DEV Community: Elias Larsen

Higgsfield Pricing Review 2026: 2 Credits per Matcha Image

Elias Larsen — Sun, 18 Oct 2026 15:53:47 +0000

Higgsfield debited two credits per matcha image; Kyncept debited one. The dollar figures below are dated plan allocations, not charges on the tested accounts.

Independent review: both products were run on my own accounts — a free tier and a paid plan — with no vendor-provided access.

I compared visible balances across seven completed text-only jobs on October 1–2, 2026: two images and two videos on Kyncept, two images and one video on Higgsfield. See https://kyncept.com and https://higgsfield.ai for the respective products; the receipts below apply only to these settings.

October 1 screenshots show $59/month and 1,200 credits on both products' separate plans; no October 2 price check or plan purchase occurred. Higgsfield video and image jobs used different free accounts; Kyncept used an account with an existing balance. The dollar figures allocate dated monthly allowances within each product, not a paid test invoice, free-tier quota, or exchange rate between companies.

My verdict in 60 seconds

Scope: seven completed jobs, one matcha-miniature brief, text-only input, selected 16:9; not a platform-wide benchmark.
Dated plan snapshot: Kyncept Pro and Higgsfield Plus each displayed $59/month and 1,200 credits on October 1.
Image receipt: Kyncept debited one credit per finished image; Higgsfield debited two for each of its two finished images.
Image scenario: at each product's own screenshot-plan allocation, approximately $0.0492 per Kyncept image versus $0.0983 per Higgsfield image for these settings.
Video receipt: Kyncept debited 10 credits for each delivered 8.000-second clip; Higgsfield debited nine for one selected five-second Wan 2.5 Fast clip, delivered at 5.042 seconds.
Video limitation: nine Higgsfield credits per five-second selection is not a cheaper same-length alternative to 10 Kyncept credits per eight-second output.
Audio: the files contain AAC audio tracks with non-silent peaks, but nobody listened to verify requested effects or synchronization; no sound-performance grade follows.
Decision: this specific image workflow favors Kyncept on credits. For video, decide first whether the selected five- or eight-second deliverable meets the job, then budget the distinct clips.

No score out of 100 is warranted from seven jobs, different video lengths, and unverified audio and licensing. The balance changes are more useful than a decorative rating.

The cover juxtaposes original outputs and their observed image debits; it does not depict an equal-duration video comparison.

The $59 comparison has a date and a condition

October 1 screenshots label Kyncept Pro and Higgsfield Plus at $59 and 1,200 credits each. On either pictured plan, $59 ÷ 1,200 is $0.04917 per credit as a conditional allocation. Higgsfield credits cannot be spent at Kyncept or compared without the model-specific debit and deliverable; no paid-plan charge occurred in these trials.

The October 1 Kyncept Pro monthly screenshot supplies the $59 and 1,200-credit assumption; it is not proof of today's checkout price.

The matching dated Higgsfield Plus screenshot supplies its separate $59 and 1,200-credit assumption.

Higgsfield and Kyncept prices and model debits can change. Neither checkout, tax, renewal, rollover, nor remaining monthly quota was verified. Check those before budgeting; the allocations below are not a billing audit.

Three wallets, not one combined test balance

Kyncept's existing-balance account showed 311 → 310 → 309 for images and 309 → 299 → 289 for videos. Higgsfield's video free account showed 10 → 1; a separate Higgsfield image free account showed 10 → 8 → 6. Higgsfield never had one wallet funding all three jobs in this test.

This cannot answer whether one new user could make a Higgsfield image and a Higgsfield video from the same free grant. The video wallet ended one credit short of another nine-credit job, while the image wallet's remaining six could not be transferred to it. Kyncept's starting 311 is not evidence of a comparable free allowance.

Kyncept's image-side starting balance was 311 on the existing-balance account.

This Higgsfield image account started at 10; it was not the account used for the Wan video.

The image debit is the cleanest price receipt

Both image pages received the identical text-only matcha prompt at selected 16:9 and returned two 1376 × 768 stills each. Both displayed Nano Banana Pro, but a shared label does not verify identical underlying versions. Higgsfield selected 1K.

Kyncept's balance moved 311 → 310 → 309: one credit per image. Higgsfield's separate image wallet moved 10 → 8 → 6: two per image. The observed two-image batch therefore used two Kyncept credits versus four Higgsfield credits. Under each company's own dated $59/1,200 plan, that would allocate about $0.0983 versus $0.1967 for two. Higgsfield's two-credit debit is specific to this selected path, not every image model.

The selected image settings and two-credit action are visible before submission; the balance sequence establishes the actual debit.

The first completed Kyncept image accompanies the 311 → 310 balance change.

The Higgsfield image wallet shows eight after its first two-credit job.

The same image wallet shows six after the second job: 10 → 8 → 6, not a mixed image/video ledger.

What a credit buys is not only a pixel count

Kyncept delivered JPEGs of 780,403 and 827,821 bytes; Higgsfield delivered RGBA PNGs of 1,660,430 and 1,801,245 bytes. All four are slightly wider than mathematical 16:9. Different encodings make bytes a delivery specification, not a quality score.

All four depict the plate, pancake, crane, steam, butter and scaffold, but miss exactly three workers: both Kyncept images and Higgsfield's first show roughly four; Higgsfield's second shows more. Both Higgsfield originals contain a burnt-in platform watermark; the tested Kyncept originals do not show comparable branded text. This proves neither Kyncept free-tier policy nor whether a paid Higgsfield plan removes the mark. Verify files and rights under the intended account before client use.

One unmodified Kyncept JPEG original. The image, not the model name, is the deliverable to inspect.

One unmodified Higgsfield PNG original; the burnt-in platform mark is visible in this tested output.

Video receipts: nine credits bought five selected seconds

The text-only video prompt requested eight seconds, 16:9 and synchronized Foley; no photo or frames were supplied. Higgsfield's sole tested mode, Wan 2.5 Fast, had five seconds, 720p, 16:9 selected with Enhance off. The nine-credit job completed, moving that free wallet 10 → 1, and the MP4 measures 5.042 seconds. Higgsfield succeeded at the selected duration; other free modes were not swept.

Kyncept Plus debited 10 credits per job and delivered two different 8.000-second MP4s. On the respective dated plans, Kyncept's clip would allocate about $0.4917; Higgsfield's five-second clip about $0.4425. They are not equal-length purchases, and two different Kyncept files do not imply a population success rate.

Kyncept's video-side receipt begins at 309 and shows a 10-credit action for the selected path.

The selected Higgsfield Wan submission shows the nine-credit action; its separate video wallet started at 10.

The free Higgsfield video wallet ended at one credit after its nine-credit job.

Per delivered second is an accounting lens, not a tariff

Dividing whole-job debits by selected nominal lengths gives Kyncept 10 ÷ 8 = 1.25 credits per second, about $0.0615/second on its pictured plan; Higgsfield 9 ÷ 5 = 1.8, about $0.0885/second on its own plan. Higgsfield delivered 5.042 seconds; five is the UI setting used for this planning ratio. Neither product is proven to bill per second; these are conditional averages across unequal clips.

A usable second also depends on the image and sound. Kyncept video two shows a giant hand with a brush; Higgsfield frames show a large brush, yellow liquid and white chef caps rather than uniformly yellow hard hats. Both files contain AAC with non-silent peaks, but the requested Foley, synchronization and absence of speech/music were not checked by listening. Higgsfield and Kyncept sound correctness therefore remains ungraded.

Samples from one 8.000-second Kyncept video; still frames cannot establish whether its sound effects are correct.

Samples from the completed 5.042-second Higgsfield video, not evidence of an equal-length eight-second offer.

Scenario A: budget 100 Higgsfield or Kyncept matcha images

For 100 images at unchanged selected debits, Kyncept would use 100 credits and Higgsfield 200 credits. Applied separately to the October 1 plans, 100 × $59 ÷ 1,200 ≈ $4.92 versus 200 × $59 ÷ 1,200 ≈ $9.83: a $4.92 allocation difference, not an extra bill on the tested free account. A full hypothetical 1,200-credit allowance yields 1,200 Kyncept or 600 Higgsfield images if debits remain fixed.

