DEV Community: lyric li

Reading ages, page counts, and why the text draft is free of credits

lyric li — Tue, 08 Sep 2026 03:41:35 +0000

Reading ages, page counts, and why the text draft is free of credits

If you build tools—or you just think like someone who does—you eventually ask the same question of a kids story product: which knobs are configuration, and which knobs are billable?

ToviStory makes that boundary unusually explicit. Reading age, page count, language, tone, and story text live on the free side of the ledger. Successful images spend credits. This post is a practical map of those levers so you do not burn picture credits debugging vocabulary.

The knobs that cost 0 credits

Public docs and pricing agree on a clean rule:

Work	Credits
Create account	0 (free signup, no card required)
Story draft	0
Text revisions	0
Character profile analysis	0
PDF download	0
Successful image (portrait / page / background / style variant)	1
Failed image	returned automatically

That is why the text draft being free of credits matters. You can treat the manuscript like a spec: change reading band, cut sentences, swap tone, regenerate wording—without touching the image budget. Most “AI story” frustration I see in parent chats is actually spec thrash paid for as image thrash. This metering model pushes you to thrash on the free side first.

Reading ages (product bands)

Homepage generator + FAQ:

Band	Ages
Toddler	1–3
Preschool	3–5
Kids	59
Pre-teen	912

FAQ also mentions teen and adult writing options. The site frames these bands as guiding vocabulary, sentence length, and words-per-page. Separately, you set page count: 8–16 body pages, plus cover and back cover.

Builder takeaway: age is not a prompt adjective you hope the model notices. It is a first-class control. Page count is another. Do not discover you wanted 8 pages after you have already illustrated 14. Covers still generate as images when you illustrate them—so length decisions cascade into credit math.

Language and style are also “set once” controls

Per book:

One story language (editor offers 19, including English, Spanish, French, German, Simplified Chinese). You can change language later and regenerate text.
One art style guiding portraits and page art. Marketing claims 15 styles; the public carousel names at least: 3D Animated, 2D Cartoon, Watercolor, Storybook Classic, Soft & Dreamy, Comic, Paper Craft, Pencil Sketch.

Pick style before you pay for a portrait. A watercolor portrait will not happily become your comic-page reference without another image spend (style variants are called out as credit-using images on pricing).

Homepage copy even gives directional hints: watercolor for soft bedtime, 2D cartoon for silly adventure, storybook classic for a keepsake. Use those as defaults if you do not want to bike-shed.

Why free text still is not a free illustrated book

The official guide is blunt: a new cast may need a portrait in the selected art style before you can request a draft, and free text generation does not mean a fully illustrated first book is free.

Example cost from that guide:

1 portrait
+ 8 story illustrations
+ front cover
+ back cover
= 11 credits

Extras and successful regenerations add 1 credit each. Failed jobs refund.

Packs (one-time, valid 12 months, no auto-renewal on the pricing page):

Pack	USD	Credits
First Spark	$1.99	50
Story Maker	$15.99	400
Story Shelf	$46.99	1200

About ** $0.04 / image** appears as a site shorthand. Terms still contain generic subscription-renewal wording; buy against the pricing pages one-time pack facts. Credit history shows expiration per batch—plan a shelf inside the 12-month window rather than assuming infinite rollover.

A builder-friendly order of operations

Create / reuse cast — preset, describe, or authorized photo; prepare portrait (this spends credits).
Configure free dials reading age, page count, language, tone/genre (e.g. Bedtime + warm + calm).
Draft and edit text — still 0 credits. Include a small problem and an ending direction (guide suggestion).
Illustrate in the page editor — scene + character references; inspect early pages before batch-spending.
Adult review — every page before a child sees it.
Read in library or download PDF — PDF does not spend credits.

Reusable character references are the architectural point of the product versus a text-only chat workflow: save the cast, bring them into the next story, keep books in a library. The homepage comparison table lists exactly those gaps in chat-only setups (no saved cast, no page editor, no reading-level control, no PDF path, no library).

