2.5 GLBs on this Free account. Dollar value, subscription price and paid-plan access were not measured. The operational problem is specific: a finished preview and enabled-looking Download control did not deliver a file on this tier.

Was Gen-1.5 a free unlock of that preview?

The image entered a separately configured Gen-1.5 Detail task; this was a new generation, not the former preview converted to an allowed file.

No. Gen-1.5 was a new task; it even auto-described the container as a drinking cup, unlike Gen-2.5’s caption. Its observed preview took 25.955 seconds and left the balance 1.5→1.5. Our two-task route ultimately spent 1.0 Hyper3D credit, but someone choosing Gen-1.5 immediately would avoid the first 0.5-credit detour. Neither its model settings nor its caption matched Gen-2.5 exactly, so visual differences cannot be assigned solely to version.

Where did the Gen-1.5 half-credit go?

Geometry Confirm took 24.499 seconds and changed the isolated Hyper3D balance from 1.5 to 1.0.

Preview plus confirmation took 50.454 seconds, excluding material, upload, pauses and delivery — not “a full PBR in 50 seconds.” The first legal Gen-1.5 geometry GLB was 4,484,160 bytes with zero embedded maps. Comparing that white-stage file to SupaVoxel’s textured 58.07 MB export would be dishonest. I continued through both material actions to obtain a comparable colored file.

Did Material Generate and Confirm cost extra?

The verified balance was 1.0 before the separate material actions and remained 1.0 after both.

After Material Confirm at unchanged 1.0 credits, PBR export was enabled for this case.

Zero additional credits in this run: separately checked 1.0→1.0 across both Material Generate and Confirm. Thus the completed Gen-1.5 PBR file follows one verified 0.5-credit geometry debit here. We did not measure the material stages’ complete durations or prove they are always free. The finished PBR has three 204⁸² PNG maps, unlike its textureless geometry sibling.

How many bytes left Hyper3D’s website?

The tested GLB export with Shaded and PBR selected delivered a ZIP containing two variants, not just the standalone PBR member.

21,471,080 bytes — the full browser ZIP. Inside it is a 14,038,528-byte PBR GLB and a 7,432,300-byte Shaded GLB. The PBR file is what I used for the comparable textured renders and parsed mesh counts. After geometry confirmation, the route was Material Generate → Material Confirm → GLB choice → Download, four recorded interface actions, then extraction to obtain the standalone PBR member. At a hypothetical ideal 12 Mbps, the entire ZIP’s pure transfer floor is 14.31 seconds; if I host only the extracted PBR member separately, its different 9.36-second floor applies. Swapping those denominators would understate the payload a Hyper3D user actually received. There is a real byte win here against SupaVoxel’s chosen Original size branch, even after counting the whole ZIP.

What did SupaVoxel’s own transaction say?

The image-matched SupaVoxel request was linked to exactly one −3-credit transaction; the recorded request interval was 328.196 seconds.

One source-PNG-matched request, one successful outcome, one linked 3-SupaVoxel-credit debit. Its database created-to-updated delta is 328.196 seconds, including its queue and excluding upload and export. That is not equivalent to Hyper3D’s 50.454-second subset of Gen-1.5 preview plus confirm; any full-delivery-speed ranking from these different time windows would be fiction. These are also two different companies’ credits. Multiplying 3 by 100 and 0.5 by 100 creates two ledgers in different units, not a relative dollar price. The SupaVoxel interface estimated five to ten minutes, a forecast consistent with this single request’s narrower lifecycle; it was not a promise of total time to file. I would use the actual transaction for this shared testing account rather than its later top-bar balance, which included other cases.

Did its GLB menu deliver the file in my comparison?

The existing SupaVoxel project was reopened without regeneration and the UI’s Original size option yielded the actual 58,065,236-byte browser blob.