Those counts exclude rejected images, retries, curation, watermark policy and rights. Both sides missed the requested worker count in this case; Higgsfield's tested originals carry a mark. Neither calculated image count promises 100 usable client deliverables.

Scenario B: collect 60 raw seconds on Higgsfield or Kyncept

To gather at least 60 seconds of raw clips, not one continuous shot, eight Kyncept clips × 8 = 64 seconds / 80 credits; twelve Higgsfield clips × 5 = 60 seconds / 108 credits. On the respective dated plans, that allocates approximately $3.93 versus $5.31. Higgsfield and Kyncept totals have different lengths because clips are indivisible. These figures assume repeatable debits and eligible credits; they omit failure, retries and plan restrictions, and are not invoices.

Higgsfield's tested free video account ended at one credit, too few for a second identical job. Kyncept video two took 104.8 seconds on the UI; its first reliable time is unknown. Higgsfield's result was first observed by 127.05 seconds, an upper bound rather than an exact completion time. Neither the speed records nor the audio metadata promise a usable, coherent minute.

What the plan balance could buy, in theory

At unchanged per-job debits, 1,200 credits could hypothetically fund 120 Kyncept eight-second clips, or floor(1,200 ÷ 9) = 133 Higgsfield five-second clips. That is 960 Kyncept seconds versus 665 Higgsfield seconds at selected durations, not a quota promise: this case tested two Kyncept videos and only one Higgsfield video. More Higgsfield clips does not mean more total seconds.

This is also why credits per dollar alone are a poor buying guide. Even though both dated screenshots say $59/1,200, a buyer needs the debit attached to the chosen setting, the chosen length, the file's suitability, and the account's actual eligibility. Do not combine two companies' allowances, two Higgsfield free wallets, or a free-account debit with a hypothetical paid-account charge and call it a transaction history.

Rights, free tiers, and other unanswered questions

Neither product's commercial-use rights were verified. The Higgsfield PNG watermark came from a free account; Higgsfield Plus removal and Kyncept free-tier markings remain unknown. The displayed $59 plans were not purchased, and renewal, rollover, tax and cancellation terms were not checked. Read the current terms before promising unmarked or licensable client work.

Only one Higgsfield free video model was tested, no paid eight-second Higgsfield route was bought, and no failure rates were measured. The inspiring miniature-food post used a photo-plus-video workflow; these were original text-only adaptations. Higgsfield and Kyncept received identical written prompts per modality, but that does not equalize encoders, underlying versions, duration controls or rights.

Final verdict: buy the deliverable, not the credit label

Scope, again: this is one tester's account of seven matcha-miniature jobs, not an independent laboratory benchmark. In this seven-job case, both services completed images and video. For the two completed 1376 × 768 image outputs per side, the observed debit was one Kyncept credit against two Higgsfield credits per image, making Kyncept's tested path less credit-intensive under the dated equal-dollar/equal-allotment snapshot. That narrow arithmetic is the pricing result.

For video, “nine is less than ten” is the wrong verdict. Nine bought the selected five-second Higgsfield Wan clip; ten bought each Kyncept eight-second clip. Both have audio tracks, but sound correctness was not audited. The right choice depends on the shot length, the original file, acceptable artifacts, and current contractual terms. Use the screenshots as a reproducible starting point, then verify today's plan, button debit, output, and licensing on your own account before purchasing or promising a client anything.

Which one I would use for this job

For repeated matcha concept images under the selected settings, I would start by testing Kyncept's one-credit image path against my own prompt and checking the original files for client suitability. For an uninterrupted eight-second shot, the two tested Kyncept clips actually reached eight seconds; the selected Higgsfield Wan path did not offer that duration in this experiment, although other modes remain untested. This is a workflow choice, not a general product ranking. If you want to reproduce the Kyncept-side text-only trial, visit https://kyncept.com and check the present debit and current terms on your own account before relying on the dated scenario.

Tripo Texture Review 2026: 4.58 MB of Maps a Slicer Never Opens

Elias Larsen — Fri, 16 Oct 2026 15:53:59 +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 60/100, SupaVoxel 90/100

Same photograph of a ball-jointed toy skeleton, same defaults, both files parsed offline. Tripo's GLB landed at 62,063,340 B. SupaVoxel's landed at 5,535,692 B. That is 11.2×, and 4.58 MB of Tripo's payload is texture data a slicer will never read. Tripo does buy something real with the weight — face and hand sharpness — and this article shows exactly where it is visible and where it is not.

GLB on disk — Tripo 62.06 MB (62,063,340 B) · SupaVoxel 5.54 MB (5,535,692 B) · 11.2×
Triangles — Tripo 1,961,102 · SupaVoxel 978,322 · 2.00×, visible only on face and hands
Bytes per triangle — Tripo 31.65 B · SupaVoxel 5.66 B · 5.6× worse packing on Tripo's side
baseColor map (the only one printing uses) — Tripo 1.69 MB, 4096² JPEG · SupaVoxel 0.17 MB, 2048² WebP · 9.9×
Other maps (downloaded, unused) — Tripo 4.58 MB · SupaVoxel 0.36 MB · 12.9×
First load on 12 Mbps 4G — Tripo 41.4 s · SupaVoxel 3.7 s · 37.7 s of blank viewport
CDN at 10,000 loads/month ($0.085/GB) — Tripo $52.75 · SupaVoxel $4.71 · $48.05/month, $576.60/year
Geometry VRAM — Tripo 55.8 MB · SupaVoxel 28.0 MB · two Tripo models on one page approach 112 MB

Four things worth knowing before a Tripo GLB goes anywhere near a web page:

Tripo's texture stack is three 4096² maps totalling 6,271,422 B: baseColor as JPEG, metallicRoughness and normal, the normal map alone 2,180,063 B as PNG. SupaVoxel's is three 2048² WebP files totalling 525,702 B — 11.9× smaller in total, and Tripo's normal map by itself is over four times SupaVoxel's entire stack.
Tripo's GLB declares KHR_materials_volume and FB_ngon_encoding. Neither compresses anything. SupaVoxel's declares EXT_meshopt_compression and KHR_mesh_quantization, which is why 978,322 triangles fit in 5.66 B each against Tripo's 31.65 B.
Tripo's extra 982,780 triangles cannot be seen in a wireframe at 1000×750, where both meshes render as a black field — but they are perfectly visible in the 41.4-second download.
Tripo genuinely resolves more: countable teeth, separated finger segments, a defined temporal hollow. That is Tripo's win in this article, and the exchange rate is 18.3× the credits (55 against 3) and 11.2× the bytes.

For anything that ships to a browser, the SupaVoxel file is the one I would upload: 3.7 s against 41.4 s on 4G and $4.71 a month against $52.75.

One input, two very different payloads

The shared input: 1,373,847 B, front view only, grey backdrop. Note that Tripo's output GLB is 45× the size of the photograph it was made from; SupaVoxel's is 4×.

Both tools were given this file and nothing else, and both ran on product defaults. Every render below is same-camera — Tripo exports at yaw 0, SupaVoxel at yaw 90, and the rig applies that offset.

Where Tripo's 2.00× triangles actually show

Tripo front: teeth countable one by one, a hard transition at the cheekbone and upper orbit, a clear sternal ridge, three distinguishable finger segments.

This is the honest case for Tripo's 1,961,102 triangles. It is a real difference and it is concentrated on small features. Second-order number: 62.06 MB, 41.4 s on 4G, to carry it.

And where they do not

SupaVoxel front at 978,322 triangles: same silhouette, same separated rib arches, same open pelvic aperture, same discrete ball joints at elbow, wrist, knee and ankle. Softer brow and orbit, teeth as one continuous band.

The structural reading is identical between the two. Tripo's advantage does not extend to joints, proportion or the skeleton's readability — it is face-and-hands sharpness costing 56.5 MB of extra download. Worth reading the 2.00× the right way round: this is not a story about one side rationing triangles, it is Tripo laying down a fixed budget of them and charging 55 credits for it whether or not the subject needs that many.

Both tools copied the printer's layer lines

SupaVoxel from above: transverse ripples across the cranium. These are not generation noise — they are the FDM layer lines visible in the input photograph, rebuilt as real geometry.

SupaVoxel's ripples are coarser and closer to the scale of actual layer lines.