Limits worth coding into your expectations

Adult-managed accounts (parents, guardians, educators, other adults). Kids contribute; they do not get independent accounts per Privacy/Terms framing.
Character consistency = reference help, not a guarantee of identical details every page. FAQ: AI details can vary; review and regenerate.
Authorized photos only; project storage covered by Privacy Policy; site states customer stories/reference images are not used to train their own general-purpose models.
Review before share. Safety is a human gate, not a marketing badge.

Try it, then tell me what you would meter differently

If you want to poke the levers yourself: tovistory.com (Write my story) and pricing. Walkthrough: How to Make an Illustrated Story.

I am curious—if you were designing this metering model, would you still keep text at zero, or would you meter draft regenerations too? Feedback welcome in the comments.

Text, first frame, or references — pick the MiniMax H3 path before you render

lyric li — Mon, 07 Sep 2026 11:17:14 +0000

Text, first frame, or references — pick the MiniMax H3 path before you render

If you open an AI video tool and immediately smash “generate,” you are making a production decision by accident. On TryH3, MiniMax H3 exposes three input paths—Text to Video, Image to Video, and Reference to Video. They share one credit gate (opening-frame review at 0 render credits, then a quoted full render), but they are not interchangeable.

This post is a chooser: when to use which path before you spend 25, 52, or more credits on a full clip. I am writing it the way I wish dashboards labeled their tabs—by input you already have, not by buzzwords.

Shared rules (all three paths)

Read these once so the per-path sections stay short:

Clip length 4–15 s; quality 768P or native 2K; synced stereo audio generated with the scene.
Six aspect ratios on text path (16:9, 9:16, 1:1, 4:3, 3:4, 21:9); image path can follow the upload.
Prompt budget 7000 characters where a prompt exists.
Visual / opening-frame review: 0 render credits. Full render starts after you confirm the quote.
Example quotes from pricing: 25 credits for 5s · 768P; 52 for 5s · 2K; longer/reference jobs cost more.
Failed generations refund credits automatically.
TryH3 is an independent browser workflow — not affiliated with MiniMax.

Quick chooser

You have…	Start here	Why
Only words for subject, camera, sound	Text to Video	Fastest path; one written direction drives picture + motion + audio
A locked still (or start + end stills)	Image to Video	Motion from known frames; ratio follows upload
Identity / motion / rhythm references to hold	Reference to Video	Multimodal guidance with one written tie-breaker

If two rows both seem true, prefer the more constrained path. A signed-off poster frame beats a clever paragraph. A face-reference pack beats hoping text keeps identity stable.

Path 1 — Text to Video

Use when: you can name the shot in prose and you do not yet have a hero frame.

From the text-to-video generator:

Describe subject, camera, movement, and sound in one pass.
Durations 4–15 s; six aspect ratios; 768P / 2K (page note: 768P uses fewer credits).
Tips on-page: 4–6 s for one action; 10–15 s for multi-shot; strong prompts often 30–80 words.
No separate sound stage—stereo lands with the full render.

Skip this path when: the client already signed off on a poster frame. Jump to Image to Video so you are not re-rolling the still from text.

Mini workflow: write the shot → opening-frame review → if the still is wrong, change one thing (camera line, duration, or ratio) only then accept the quote.

Path 2 — Image to Video (first / last frames)

Use when: composition is already decided in a still—or you know the start and end poses.

From the image-to-video generator:

One frame = open-ended motion; two frames = controlled transition (in-betweener).
Formats: JPEG / PNG, ≤ 30 MB, max two images for a first/last pair.
Output ratio follows the uploaded frame (adaptive / auto).
Prompt should describe motion, not re-describe the image.
Match source aspect to the intended output to avoid invented edges (page FAQ).

Skip this path when: you do not have rights to the still, or the still is a tiny compressed screenshot you cannot stand behind. Garbage in still applies.

Path 3 — Reference to Video

Use when: you need identity, camera language, or timing to stay anchored across a cut—character stories, brand motion language, music-timed social beats.