Yes. Export → GLB → Original size produced a real 58,065,236-byte GLB, with 1,500,000 triangles and three 4096 × 4096 PNG maps. This is the one behind the article’s paired model renders. A smaller 8,499,144-byte compressed CDN resource exists for the same task, but nobody in this test measured that as a click on the UI’s Compressed option. A 44,556,092-byte local analysis derivative is not a browser export either. The extra files cannot silently shrink the cost of my selected route. On a hypothetical ideal 12 Mbps line, 58.065 MB takes 38.71 seconds of pure transfer against Hyper3D’s 14.31-second ZIP — a 24.40-second difference before protocol or decoding. Hyper3D wins this part. SupaVoxel’s better irregular foil costs more bytes with this exact export choice.

Does one hundred modeled planters change the bill?

Only as a scenario, not a completed batch. Repeat these exact charges and export selections 100 times, assume no failures or caps: Hyper3D Gen-1.5 implies 50 Hyper3D credits and 2.1471 GB of UI ZIP downloads; SupaVoxel implies 300 of its own credits and 5.8065 GB of Original size GLBs. Those are separate internal currencies and decimal byte totals. If I hosted only extracted Hyper3D PBR members, that would instead be 1.4039 GB; the clients’ browser ZIP transfers do not disappear retroactively. The one-off Free allowance and future account policies are not proven to cover 100 successful exports. Extrapolating Hyper3D’s known 50.454-second preview/confirm subset gives 1.4015 hours, and SupaVoxel’s 328.196-second request lifecycle gives 9.1166 hours, but those clocks omit different stages; they are not rival batch delivery estimates. The calculation helps plan storage while warning me not to invent a throughput winner.

Where does Hyper3D genuinely win the shape decision?

Hyper3D Gen-1.5 shows a blue inner floor in the rendered cavity; no physical leak test was performed.

SupaVoxel’s more irregular model shows white background through the cavity’s lower opening despite a topologically watertight welded shell.

Hyper3D’s blue floor is the better initial geometry if the job is a container candidate. Both meshes are watertight, consistently wound and one face-connected shell after UV welding, but topological watertightness does not mean a pot can hold soil or water. SupaVoxel’s through-passage is visually obvious; it would need a deliberately modeled base before this specific file meets a vessel brief. No print, slicer or water test was done on Hyper3D either, so I will not call its apparent bottom leakproof. The exchange for SupaVoxel’s attractive uneven exterior is therefore unknown base-design and validation labor plus a 58.065 MB Original size file. For a digital prop the missing base may be a simple art-direction change; for an actual planter it is the primary unresolved requirement.

Final verdict: which credit purchase led to a useful file?

My Gen-2.5 Hyper3D half-credit bought a preview but no Free-tier download. Another half-credit on a separate Gen-1.5 run, plus two independently zero-extra-credit material actions, led to a real textured ZIP. The SupaVoxel request’s separate three-credit transaction led to a genuine GLB with stronger irregular folds. Hyper3D gave a smaller download and an apparent interior bottom; SupaVoxel gave the better foil surface and a simpler recorded one-job digital delivery. For the recorded digital acquisition path, SupaVoxel won without erasing the difference between the two deliverables. To quote either as a purchasable, leakproof planter would be getting ahead of every physical test we did not run.

Use SupaVoxel if the deliverable is visual foil first

For an irregular blue foil study you will rotate on screen or edit before fabrication, try your image in SupaVoxel and inspect the actual exported file, not only the result card. This tested route delivered the crossed, uneven folds in the visual brief. If you intend to put soil in it, add a real base-design and leak-validation stage before making promises. If your immediate requirement is an apparently bottomed file on a Free Hyper3D account, go directly to tested Gen-1.5, not Gen-2.5’s gated preview.

How I tested it. I isolated Hyper3D balance changes around each task and material step, matched the SupaVoxel transaction to this exact input’s request, and measured each browser-delivered artifact separately from the extracted PBR member and other derivative copies. Credit numbers are measured within their own services only. Transfer and ×100 scenarios are arithmetic using explicitly stated bandwidth and identical-repeat assumptions. Full upload-to-file elapsed time, cash cost, subscription entitlement beyond this Free account, floor thickness, a successful physical pot and publication rights for the reference-influenced source were not measured.