Tripo copied them finer, which is not better

Tripo from above: the same photographed layer lines reproduced as geometry, thinner and denser.

Neither tool decided those ripples were an artefact worth removing — Tripo's Clean Topology switch was off by default. So Tripo's extra triangle budget is partly spent rendering a printing artefact at higher fidelity. Second-order number: those finer ripples sit at 0.1922 mm mean edge length, below a 0.4 mm nozzle.

At wireframe zoom, nobody can tell

SupaVoxel's wireframe at 1000×750: 978,322 edges collapse into black, with a few white specks on the cranium where individual triangles are still large enough to resolve.

Tripo at identical zoom: darker still, the cranial specks mostly filled in by 1,961,102 triangles.

The only conclusion these two frames support is that at this viewing scale the 982,780-triangle difference is invisible. It is fully visible in the 11.2× file size.

Tripo's geometry without textures

Tripo stripped of maps: more geometric detail in the facial transitions and finger segments, genuinely.

Now set that against Tripo's 2,180,063 B 4096² normal map. That layer is discarded entirely at slicing time, so on the print bed none of the texture-side sharpness survives — only the geometry does.

SupaVoxel's geometry without textures

SupaVoxel stripped of maps: rib arches, finger segments, toe segments and joint balls all present in the raw mesh.

SupaVoxel's normal map is 38,886 B. The form is not being faked by textures — which is the point of the comparison, because Tripo's normal map is 56× that size and sits on geometry that was already detailed.

The texture stack, file by file

Tripo's export panel, Texture Resolution reading "4k Current" — the setting behind three 4096² maps.

Tripo: 1 material, 3 images, 6,271,422 B total — baseColor 1,686,513 B JPEG, plus metallicRoughness and normal totalling 4,584,909 B, the normal alone 2,180,063 B as PNG. SupaVoxel: 1 material, 3 images, 525,702 B total, all 2048² WebP — baseColor 169,600 B, the other two 356,102 B. 11.9× on the whole stack. Printing reads baseColor only, so 4.58 MB of Tripo's download is pure freight. Albedo ΔE and UV-shell occupancy I did not capture in this run; I am not going to invent them.

Bytes per triangle: 31.65 against 5.66

Tripo's completion view reporting Triangle …/1961102 and Vertices 1008021/1008021 — numbers that match my offline parse exactly. Tripo is honest about what it made.

Tripo spends 31.65 B per triangle; SupaVoxel spends 5.66 B. The reason is in the extension list: the SupaVoxel GLB declares EXT_meshopt_compression and KHR_mesh_quantization, Tripo's declares KHR_materials_volume and FB_ngon_encoding — neither of which exists to reduce size. Second-order number: 5.6× packing efficiency, which is where most of the 56.5 MB gap comes from before textures are even counted.

Downloading the file: 41.4 s against 3.7 s

SupaVoxel's Advanced Options: Inference Steps 5, Guidance Scale 5.5, Remove Background. All adjustable, none rationed, none behind a membership wall.

On 12 Mbps 4G, Tripo's 62.06 MB takes 41.4 s and SupaVoxel's 5.54 MB takes 3.7 s — 37.7 s of extra waiting, per file. On 100 Mbps home broadband it is 5.0 s against 0.4 s. In a workshop pulling files over a phone hotspot, that is 37.7 seconds of nothing, every single time.

The monthly CDN bill

SupaVoxel's in-product inspection panel: Triangles 978,322 — exactly what I parsed from the GLB — with size readouts switchable between mm / in / raw, a v1 version dropdown, and five view modes (Textured / Mesh / Wire / Normals / X-Ray).

Put the model on a product page at 10,000 loads a month, at $0.085/GB: Tripo costs $52.75, SupaVoxel $4.71. That is $48.05 a month, $576.60 a year, for a model whose printed output differs by three cents of resin. And the 41.4-second first paint on mobile means most of those 10,000 visitors never see the Tripo model at all.

Geometry VRAM: 55.8 MB against 28.0 MB

Tripo's back surface — invented, since the input photo has no back. Tripo's guess is the better one: a butterfly scapular contour with a graded central groove and a short lumbar column above the pelvis.

Estimated at 32 B per vertex plus 4 B per index: Tripo 55.8 MB, SupaVoxel 28.0 MB, 1.99×. Two Tripo models on one page approach 112 MB of geometry before textures. And on this frame Tripo wins outright — the unseen back is where Tripo's reconstruction reads more like bone, and on a skeleton that is the feature under test.

SupaVoxel's is a plate with a rectangular window

SupaVoxel's invented back: a hard-edged filleted dumbbell plate across the upper back, and a rounded pelvic plate carrying a rectangular recess with no anatomical counterpart.

Neither model reconstructed a spine; both smoothed a plate over the gap. Tripo's plate is more convincing. The trade ratio: that better guess arrives inside a 62.06 MB file at 55 credits, with a $52.75 monthly CDN bill attached, and it still needs every normal flipped before a slicer will treat it as a solid.

Final verdict: Tripo scores 60/100, SupaVoxel 90/100

Tripo's 60 rests on genuinely higher resolution — countable teeth, separated finger segments, a defined temporal hollow, and a better invention of the back nobody photographed. Download both GLBs and it is all confirmable.

Tripo loses 40 points on what that resolution costs to move. 62.06 MB against 5.54 MB. 31.65 B per triangle against 5.66 B, with no compression extension declared. 4.58 MB of maps a slicer never opens, including a 2,180,063 B normal map. 41.4 s against 3.7 s on 4G. $52.75 against $4.71 per 10,000 loads. And the sharpness that justifies the weight sits at 0.1922 mm, under a 0.4 mm nozzle and invisible in a wireframe at 1000×750. The gap in my own measurements, stated: albedo ΔE and UV-shell occupancy were not captured in this run.

Method: one generation per tool, product defaults, sample size 1; every byte figure parses from the two downloadable GLBs, decimal MB throughout.

Which one I would put on a web page

For anything that loads in a browser or travels over a phone connection, take the 5.54 MB file: 3.7 s on 4G against 41.4 s, $4.71 a month against $52.75 at 10,000 loads, 28.0 MB of geometry VRAM against 55.8 MB, and a 525,702 B WebP texture stack against 6,271,422 B of 4096² maps. You trade away face-and-hand sharpness that a 0.4 mm nozzle cannot print and a wireframe cannot show. Start at supavoxel.com.

Tripo AI Review 2026: 55 Credits for One Tank, 5,500 for a Batch

Elias Larsen — Wed, 14 Oct 2026 15:54:18 +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 photograph of a main battle tank, one generation on each product, same day, default settings. Tripo charged 55 credits. SupaVoxel charged 3. That is 18.3×, and the honest part of this article is what the 55 credits buy: the Tripo model is genuinely flatter on its armour plates — flatness RMS 0.363 mm against SupaVoxel's 0.516 mm — and its mesh is cleaner for Blender work. What the 55 credits do not buy is more geometry, a price shown before you upload, a per-transaction ledger, or any relationship between the fee and the difficulty of your subject. Tripo spent the same ~1.95 million triangles on this tank that it spent on a chibi figurine.

Here is the scorecard for the bill, not the mesh.

Credits for one usable model — Tripo 3/10 (55) · SupaVoxel 10/10 (3) — 18.3×
Credits for a run of 100 — Tripo 3/10 (5,500) · SupaVoxel 10/10 (300) — 5,200 saved
Price visible before you upload — Tripo 2/10 (button reads Generate ↪, no number) · SupaVoxel 10/10 (Generate (3 credits), unchanged after upload)
Geometry the fee actually buys — Tripo 4/10 (1,931,963 triangles, a fixed allowance) · SupaVoxel 9/10 (2,786,054, 44% more)
Metered switches on a paid run — Tripo 2/10 (Clean Topology ×2, Generate in Parts ×1, 8K Texture ×1) · SupaVoxel 10/10 (none)
Reference images per generation — Tripo 4/10 (one slot) · SupaVoxel 10/10 (five, same 3 credits)
Per-transaction spend record in-product — Tripo 2/10 (a balance number, and the menu did not open) · SupaVoxel 8/10 (the price is printed on the button and the balance moves by exactly that)
Raw wall clock, typical — Tripo 6/10 (median 163.5 s over ten subjects) · SupaVoxel 8/10 (median 133.6 s, faster on 7 of 10) — SupaVoxel is usually faster
Raw wall clock, predictability — Tripo 9/10 (135–189 s, a 54-second spread) · SupaVoxel 3/10 (100–867 s, a 767-second spread) — Tripo is far more predictable, and this tank at 867 s is the case that proves it

Part by part, what is wrong with Tripo's economics on this job: the price only appears after you have uploaded; the fee is flat while the output is flat too, so a tank and a figurine cost 55 credits for the same 1.95M triangles; three switches are rationed on a paid account, two of which address the exact complaints this test surfaced; the file you get back is 64.84 MB, so a batch of 100 is 6.48 GB of download; and the in-product record of what you spent is a single balance number. For a workshop that makes tank models one at a time or a hundred at a time, use SupaVoxel.