From the reference-to-video generator:

Images ≤ 9 · video ≤ 3 · audio ≤ 3 (limits depend on selected model).
Audio cannot be used by itself.
Roles on-page: image → identity/style; video motion/camera; audio → rhythm/timing.
Keep one written direction as the tie-breaker; combine with one job per reference intent rather than dumping every file you own.

Skip this path when: you only have a single still and no consistency problem—Image to Video is lighter and clearer.

Decision examples (no invented case studies)

These are decision patterns, not testimonials:

Social vertical, idea only → Text to Video, 9:16, 5 s, review opening frame, quote (~25 at 768P).
Product hero already shot as a still → Image to Video, one frame, prompt the move/light shift only.
Same character across two beats → Reference to Video with identity images; add a short motion video reference only if camera language must match.
Music bed timing matters → Reference to Video with audio plus image or video (audio alone is blocked).

Homepage use-case lanes (cinematic, product ads, character stories, social) map cleanly onto that list. The FAQ’s “when is H3 right?” answer—short directed scenes where motion, composition, and synced sound matter together—is the filter for whether you should be here at all.

The credit gate (do this after you pick a path)

Pick Text / Image / Reference.
Set length + ratio + 768P/2K.
Review the opening frame (0 render credits).
Read the quote; confirm only if the still works.
Wait for the MP4 (homepage FAQ: usually 1–3 minutes).

Packs (optional): first-buy 30% OFF live prices Boost $20.93 / Creator $34.93 / Studio $69.93, credits valid 12 months. Subscriptions available; sub credits do not roll over. Welcome credits = evaluation only; pack/sub credits support commercial use per site FAQ.

Honest limits

Not a MiniMax affiliate product page—independent workflow (About).
4–15 s ceiling; not long-form.
Reference audio never solo; image uploads capped as above.
Full renders cost credits; only opening-frame review is free of render credits.
Upload only media you have rights to use.
Cash refunds case-by-case; failed gens auto-return credits.

Try it

Pick the path that matches your inputs, then review before you render: https://tryh3.com/. Feedback welcome in the comments—especially if your chooser table differs in production.

Export one TikTok frame, then rebuild — this tool is stills-only

lyric li — Sun, 06 Sep 2026 02:50:35 +0000

You have a TikTok (or Reels) clip stuck in matcha mode. The whole video is green liquid art. You want a normal-looking photo of the person in it.

Here is the constraint up front: Rayopix Matcha Filter Remover processes still images only. It does not take a full video and remove the filter frame-by-frame. The supported creator path is: export one sharp frame → rebuild that still → repeat if you need more moments.

That sounds like a downgrade until you notice most "I need the photo back" requests are really about one or two frames anyway — a face, a couple pose, a product

Why full-video removal is a different product

Matcha, as Rayopix describes it, redraws surfaces as liquid and foam AI art. Doing that consistently across hundreds of frames is a video pipeline problem (temporal consistency, encode time, cost). The public FAQ answer is plain:

Still images only. Export one sharp frame from TikTok/Reels. Whole-video not offered yet.

Homepage use cases even list "TikTok video frames" with the same boundary: the page rebuilds photos, not complete moving video.

So this post is a DIY stills workflow, not a pretend After Effects tutorial.

DIY: pick a frame worth rebuilding

Play the clip and pause where the subject is readable.
Prefer a moment with visible face or product edges — not a whip-pan blur.
Avoid frames dominated by stickers, captions, or UI chrome cropping the head.
For couples/groups, pause where faces are separated if you can; merged liquid faces are harder (site note).
Export/save a still. Keep quality high. You will upload JPG, PNG, or WebP ≤ 10 MB, one image per run.

If your only export path is a compressed screenshot, try again for a cleaner capture. Soft edges make reconstruction guess more.

Rebuild the still (Auto + 1K first)

Open https://rayopix.com/.
Optionally drag free sample sliders to see Matcha art ↔ Natural rebuild. Those demos do not process your upload.
Sign in (required for your own image).
Upload the frame.
Aspect Auto, resolution 1K.
Click Remove matcha filter. Expect about a minute on the paid path.