Also in this series
Hyper3D Rodin Planter Review 2026: A Base Matters More Than a Mesh
Hyper3D Rodin 3D Review 2026: The Foil Folds Turn Into a Mask

Claude Code in VS Code: The Official Extension, Set Up Properly

Noor Halvorsen — Wed, 07 Oct 2026 03:11:48 +0000

Quick correction before anything else, since it still trips people up: if you search "claude code vscode", a lot of what comes up tells you there's no official Claude Code integration and you should wire Claude in through a third-party extension and an API key. That was true once. It isn't now. Anthropic ships a Claude Code extension for VS Code, its docs call it the recommended way to use Claude Code there, and you don't need an API key to use it.

I spend most of my week inside existing repos: reading code someone else wrote, fixing it, and reviewing what Claude proposes before it lands. That's where this extension earns its keep. (For brand-new throwaway sites I don't open VS Code at all. I start in begin, which turns a prompt into a Next.js/React/Tailwind project I can then pull into VS Code like any other repo.)

What follows is how I set the extension up, the settings I change on day one, and a few things that aren't obvious from the marketplace page.

Claude Code VS Code extension: what you need first

From the official VS Code docs, the prerequisites are short:

VS Code 1.94.0 or later
A paid Claude subscription (Pro, Max, Team or Enterprise) or a Claude Console account. No API key is required; you sign in through the browser the first time you open the panel.

If your company reaches Claude through Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry, that's supported too. You tick Disable Login Prompt in the extension settings and configure the provider in ~/.claude/settings.json.

Installing is the usual Extensions view search for "Claude Code", or from a terminal:

code --install-extension anthropic.claude-code

Two details I wish I'd known earlier:

The extension's version number is the Claude Code version it bundles. When the docs say a feature needs v2.1.280 or later, check the extension version in the Extensions view. That's the number that matters.
It does not put claude on your PATH. The panel uses a private bundled copy of the CLI. If you also want claude in the integrated terminal, do the standalone CLI install separately.

Cursor and other VS Code forks can install it too, either from their own extension search or from Open VSX.

The first ten minutes

Open any file, and a Spark icon shows up in the top-right editor toolbar. That's the fastest way in. It only appears when a file is open; having just a folder open isn't enough, which confused me for a minute. You can also click ✻ Claude Code in the status bar, which works with no file open, or use the Spark icon in the Activity Bar to see your sessions.

The habit that changed the most for me is selection-as-context. Whatever you highlight, Claude already sees. Press Option+K on Mac or Alt+K on Windows/Linux and it inserts a precise reference like @app.ts#5-10 into the prompt. That beats pasting code into a chat window, because Claude knows exactly which file and lines you mean and can go read the surrounding code itself.

Other context tricks worth learning on day one:

@src/components/ (with the trailing slash) pulls in a folder; @auth fuzzy-matches auth.js, AuthService.ts and friends.
@terminal:name references a terminal's output by its title, so I stop copy-pasting stack traces.
Paste a screenshot straight into the prompt box. Hold Shift while dragging to attach files.

Pick a permission mode on purpose

This is the part older tutorials skip entirely, and it matters more than prompt phrasing. The mode indicator at the bottom of the prompt box has four options:

Mode	What happens	When I use it
Plan	Claude writes a plan and waits. VS Code opens it as a Markdown doc you can comment on inline.	Anything touching more than two files
Manual	Asks before file edits and most shell commands; shows a side-by-side diff per edit.	Unfamiliar or sensitive repos
Auto	A classifier reviews most actions instead of asking you.	Well-tested repos where I trust the guardrails
Edit automatically	Edits files without asking.	Scratch branches only

With Claude Code v2.1.283 or later, Auto is the built-in starting mode. I override that. Plan mode with inline comments is the single best feature in the extension: I correct the approach in the plan, before any code exists, instead of arguing with a diff afterwards. You can also type /plan fix the auth bug to jump straight in.