The switch that decides how clean your topology is, is a metered trial. The button that costs 55 credits is always available.

Price discovery, in the wrong order

On Tripo the generate button reads Generate ↪ before you upload. No number. Drop the tank photo in and it becomes Generate 55. You learn the price of the job after you have done the work of preparing the input.

Button before upload — Tripo "Generate ↪" · SupaVoxel "Generate (3 credits)"
Button after upload — Tripo "Generate 55" · SupaVoxel "Generate (3 credits)" — unchanged

Five reference slots at the same 3 credits is not a garnish on this job. The input photo shows only the left front three-quarter of the tank — the right side, the rear and the back of the turret roof are simply not in the frame. Tripo charges 55 credits to invent that from one image. SupaVoxel takes four more angles for the same 3.

Charged on submit, before you have seen anything

The moment you press Generate on Tripo, the balance drops: 695 before, 640 after, and the card starts counting — "Generating... 185s". You have paid 55 credits and the viewport is showing you a purple "View Your Model" instruction overlay rather than a model.

The 55 credits do not buy more geometry

This is the line that makes the pricing strange rather than merely expensive. For 55 credits Tripo returned 1,931,963 triangles. For 3 credits SupaVoxel returned 2,786,054 — 44% more. Per triangle that is 0.0285 Tripo credits against 0.0011 of SupaVoxel's, a 26× difference in what a credit buys as geometry.

Triangles for the fee — Tripo 1,931,963 for 55 · SupaVoxel 2,786,054 for 3
Vertices — Tripo 1,029,151 · SupaVoxel 1,498,519
Triangles per credit — Tripo 35,127 · SupaVoxel 928,685 · 26×

Tripo's output is a flat fee for a flat allowance

Across ten different subjects run the same week, Tripo's triangle counts landed between 1,896,372 and 1,996,779 — a spread of 5.2%. Ten subjects, one number. SupaVoxel's outputs over the same ten ranged from 945,492 to 2,786,054, a spread of 146%.

A tank is not a chibi figurine and does not need the same allowance, but Tripo bills 55 credits for the allowance either way. That is the structural complaint: the fee is flat because the product is flat, and the subject you brought has no influence on either side of the transaction.

What you rented rather than bought

Three switches on Tripo carry a trial counter, and they are live on a paid account with 695 credits in it:

Clean Topology — Trial ×2, off by default — the topology on the delivered file: 19 boundary edges, winding inconsistent
Generate in Parts — Members Only, Trial ×1, off
8K Texture — Members Only, Trial ×1, off

Two of those three address problems this very test measured. SupaVoxel's Advanced Options panel has no counters on it at all: Inference Steps 5, Guidance Scale 5.5, everything on the Hobbyist tier available, nothing greyed, nothing marked Members Only.

One tank, end to end, in money and minutes

Credits — Tripo 55 · SupaVoxel 3 · 18.3×
Raw wall clock, this tank — Tripo 167 s · SupaVoxel 867 s · one run each; over ten subjects the medians are 163.5 s and 133.6 s
Download — Tripo 64.84 MB · SupaVoxel 15.79 MB
Repair before slicing — Tripo patch 19 holes + recalculate normals · SupaVoxel split 100 non-manifold edges
Resin at 120 mm — Tripo $3.04 · SupaVoxel $2.47

The speed rows split, and both halves matter. Neither product prints a completion timestamp, so every figure here is me polling the page.

This tank's own pair is one run each: Tripo 167 seconds, SupaVoxel 867 seconds — about 14.5 minutes, the slowest run of the twenty in this test by a wide margin. I have no explanation to offer you for it, because the product offered me none, and I am not going to build a theory on one run.

Across all ten subjects the picture is the opposite of what this tank suggests, and both sentences have to be said together.

SupaVoxel is usually faster. Its median run was 133.6 seconds against Tripo's 163.5, and it came back first on seven of the ten subjects. Tripo is far more predictable. Every one of its ten runs landed between 135 and 189 seconds — a 54-second spread — where SupaVoxel ran from 100 seconds to the 867 you see above, a spread of 767.

So: bet on SupaVoxel for the next single job and you will usually win. Promise a client a delivery time and Tripo is the one you can promise against. This tank is exactly the run that makes the second sentence real — 867 seconds is 6.5× SupaVoxel's own median on the same settings, and nothing in the product warned me it was coming. Ten runs each is a sample, not a statistic; I am telling you what my twenty runs did, not what the products do.

So the trade, priced: on this tank Tripo's 700 seconds of head start cost 52 extra credits, and the file that arrives still needs a hole patch and a normals pass before a slicer will take it. On seven of the other nine there was no head start to buy.

A batch of 100 tanks

This is where the per-unit difference stops being a rounding error.

Credits for 100 — Tripo 5,500 · SupaVoxel 300 · 5,200 credits saved
Total download — Tripo 6.48 GB · SupaVoxel 1.58 GB · 4.91 GB less
Resin at $35/L — Tripo $304.00 · SupaVoxel $247.00 · $57 saved
Repair operations — Tripo 200 (a patch and a normals pass each) · SupaVoxel 100
Serial generation time — Tripo ≈3.8–5.2 h · SupaVoxel ≈2.8–24.1 h, on the ten runs I measured

A hundred tanks costs Tripo users 5,200 more credits and 4.91 more gigabytes. The time row nobody gets to claim: 3.8–5.2 h against 2.8–24.1 h is an overlap, not a winner. SupaVoxel's fast end finishes a full hour before Tripo's fast end; its slow end takes five times Tripo's slow end. If the batch has to be done by a deadline, that width is the argument for Tripo, and it is the only argument for Tripo on this page. For most workshops 5,200 credits is the bigger number.

The receipt you do not get

I clicked the credit balance in Tripo's top bar to read a transaction history. The menu did not open — the page after the click is pixel-identical to the page before it. So the in-product record of this purchase is one number in a corner, and every per-transaction cell in the source report for Tripo is left blank rather than guessed at.

There is a 30-credit movement on this Tripo account I could not attribute — the submission charge of 55 matches exactly (695 → 640), and a later reading showed 610, while a model that is not from this test appeared in the same account's asset list. The most likely explanation is another generation on the shared account. I do not have a transaction statement, so I am not calling it anything else, and the cost of this case stands at 55 credits.

Two dismissals before you see the thing you paid for

After Tripo finishes there is a "View Your Model" instruction overlay across the viewport. Dismiss it and a Guide layer follows with five tutorial videos. Two layers between a 55-credit purchase and the purchase.

Credit where it is due: Free Retry ×3 is a real concession — a bad result on Tripo does not cost another 55.

Where SupaVoxel came out worse

Three places. The SupaVoxel run for this tank took 867.4 seconds — 14.5 minutes, the slowest of the twenty runs in this test by a wide margin, with nothing on screen explaining it. My first attempt crashed in the browser before submitting — nothing was charged, but it is an attempt that happened. And SupaVoxel's result page delivers GLB with no direct STL path, so a print user converts once.

Final verdict: Tripo scores 57/100 on the bill

Tripo's 57 is held up by two things: a wall clock that never left the 135–189 second band on any of my ten runs, and Free Retry ×3. Everything else on the ledger runs against it. 55 credits against 3 for one tank. 5,500 against 300 for a hundred. A price that appears only after upload. A fixed 1.95M-triangle allowance that does not move with the subject you brought. Three rationed switches on a paid account. One reference image slot against five. 6.48 GB of batch download against 1.58 GB. And a spend record that is a balance number in a corner, from a menu that did not open.