What the model aims to do: rebuild natural color and texture while anchoring visible subject, pose, and framing. No prompt required. Output is labeled as an AI reconstruction not recovered hidden pixels.

hold.

Review order that saves credits

Identity / subject count
Pose and framing
Color and materials
Fine detail

Credits: 3 / 5 / 8 for 1K / 2K / 4K. Packs are one-time (Mini $5/42, Plus $12/108, Max $25/240, Ultra $49/480), never expire, no subscriptions. Stay on 1K until the frame's structure looks right, then upscale.

If you need another moment from the same video, export a new still and run again. That is the supported "multi-frame" approach today: multiple still jobs, not one video job.

Creator checklist

[ ] Clip is matcha-heavy (liquid/foam), not just a mild green tint
[ ] I exported one sharp still (not the whole MP4 uploaded)
[ ] File is JPG/PNG/WebP ≤ 10MB
[ ] I have permission to process this image
[ ] First pass: Auto + 1K
[ ] I judged identity before buying 2K/4K
[ ] I keep the matcha source beside the rebuild
[ ] I will not treat the rebuild as legal/medical/ID evidence

Mild green cast with texture intact? Use ordinary color correction; skip credits. Still have the camera original? Use that; Rayopix is for matcha-only copies.

Optional: a tiny export script mindset (no fake APIs)

Rayopix does not publish a public video API in the materials I am using here. Do not wrap the website in a scraper and call it a pipeline. Keep the automation on your side of the fence:

Use your editor or OS tools to dump candidate frames from the local MP4 you already own
Rank frames by sharpness (Laplacian variance, or simply eyeballing at 100% zoom)
Upload only the winners

Pseudo-flow:

# Illustrative local idea only — use whatever frame tool you already trust
# 1) extract a few candidates at scene pauses
# 2) pick the sharpest face-forward still
# 3) convert to JPG/PNG/WebP under 10MB
# 4) upload that one still in the browser tool

The credit meter stays honest because each browser job is still one image. Batching twenty frames means twenty still jobs if you choose to run twenty — not a hidden video discount.

What success looks like for a creator handoff

When I hand a rebuilt still to design or social, I label it as an AI reconstruction. I send the matcha source in the same folder. I note the resolution I paid for (1K vs 2K vs 4K). That prevents someone downstream from treating the file like an unaltered camera original.

If the team later needs a different second from the same TikTok, we export again. We do not pretend the first rebuild "contains" the rest of the video.

Compared with prompting a general chat model for twenty minutes (the pricing page's comparison category), the stills tool removes prompt thrash and keeps a restoration ledger in-account — at the cost of accepting the stills-only boundary.

Honest limits (read before you comment "it failed on my 4K video")

No full-video processing
Not exact original pixels
Sample homepage sliders ≠ your photo processed
Heavily crushed stills may not come back
Small text/logos and fine patterns under liquid art may be re-imagined

Try it + feedback

Export one frame, run Auto+1K, and tell me in the comments what export method you used (in-app save, screenshot, desktop grab). I am collecting practical frame-picking tips — not looking for miracle video promises the product does not make.

Tool: https://rayopix.com/ · Packs: https://rayopix.com/pricing

One reference photo to a reusable cast sheet for a short comic

lyric li — Sat, 05 Sep 2026 04:43:17 +0000

I used to "cast" a short comic by generating panel one, liking the face, then praying panel four still matched. Spoiler: it rarely did. What fixed the chaos was boring and technical: build a cast sheet first, then draw panels against that kit.

This is the workflow I use with Consistent Character AI — reference in, pose/expression sheet out, then scenes. Friendly for indie comics and game pre-prod. Not a promise of perfect clones.

What I mean by "cast sheet"

A small set of stills of the same character:

Neutral portrait (your north star)
3/4 turn, profile if you dare
4–6 expressions (smile, frown, surprise, tired…)
23 signature poses for action pages

Features page use cases literally list comics & storyboards and games & character sheets. That is the job. Marketing and product-mascot work uses the same kit logic: one face, many contexts.