In Manual mode, diffs now have per-change Accept this change and Reject this change buttons (v2.1.275+), so you're no longer forced into all-or-nothing on a file. One caveat: diffs with more than 100 changes open without those buttons.

Settings I change on day one

There are two settings files, and mixing them up cost me some time:

VS Code user settings for the extension's behaviour (claudeCode.* keys).
~/.claude/settings.json for things shared with the CLI: permissions, hooks, MCP servers, environment variables.

My VS Code side:

{
  "claudeCode.initialPermissionMode": "plan",
  "claudeCode.preferredLocation": "sidebar",
  "claudeCode.useCtrlEnterToSend": true
}

initialPermissionMode is only read from user settings. Workspace values are ignored, so don't bother committing it to .vscode/settings.json. useCtrlEnterToSend stops me from sending half-written prompts when I hit Enter for a new line (Shift+Enter also works).

On the Claude Code side, the docs suggest adding "$schema": "https://json.schemastore.org/claude-code-settings.json" to ~/.claude/settings.json so VS Code autocompletes and validates it. Then I add a Read deny rule for .env. The docs point out that a matching deny rule stops both selected text and the open-file notice for that file from reaching Claude.

Checkpoints: the undo button I lean on

Hover over any message and you get a rewind button with three choices: Fork conversation from here, Rewind code to here, or both. "Rewind code to here" is what makes Edit automatically mode survivable. If Claude goes off in a bad direction across six files, I roll the files back to before that message and keep the conversation, so I can explain what went wrong.

Panel or terminal?

Both, honestly. The extension and the CLI share conversation history, so I start in the panel and run claude --resume in the terminal when I need something CLI-only. Per the docs, the gaps are:

The panel exposes a subset of commands and skills (type / to see them); the CLI has all of them.
The ! bash shortcut and tab completion are CLI-only.
MCP config and checkpoints work in both. In the panel, /mcp opens a dialog to add and manage servers.

If you just prefer the terminal UI, the Use Terminal setting launches Claude in terminal mode inside VS Code.

Things that bit me (so they don't bite you)

ANTHROPIC_API_KEY set in your shell but still getting the sign-in screen? VS Code may not have inherited your shell environment. Launch it from a terminal with code ..
No Spark icon? Open a file, check you're on 1.94.0+, reload the window, and temporarily disable other AI extensions. The docs specifically name Cline and Continue as possible conflicts, which is ironic given how many guides tell you to install Continue for this.
Untrusted repos. With auto-edit on, Claude can modify .vscode/settings.json or tasks.json, which VS Code may execute. The docs recommend Restricted Mode for untrusted workspaces (the extension doesn't run there) and Manual mode for edits.
Usage surprises. /usage shows your plan's limit bars, with reset times, and flags what's eating them, like cache misses or subagent-heavy sessions.

Wrapping up

The official extension does what people were hacking together with third-party plugins a year ago, plus things those never had: plan review with inline comments, per-change diff review, checkpoints, and one session history shared with the CLI. Set the permission mode deliberately, learn Alt+K, and keep secrets behind a deny rule.

And when the job is a fresh prototype rather than an existing codebase, I still let begin produce the first build, then open that repo in VS Code and let Claude Code take it from there.

FAQ

Is there an official Claude Code extension for VS Code?
Yes. It's published by Anthropic (anthropic.claude-code), and Anthropic's docs call it the recommended way to use Claude Code in VS Code.

Do I need an Anthropic API key for Claude Code in VS Code?
No. A paid Claude plan (Pro, Max, Team, Enterprise) or a Claude Console account works; you sign in through the browser.

Why can't I run claude in the VS Code terminal after installing the extension?
The extension bundles a private CLI and doesn't add it to your PATH. Install the standalone CLI to use claude in any terminal.

Does Claude Code work in Cursor?
Yes. Install it from Cursor's extension search or Open VSX. If your editor can't install it, run the CLI in its integrated terminal.