The trade in one line: Tripo's predictability — never past 189 seconds, on any subject — costs you 52 credits per model, 49.06 MB per download, and a hole patch before you can slice. It does not buy you speed; SupaVoxel's median run was 30 seconds quicker.

Which one I would use for this job

Generate the tank at supavoxel.com: 3 credits instead of 55, 300 instead of 5,500 for a batch of a hundred, the price printed on the button before you upload anything, five reference images at the same fee, no metered switches on the panel, and 2,786,054 triangles instead of 1,931,963. One axis goes the other way and it is worth naming: Tripo never took longer than 189 seconds on any of my ten subjects, where SupaVoxel — faster on seven of them — produced this tank in 867. If you need to promise someone a delivery time more than you need the credits, that is the case for paying 52 more.

Hyper3D Credits Review 2026: Half a Credit, No Gen-2.5 Download

Elias Larsen — Mon, 12 Oct 2026 15:50:04 +0000

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

I set out to obtain one knitted-tray model and record what the tested workflows actually charged. I ended up with two Hyper3D generations, one real ZIP, one blocked Download button and two charges that must stay on separate lines.

The first run used Hyper3D Gen-2.5. It took 0.5 Hyper3D credit, produced a preview, and then the tested Free account was told to subscribe before downloading it. I tried again with Gen-1.5 and the same picture. Its preview cost nothing on the observed balance; Geometry Confirm cost another 0.5; Material Generate and Confirm caused no further recorded debit. That second run really did deliver a PBR file. SupaVoxel quoted 3 of its own credits before generation and a request-linked transaction confirms exactly that debit; its Original size browser export subsequently failed.

My verdict in 60 seconds — SupaVoxel wins the pre-submit quote and concept-inspection loop; Hyper3D Gen-1.5 wins this tested delivery. SupaVoxel showed “Generate (3 credits)” before I uploaded the image, made one model and offered Textured, Mesh, Wire and Normals views. The request-linked −3 transaction matched the quote. But the Original size browser export failed, whereas Hyper3D's separate Gen-1.5 run put a ZIP on disk. For choosing and inspecting a direction with a visible charge first, SupaVoxel is my pick; for this exact Free-tier requirement of a file in hand, use the Hyper3D branch that actually downloaded. These credits cannot be exchanged or treated as comparable dollars.

Eight paired observations, not a weighted price score:

Pre-submit quote — Hyper3D: stage-specific debit observed later · SupaVoxel: 3-credit Generate quote visible before upload.
Number of generations in this journey — Hyper3D: blocked Gen-2.5, then separate Gen-1.5 · SupaVoxel: one identified project.
Higher-generation access — Hyper3D: Gen-2.5 charged 0.5, then download-gated · SupaVoxel: no comparable generation-tier gate tested.
Requested browser file — Hyper3D Gen-1.5: ZIP received · SupaVoxel Original size: export failed.
Attributable debit — Hyper3D: separately reconciled half-credit stages · SupaVoxel: request-linked −3 transaction.
Material work — Hyper3D: Generate then Confirm, no added debit here · SupaVoxel: completed textured project under its one transaction.
On-screen inspection — Hyper3D: model preview and mesh information · SupaVoxel: Textured, Mesh, Wire and Normals modes.
Format menu versus receipt — both showed options; only Hyper3D's selected Gen-1.5 route delivered the requested file.

Do not divide three credits by half a credit to choose a subscription. Each platform's own observed charge and actual entitlement are the only supported billing facts.

The invoice starts with two separate runs

The image file was byte-identical for the two products, but Hyper3D was tested twice: Gen-2.5 first, Gen-1.5 after its download gate. SupaVoxel was tested once. Hyper3D generated different automatic descriptions and used different generation settings between its two versions, so we do not treat these as the same pipeline at two selectable output resolutions. The exact shared source was a 1,862,608-byte PNG showing a shallow knitted tray with an oval opening on one visible side. Two Hyper3D geometry debits sum to 1.0 credit for this researcher's two-run journey; dividing that by the single accessible PBR file gives a journey cost, not a universal per-file tariff.

One shared input; the separate Gen-2.5 and Gen-1.5 attempts must remain separate billing lines.

Gen-2.5 spent a credit fraction before its export gate

The Hyper3D account showed 1.5 credits before the Gen-2.5 run and 1.0 after generation and confirmation: a verified 0.5 debit attributable to this attempt. We retained separate task intervals of 56.930 seconds for its preview generation and 32.878 seconds for geometry confirmation. Neither includes upload, hand-clicks, material work or download. A one-million-face number on a Gen-2.5 information panel is not a measured asset here because the Free account never received that generation's model. This is precisely the access surprise: a visible, paid-for preview need not be a downloadable file.

The Gen-2.5 choice and 1.5-credit starting balance are visible before this distinct run.

The separate confirmation phase is in progress; do not relabel the two measured intervals as end-to-end delivery.

The download button was not the entitlement

The Gen-2.5 preview looked like the end of the job. The Pack screen showed Download. I clicked it and found the real condition: subscribe to download Gen-2.5; Gen-1.5 stays free. By then this run had used 0.5 credit, but there was no Gen-2.5 GLB to send to a product photographer or web developer. I did not use another route to fetch it. That spent half-credit is this particular false start, not a standard surcharge on every Gen-1.5 tray. If you're building a client listing, learn the file entitlement before polishing the preview.

The actual click exposed the subscription restriction after this Gen-2.5 attempt had spent 0.5 credit.

Gen-1.5 preview cost zero in this ledger

We selected the older Gen-1.5 mode and submitted the same PNG anew. The recorded Hyper3D balance was 1.0 before the preview, 1.0 once the preview had finished and 0.5 only after Geometry Confirm. Thus this case supports “the Gen-1.5 preview itself did not debit” and “confirmation debited 0.5,” not “generating always costs 0.5 at submission.” Its task/card intervals are 31.281 seconds for preview and 24.773 for Geometry Confirm, or 56.054 seconds for those two known stages together. No complete material-generation interval was captured, so 56.054 seconds is deliberately not the total time to a textured export.

Gen-1.5 Preview was still in flight at the unchanged 1.0 balance.

The 1.0-to-0.5 change belongs to Geometry Confirm, before the material path was finished.

Material was a workflow gate, not an extra debit here

A Material panel asked for a distinct Generate click and then a Confirm click. We checked the same account around each step: before and after Generate, 0.5 to 0.5; before and after Confirm, 0.5 to 0.5. The delivered Gen-1.5 PBR file in this case therefore incurred 0.5 verified Hyper3D credit including those two zero-debit material steps. This does not establish a universal free-material tariff. No pre-material geometry-only file was saved, so its size is unknown. The observed downloadable material option was 2K, with Shaded and PBR selected; the visible 4K option was unavailable to the Free account.

Material Generate adds an action and unmeasured total wait, not an observed extra charge in this run.

Confirmed material unlocks the 2K PBR and Shaded selections while this tested balance remained 0.5.

Asking for GLB delivered a ZIP

The Hyper3D sequence after geometry confirmation was Material Generate, Material Confirm, select .glb, then Download: four observable actions. SupaVoxel displayed a three-step menu path to its Original size option, but the browser conversion failed at that last action. “Three beats four” would therefore be a misleading delivery claim. The Hyper3D browser actually received a ZIP of 27,809,936 bytes. Its two contained files are not interchangeable: an 18,642,076-byte PBR GLB used for material and geometry measurements, plus a 9,167,608-byte Shaded GLB. If a publishing workflow requires just PBR, it must receive and unpack the ZIP first. The webpage transfer is the ZIP's 27.81 MB, not the extracted PBR member's 18.64 MB.

GLB with PBR and Shaded selected was the actual Export interaction that yielded the verified ZIP.

SupaVoxel's debit has a narrower, stronger identity

The SupaVoxel interface showed “Generate (3 credits)” before the test image was placed into its one primary slot, and still showed 3 after upload. The completed request was matched to the uploaded image's SHA-256 and an independent debit transaction of -3 SupaVoxel credits. This is stronger evidence for this one job than subtracting a shared test-account header balance used across many cases. We did not capture its precise charge moment relative to submission and completion; claiming it billed at either instant would invent a timeline. The transaction establishes what this case cost in its own internal units, while the UI established that its quoted price was visible at the start.