I keep the sheet in a folder named after the cast member. Panels that cannot survive a side-by-side with the neutral portrait get deleted, not "fixed in lettering."

Step 1 — Pick a reference that can survive a sheet

Upload rules on Generate: PNG / JPG / WebP ≤ 20MB. Front-facing, well lit, face readable. Stylized characters usually behave better than celebrity photos; the UI warns real people, logos, watermarks, or sensitive content may be blocked (failed tasks refund credits).

New accounts: first character image free, no card — use it on the neutral portrait if you are inventing from Create Character first (text → character, Lite/Standard/Professional, shown as 1 credit), or on the first pose from a photo you already own.

If you are inventing from text, treat the best Create Character output as the new reference. Do not keep regenerating text prompts for every panel; that reopens the cousin problem.

Step 2 — Mint the sheet with Batch (one image, many prompts)

Open Batch Generate:

Mode: one image · multiple prompts (the other mode — multiple images one prompt — is handy later for outfit tests)
Attach the same reference
Write short pose/expression prompts (keep identity out of the way — the photo carries it)
Check estimated credit cost (about 1 credit per image on standard)
Run and download the ZIP

Example prompt atoms I reuse:

neutral standing portrait, soft comic flat color
three-quarter view looking left, same outfit
closed-mouth smile, eyes half-closed
running pose, side view, backpack visible
holding a coffee cup, waist-up, looking at camera

You are not writing novels. You are filling cells on a sheet. If a cell drifts, regenerate that prompt only — do not rebuild the whole ZIP.

Step 3 — Optional: Pro when one photo is not enough

Generate Pro accepts up to 14 reference images, quality 1K / 2K / 4K, optional web search, prompts ≤ 2000 characters. Credit cost scales with quality in the UI.

I reach for Pro when:

I already have 34 approved sheet cells and want to lock them as multi-ref
Print resolution matters (2K/4K)
A single photo keeps losing a scar, hair accessory, or logo-free emblem you designed

Standard Generate stays fine for most webcomic drafts: one reference, fixed resolution path. llms.txt positioning still holds either way: same character across poses/scenes from a reference — no LoRA training or fine-tuning required.

Step 4 — Panels against the kit, not against vibes

Only after the sheet exists do I generate story scenes. Options:

Manual Generate: reference + scene prompt
Batch again for a sequence of panel descriptions
Homepage Character Agent if I want a free chat to plan shots, then approve before spending credits ("Chatting is free — the Agent only generates images when you approve.")

Each panel still gets a human glance next to the neutral portrait. Site FAQ + guide: consistency retains traits but is not a mathematical lock. Reject cousins early.

A practical rule: if the cast sheet took 12 credits, do not "save money" by skipping review on panels. The sheet is the cheap insurance.

Hard boundary for game folks

If you came from For Game Developers: this is 2D concept / reference continuity across portraits, poses, costumes, environments. It is not production-ready 3D game assets. Use the sheet in pre-prod, narrative docs, and marketing validation; do not pretend it ships as a rigged mesh.

That honesty is why I like the page — it stops a bad handoff to a 3D contractor.

Credits without the mystery

Free: first image free; Standard 1K; shared queue; no Pro/video/upscale
Packs (pay once, 12 months, unlock Pro tools while active): Starter $9.99 / 50; Pro Pack $24.43 today (30% off first pack) / 400; Max $62.93 today / 1,200
Plans exist if you sheet characters weekly — see Pricing (Basic/Pro/Max with yearly Save 33% framing)
Paid includes commercial use subject to terms; keep rights to your reference
About: reference uploads deleted within 24 hours — keep local copies of the sheet

Recap checklist

One clear reference (photo or best text-to-character still)
Batch → expression/pose ZIP
Promote best cells; optional Pro multi-ref
Generate panels against the kit (Agent optional for shot planning)
Review every face against the neutral portrait

Try it on https://consistentcharacterai.com/ and tell me in the comments how many sheet cells you actually need before a 6-page comic stops drifting — I'm still tuning my own minimum.