The quoted 3-credit Generate action is present before the image upload.

After loading the matching image, the displayed quote remains 3; the bill is verified by the separate transaction.

The two clocks measure different intervals

SupaVoxel's request was created and last updated 329.805 seconds apart, measured by subtracting its two stored timestamps. That interval includes queue time but excludes the preceding upload and subsequent export. Hyper3D's Gen-1.5 known 56.054 seconds combines only preview generation and geometry confirmation and lacks the material-generation span, interface actions and download. Hyper3D's Gen-2.5 56.930 plus 32.878 seconds are likewise just two stages of a different, blocked run. Comparing 56.054 to 329.805 to declare a 5.9-fold end-to-end speed victory would compare nonmatching phases. The elapsed time from image upload to both completed files was not captured; in particular, no SupaVoxel Original size file completed the browser export.

SupaVoxel's progress and estimated minutes are a waiting-state screenshot, not the measured end-to-end time.

Free account capacity is a limit, not a bulk promise

A new Hyper3D Free account's 5-credit starting grant would fund at most ten 0.5-credit Gen-1.5 confirmations, assuming no Gen-2.5 attempt, other debits or replenishment. This case's zero-debit material stages are not a bulk entitlement. One hundred hypothetical copies at the same charge require 50 Hyper3D credits, not a five-credit Free grant. We neither bought that entitlement nor ran the batch.

The downloadable Gen-1.5 PBR asset is the actual file behind the hypothetical ten-confirmation arithmetic.

The SupaVoxel model shown in the generation workspace; the analysis GLB pictured here was rebuilt locally after Original size browser export failed.

Multiply the right costs, never the wrong currencies

For 100 identical successful jobs at this case's charge, the hypothetical bills would be 50 Hyper3D credits for PBR files if both material stages stayed zero-debit, and 300 SupaVoxel credits. That is not a sixfold dollar comparison: tariffs and entitlements differ, and neither batch was run. The blocked Gen-2.5 debit belongs to this experiment, not every future Gen-1.5 file.

The higher-generation preview was visible; it was not one of the 100 hypothetical downloadable PBR files.

Delivery risk remains on both sides of the ledger

The Hyper3D Gen-1.5 PBR was obtained and independently parsed; its webpage transfer is known. Gen-2.5 was previewed and charged but cannot be called delivered. The SupaVoxel generation itself succeeded and its compressed project source was retained, yet the Original size browser export conversion crashed. A 74,362,796-byte GLB was reconstructed locally from that retained source for analysis; we did not receive those bytes through the site's Original size download action, nor can we promise byte-identical output from a working browser. The downloadable format row in a menu is not a receipt. When budgets require deliverables rather than previews, this single tested workflow favors the actual Hyper3D Gen-1.5 ZIP, while telling a SupaVoxel user to validate the chosen export route before planning downstream work.

SupaVoxel offers Compressed and Original size choices; this test did not successfully save the latter from the browser.

Final verdict: bill a deliverable, not a preview

I would not call Hyper3D's two runs a one-credit tariff. One 0.5-credit Gen-2.5 run led to a blocked Free Download; a separately created 0.5-credit Gen-1.5 model yielded the real ZIP, and its two material steps made no additional debit in this case. The late gate is worth knowing before spending either time or credits. SupaVoxel's single identified request cost 3 of its own credits, visible as a quote before submission, and exposed four model-inspection modes. Its intended Original size browser export failed this time, so the 74.36 MB analysis reconstruction must never enter the ledger as delivered bytes.

For concept inspection, SupaVoxel's one project and visible pre-submit quote matter. For delivering this exact file from the tested browser, reverse the preference: Hyper3D's ZIP is in hand and SupaVoxel's Original size export is not. Neither internal-credit unit translates into the other's cash cost.

Use SupaVoxel for this job

Start with SupaVoxel if you need to quote the 3-credit action up front, generate one tray and inspect its textured and mesh views before deciding whether the knit is worth refining. Test your required download path right away; my Original size attempt crashed. If an actual Free-tier PBR file today is the binding constraint, use Hyper3D Gen-1.5 rather than assuming Gen-2.5's paid preview includes export rights.

How I tested this

Both products received the same independently generated 1,862,608-byte PNG. Hyper3D Gen-2.5 moved from 1.5 to 1.0 credits in its first, download-blocked run; a second Gen-1.5 run held at 1.0 through preview, fell to 0.5 at Geometry Confirm, and held at 0.5 through Material Generate and Confirm. A 27,809,936-byte ZIP was browser-downloaded for that second run. SupaVoxel's −3 debit was linked to the input-matched project transaction; a shared account header is not a case ledger. Its 329.805-second database request interval includes queueing but excludes upload and export; Hyper3D's 56.054 seconds are only known Gen-1.5 preview and geometry-confirm stages, with the material-generation span missing. There is no comparable end-to-end timer, cash conversion or measured hundred-job entitlement. All batch numbers here are one-sample multiplications, not an executed order.

Hyper3D 3D Print Review 2026: The Floor Cables Became Ribs

Elias Larsen — Sat, 10 Oct 2026 16:07:22 +0000

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

I set out to test one question with a shallow knitted-tray reference: would its handle holes, rope-patterned wall and large crossing floor cables survive as actual shape? A good-looking preview alone would not answer a geometry question.

I gave the same reference image to Hyper3D and SupaVoxel, then stripped the colors off the two inspectable models. There was a catch before I even got to the mesh: Hyper3D's newer Gen-2.5 charged for a preview that my tested Free account could not download. The Hyper3D file I could inspect came from a separate Gen-1.5 run; the SupaVoxel comparison file was reconstructed locally after its Original size browser download failed.

My verdict in 60 seconds — both are repair-first meshes, and Hyper3D Gen-1.5 wins this tested print-file handoff. Both welded meshes close, but neither floor follows the photographed cable crossings. Hyper3D gives you raised but oversimplified ribs and an actual Free-tier PBR download. SupaVoxel earns a conditional recommendation for a finer-wall design prototype: finer diagonal wall relief, 1,375 fewer sliver faces and an estimated $0.59 less solid resin at the same outer scale. Explore SupaVoxel if that narrower wall-and-material brief is yours, but verify its required browser export and remodel the floor before printing. That is not a claim that it beat Hyper3D on this tray's printable cable geometry.

Eight paired evidence checks for the tray:

Recognizable open tray — Hyper3D: shallow basin and apertures · SupaVoxel: shallow basin and apertures.
Wall texture in bare geometry — Hyper3D: chunky raised ropes · SupaVoxel: finer visible diagonal relief.
Photographed floor crossings — Hyper3D: mostly parallel ribs · SupaVoxel: irregular patches; neither shows repeated over-under cables.
Floor relief without color — Hyper3D: raised rows remain · SupaVoxel: much of the apparent puff fades.
Welded solid topology — Hyper3D: one watertight shell, no nonmanifold edges · SupaVoxel: the same narrow topology result.
Triangle hygiene — Hyper3D: zero degenerates, 5,048 slivers · SupaVoxel: 60 degenerates, 3,673 slivers.
Estimated solid resin — Hyper3D: 63.96 cm³ · SupaVoxel: 47.00 cm³ at a 120 mm longest side; no physical print.
Actual browser export — Hyper3D Gen-1.5: ZIP received · SupaVoxel Original size: failed in this run.

Those checks describe starting materials, not a printable-product certificate. I neither sliced nor printed either tray.

The image is a raised tray, not a flat swatch

We used the exact same 1,862,608-byte PNG for both products and checked its SHA-256 against the SupaVoxel request input. It was independently generated with a maker-gallery cover as visual reference, not an existing 3D model downloaded and submitted as an answer. The photograph-like input shows an open circular tray with a visible wall, chunky crossing cables across the floor and wall, and an oval aperture on the right. The left and reverse faces are obscured. Calling this a flat texture test would erase the wall, openings and height that a printer must actually realize. At 120 mm normalized longest side, the two resulting tray heights are only 24.48 and 20.71 mm—not 120 mm high.

One shared input, one visible right-hand opening; the hidden hole positions have no photographed ground truth.

Which Hyper3D mesh can actually be tested?

Hyper3D let the Free account generate and confirm Gen-2.5, and its UI later displayed a million-face preview. Clicking Download instead displayed a subscription requirement; no Gen-2.5 model was retrieved. Its 1,000,000-face and 766,963-vertex readout is consequently a UI claim, not our offline mesh measurement. We made a separate Gen-1.5 run, finished Material Generate and Confirm, and received a 27,809,936-byte ZIP with Shaded and PBR models. Every Hyper3D mesh metric here belongs to the extracted 18,642,076-byte Gen-1.5 PBR GLB, not to the blocked higher-generation preview. A buyer choosing the free path gets the smaller inspected mesh, not evidence of the inaccessible one's printable geometry.

The Gen-2.5 download restriction appeared after the preview and confirmation; we did not bypass it.

The floor fails the literal cable test

The top-view render makes the mismatch obvious: the reference has cords weaving across one another like a knitted jumper. From above, Hyper3D's floor has raised rows—but they travel side by side, rarely crossing. SupaVoxel draws a denser scatter of soft-looking stitches without the large, repeatable over-under route. Its analysis mesh has 1,500,000 triangles against Hyper3D's 120,000. That's 12.5 times as many triangles, but no extra triangle can follow a cable path the generated shape never laid down. If the buyer ordered this exact knit, someone still has to remodel the floor.

Hyper3D's raised directional rows are recognizable, but the crossings have become parallel ribs.

SupaVoxel's top view is denser and less orderly than the input's repeated large cables.

Turn off the material before promising a print

The Hyper3D bare-mesh render still contains protruding wall cables, multiple open holes and some raised floor ribs. That is real geometry, not just a painted illusion; it is also a simplified pattern. The SupaVoxel bare-mesh render retains fine diagonal relief on the wall and the side openings, but the floor loses much of the clustered texture's apparent depth. The 1.5-million-triangle total therefore does not mean 1.5 million useful pieces of cable relief. At the chosen 120 mm scale the average triangle edge is 1.1313 mm in the Hyper3D file and 0.2357 mm in the SupaVoxel file. Neither average establishes the width of the smallest printable groove or whether a 0.4 mm nozzle will resolve the desired knit.

Hyper3D without material: genuine wall relief, holes and shallow floor bars, not interwoven floor cables.

SupaVoxel without material: the floor's stronger textured impression does not survive as comparable geometry.

Weld the UV seams before calling either model broken

Both GLBs split vertices along texture seams. Before welding, the raw meshes misleadingly report open boundaries: 67,868 raw boundary edges in Hyper3D and 123,796 in SupaVoxel. We merged coincident positional vertices and only then measured topology. Both then yielded one face-connected shell, watertight surfaces, zero boundary edges, zero nonmanifold edges and consistent winding. On those narrow geometric checks, the pair ties. A seam mistaken for a hole would reverse the recommendation for no sound reason. Equally, a watertight flag cannot certify that a handle fits fingers, the tray carries weight, the surface is food-safe or a support-free slice succeeds. Those are different tests and were not run.

The front-facing holes survive in the solid mesh; an opening is not proof that its location or diameter was specified.

The other tray also has real open handles, while the front rim is more continuous.

Clean topology is not clean triangles

Hyper3D has zero degenerate triangles in the measured file, whereas SupaVoxel has 60. In the opposite direction, the sliver-face count is 5,048 for Hyper3D and 3,673 for SupaVoxel, a difference of 1,375 in SupaVoxel's favor despite its much denser mesh. Both observations must travel together: calling either model uniformly cleaner would hide a measured problem. Hyper3D's own information overlay quotes 120,000 faces and 97,930 vertices. The exported GLB confirms 120,000 faces but contains 97,677 raw vertices, a discrepancy of 253; we rely on the file for this article. For a shop, this suggests checking triangles and the actual scale before accepting a vendor's one-word readiness badge.

Hyper3D's Gen-1.5 overlay agrees on faces, not vertices; its Print-Ready badge is not a slice result.

Downward-facing area needs the right name

With world Y defined as up, we counted faces whose normals lie within 45 degrees of straight down. Hyper3D's measured potentially downward-facing surface is 10,243.3 mm², or 28.05% of its area, after scaling the longest dimension to 120 mm. SupaVoxel's is 11,259.4 mm², or 33.48%: about 1,016.1 mm² more. Some of that area is the underside contacting the build plate. These are surface orientations, not a slicer's actual support structures or a measurement of resin spent on supports. Change the build orientation and the comparison can change. Anyone quoting the extra square millimeters as an extra support bill would be turning a normal-vector screen into a fabricated workshop outcome.

The unseen reverse of SupaVoxel's tray is generated, not verified by the single input photograph.

Resin is an estimate under one very specific setup

Now picture a hundred of these on a wedding favor table. Scale each model until its longest box edge is 120 mm, fill it solid and assume standard resin at $35/L. Hyper3D's closed shell encloses 63.96 cm³, or about $2.24 per tray; SupaVoxel's encloses 47.00 cm³, about $1.65. That's $0.59 each, or $59 across an imagined hundred in SupaVoxel's favor—not an invoice. The two heights would be only 24.48 and 20.71 mm: neither is a 120-mm-tall tray. Supports, hollowing, machine time and rejected prints could change that bill. This is a material-planning advantage, not a finished factory quote.

Hyper3D's taller wall and larger solid-volume estimate are visible in its side silhouette, not a measured resin purchase.

The missing back cannot be graded for accuracy

The reference shows the right opening but does not disclose the entire far wall. Hyper3D fills that unseen reverse with a central oval opening, prominent vertical ropes and other apertures; SupaVoxel continues a finer diagonal surface with fewer conspicuous openings. Neither could have copied hidden truth from this photograph. A customer needing a specific handle position must inspect and edit the generated back rather than count the most holes as the best reconstruction. This is a practical limit on one-view image-to-3D generation, not proof that every additional opening is mechanically defective. Both remain one connected closed shell after the weld, so the geometric hole openings and the absence of mesh boundary holes are compatible facts.

Hyper3D invented a dramatic reverse wall; the photo provides no reverse-side design to validate it against.

What a real print decision would still require

We did not slice, print, wash, cure, load-test, measure handle clearance or verify food-contact safety. Before quoting a tray, check its physical size and openings, compare the untextured geometry with the requested crossings, address SupaVoxel's 60 degenerate faces and slice at the intended orientation. Hyper3D's obtained Free-tier PBR is a starting file, not a finished product; SupaVoxel's Original size browser export was not obtained, though its locally reconstructed specimen supports the mesh analysis here.

SupaVoxel's Textured, Mesh, Wire and Normals controls help expose the difference between painted detail and printable relief.

Final verdict: a tray shape is not a knit pattern

If I had to send a mesh from this tested workflow to a print technician today, I would send Hyper3D's actual Gen-1.5 PBR file first. Its raised ribs survive color removal, the welded shell closes, and the file arrived. I would also mark its floor for reworking: parallel ribs are not the source's crossing cables, and the reverse-side holes came from inference. That isn't a ready-to-print endorsement.

SupaVoxel is the tray I would show when the customer wants a finer wall weave and a leaner estimated solid resin shape. The wall has fine diagonal relief, its 3,673 slivers beat Hyper3D's 5,048, and at equal 120 mm outer longest side the calculated solid fill is 47.00 rather than 63.96 cm³. I would then turn off the material to expose the nearly flat floor and check export again; 60 degenerate triangles and a failed Original size browser conversion are not details to hide. Neither file copies the pictured floor knit; there is no basis here for a general-purpose winner.

Use SupaVoxel for this job

If the brief is a shallow decorative concept and you can rework the floor, SupaVoxel lets you inspect the textured tray beside its bare mesh before signing off on the knit. Bring your own picture, check the hidden side, and verify your chosen export before promising a physical item. For a finished cable-knit print, neither of these one-shot results is done.

How I tested this

One independently generated 1,862,608-byte reference PNG, byte-matched across the products; one downloadable Hyper3D Gen-1.5 PBR GLB from a 27,809,936-byte ZIP; one SupaVoxel analysis GLB reconstructed locally from its retained compressed project source after its Original size browser export failed. The separate Hyper3D Gen-2.5 preview spent 0.5 credit but was blocked on Free, so it contributes no file measurements. We parsed the files and welded UV seam vertices before topology checks. Longest bounding-box side, not vertical height, was scaled to 120 mm: actual vertical spans were 24.48 and 20.71 mm. Resin is volume × $35/L under hypothetical solid fill; downward-facing area uses world Y-up normals within 45° of down, not slicer support. No slice, print, strength or fit test occurred; no sourced design for the hidden back exists.

Aqua Voice Review 2026: The 230 WPM Dictation App's Plans, Limits and Trade-offs

Elias Larsen — Thu, 08 Oct 2026 15:52:38 +0000

Aqua Voice is a desktop and mobile dictation app that turns speech into clean, formatted text in any application, with a free tier of 1,000 words and paid plans from $8 a month. I went through Aqua Voice’s feature set, its five plans, its platform coverage and the claims on its homepage to work out who should switch from the keyboard and where a voice generation tool, rather than a voice input tool, is the thing you actually need.

“We’ve typed for 150 years. It’s time to speak.” That is the first line on Aqua Voice’s homepage, and the pitch under it is precise: hold a key, speak naturally, and Aqua Voice writes what you meant — into an email, a document, a Slack update or a Codex prompt — at up to 230 words per minute against a 40 WPM keyboard. Aqua Voice has become a favourite of people who spend their day prompting coding agents, and the testimonials on its download page are almost all from that crowd. I spent a week reading Aqua Voice’s homepage, its pricing table and FAQ, and the download page, mapping every plan and every claim.

Because Aqua Voice is strictly about getting your words onto the screen, I kept a text-to-speech and voice-cloning studio, Supavocal, open in the next tab for the opposite direction — turning finished text back into a voice — and I come back to that at the end.

Aqua Voice’s pitch is one gesture — hold Space and speak — and one number: 230 WPM with Aqua 3.1, five times faster than typing.

What Aqua Voice is

Aqua Voice is a system-wide dictation app. You hold a hotkey, talk, and Aqua Voice’s model refines what you said as you talk — removing filler, fixing punctuation, formatting lists — then inserts the text wherever your cursor is. There is no separate window to paste from. Three ideas carry the product:

Works with all your apps. Aqua Voice sits at the OS level, so it dictates into Mail, Notion, Cursor, Codex, a browser form or a terminal without per-app setup.
Your thoughts set the pace. Aqua Voice is built to keep up with fast, rambling speech and to tidy it rather than transcribe it verbatim.
Your screen is its dictionary. Aqua Voice reads what is on screen — code syntax, variable names, product names — so “useState hook for selectedTaskId” comes out spelled the way your file spells it.

The homepage demo is a React component request dictated at 230 WPM with the correct camelCase identifiers, next to the same sentence typed at 40 WPM with typos. That is the Aqua Voice use case in one image.

Getting started with Aqua Voice

Aqua Voice is available for Mac, Windows and iOS, with Android on a waitlist. The Starter plan is free, needs no credit card, and includes 1,000 words with the Avalon transcription model — enough to dictate a few long prompts and a couple of emails and decide whether your voice and Aqua Voice’s model get along. The Mac build is the primary one; “Get Started on Mac” is the call to action on every page.

The download page is a wall of quotes from developers — “for coding prompting with voice nothing beats Aqua Voice”, “saved me over 2 days of typing”.

Aqua Voice pricing in 2026

Aqua Voice sells two individual plans above the free tier and two for organisations, with annual billing 20% cheaper. From Aqua Voice’s pricing page and FAQ:

Starter — free. 1,000 words, Avalon transcription model, no card required.
Pro — $8/month billed annually, or $10 month-to-month. Unlimited words, custom instructions (tell Aqua Voice how you want text formatted), an expanded custom dictionary.
Max — $24/month annually, or $30 month-to-month. Everything in Pro plus Realtime Mode (see words appear live as you speak), the “Send it” voice command that submits the message without touching the keyboard, and early access to new features.
Team — $12/user/month annually, or $15 month-to-month, for teams of 2–9. Pro features plus centralised billing, team-wide settings and an enforced Privacy Mode.
Business — custom, for 10+ seats. SSO/SAML and SCIM, advanced reporting, zero data retention, team-wide dictionaries and volume discounts.

Students get 70% off Pro and Max with a school email — Pro drops to $2.40 a month annually. Everything is cancel-anytime.

The individual tiers differ on one axis — how much Aqua Voice does beyond transcription — not on word counts, once you are past the free 1,000.

Aqua Voice for coding and prompting

This is the audience Aqua Voice is chasing hardest. The homepage’s second section is titled “Prompt faster with your voice” and shows Aqua Voice dictating into the Codex app with GPT-6 Astra selected: “Give GPT-6 Astra the full context in the Codex app. Speak your prompts with Aqua, from the first idea to the final fix.” The screen-as-dictionary feature is what makes this work — Aqua Voice sees the file, the framework and the library names, so spoken prompts arrive with the right identifiers. The testimonials back it up: “all the other transcription apps don’t work well with code like Tailwind class names”, “lowest latency, highest quality STT I’ve used so far”.

The speed and accuracy claims

Aqua Voice claims “5x faster than typing and twice as accurate”, with the 230 WPM figure attributed to its Aqua 3.1 model. Both are demo-derived rather than benchmarked in public, and the accuracy half depends heavily on microphone, accent and how much you ramble. What is fair to say is that Aqua Voice’s refinement step — cleaning the transcript as it goes — is where it differs from built-in OS dictation, and users consistently describe it as “much more accurate than macOS dictation”.

What Aqua Voice gets right

Works everywhere. One hotkey, every app, no paste step.
Context from the screen makes it the best dictation option for code and prompts.
Unlimited words from $8 — the free tier is a trial, but Pro removes the meter entirely.
Custom instructions and dictionary let you shape the output rather than correct it.
Realtime Mode and “Send it” on Max close the loop for hands-free agent conversations.
Team controls — enforced Privacy Mode, SSO, zero data retention — for regulated workplaces.

Where Aqua Voice falls short

1,000 free words is small. At 230 WPM that is under five minutes of speech; you will decide on the paid plan before you have finished evaluating.
Android is waitlist-only. Mac first, Windows and iOS second, Android not yet.
Realtime Mode is a Max feature. Seeing your words as you speak — the thing most people expect from dictation — costs three times Pro.
Cloud processing. Aqua Voice’s model runs in the cloud; Privacy Mode and zero data retention exist, but at Team and Business tiers.
Input only. Aqua Voice turns your voice into text. It does not produce a voice from text, clone one, or make audio for anything you publish.

Aqua Voice vs the alternatives

Against OS-native dictation, Aqua Voice wins on cleanup, context and speed. Against other AI dictation apps, its edge is the screen-aware dictionary and the coding focus; its costs are the metered free tier and the Max-only realtime view. Against text-to-speech tools there is no contest because they do the opposite job — which is exactly why I found myself running both.

Verdict: is Aqua Voice worth it in 2026?

For anyone who writes long prompts to AI agents, drafts a lot of email, or simply thinks faster than they type, Aqua Voice Pro at $8 a month is an easy purchase — the free 1,000 words will tell you in an afternoon. Max is for people who live in voice all day and want to send without touching the keys. Try it on your own long prompts for a day before you choose between Pro and Max.

The alternative worth keeping next to it: Supavocal

Two Aqua Voice facts kept a second tab open: it is voice in, never voice out, and it is the wrong tool the moment you need audio for a video, a course or a product. Supavocal is a text-to-speech and voice-cloning studio for exactly that direction.

30+ voices, 30+ languages, with cloning from a 15-second sample and a Voice Lab for tuning delivery.
Simple metering. 1 credit is 1 character; cloning costs 3x. Starter is $49 a month for 600,000 characters, Pro $99 for 1.5 million, Business $299 for 5 million.
Commercial use included on every paid plan, with API access for pipelines.
Built for output — narration, dubbing, voice-overs and agents that speak, all from text you may well have dictated with Aqua Voice.

Use Aqua Voice to get the words out of your head. When those words need to be spoken back to an audience, keep Supavocal open.

Photo by Klim Musalimov on Unsplash