Cleaning text off product shots before they go live

lyric li — Fri, 04 Sep 2026 10:31:36 +0000

Shipping a listing with last month's "SALE" sticker still glued on the hero image is the visual equivalent of leaving console.log in production. The fix is not glamorous. It is a pre-live cleanup step.

This post is how I run that step with AI Remove Text on photos the business owns—single shots first, then batches of similar SKUs—without pretending the model is a guarantee.

The problem in concrete terms

Product folders accumulate overlays:

Price tags from a seasonal campaign
Promo stickers ("NEW", "% OFF")
Date stamps from phone cameras
Leftover design captions from mock comps

The Ecommerce Image Cleaner page describes the same job: revise promotional text, price tags, and stickers on product photos owned by your business, then check each marketplace's current image requirements before listing. That last clause matters. Cleaning pixels does not auto-approve Amazon (or anyone else's) policy checklist—you still read the rules yourself.

Also: only remove brand logos if your business owns them or has permission. Same rights rule as the rest of the site.

Single-image flow (one SKU hero)

Open https://remove-text.com/.
Confirm ownership / permission (checkbox + notice: removing watermark/logo/attribution does not grant third-party rights).
Upload JPG / PNG / WebP ≤ 10 MB.
Scope: All detected text for stickers and captions.
Run AI → review at full size (edges, fabric/plastic texture, missed glyphs).
Download if acceptable.

New accounts get 3 free credits as watermarked previews—useful to probe whether your lighting/texture family cleans well. Paid results from credit packs are watermark-free. 1 credit = 1 image.

For a tiny leftover on a flat backdrop after AI, the Free Text Eraser is a no-credit, client-side brush fill. I use it as a detail tool, not as the main pass on busy product surfaces.

Batch flow (similar SKUs)

When ten packshots share the same sticker style:

Open Batch Remove Text.
Upload up to 10 authorized images (each ≤ 10 MB).
Set one scope + output format for the whole batch.
Start. Images process one after another (not parallel).
Review each preview. Retry failures; failures are not charged.
Download individually or all.

If SKUs need different scopes, do not force batch. Use the single-image tool so settings stay honest.

There is no public "API" documented for this product in the site intel I am writing from—so I do not script against private endpoints. Browser batch + generation history (mentioned on the batch FAQ) is the supported volume path.

Example: 24 SKUs, two sticker styles

Suppose a launch folder has 24 heroes. Twelve share sticker style A (corner price tag). Twelve share style B (diagonal SALE ribbon).

I would not dump all 24 into one batch. Shared settings mean one scope for the whole queue. Instead:

Batch 1: ten style-A shots, All detected text, review each
Batch 2: remaining two style-A + eight style-B (or keep styles pure—two batches of twelve split as 10+2 and 10+2)
Hand-fix any failure with single-image retry or Free Text Eraser on flat packing cardboard

Credit floor if everything succeeds once: 24. Realistic budget: 24 + a few retries on glossy plastic reflections.

Document which SKUs needed hand repair. That list becomes your next lighting or packaging brief—not just a cleanup chore.

Credits math for a go-live week

From the Pricing UI:

Pack	Credits	Price
Basic	600	$9.90
Plus	2,000	$13.93
Pro	4,000	$20.93

No subscription. Packs valid 1 year. Rough planning: a 40-SKU launch with one hero each is 40 credits if every pass succeeds once; budget extras for retries on hard textures. Free watermarked credits are for evaluation, not for final listing assets you need clean.

Pre-live gate (beyond the tool)

Before I mark imagery "live":

[ ] Rights cleared for every source file
[ ] Text/stickers removed or intentionally kept
[ ] Full-size review passed (or hand-fixed)
[ ] Marketplace image rules re-read for the destination
[ ] Export format matches listing needs (PNG transparency preserved when you choose PNG, per ecommerce FAQ)

Limits (so nobody is surprised)

Accuracy varies with contrast, compression, and background complexity
No promised fixed processing time or accuracy
Batch max 10; sequential
Manga-oriented bubble mode is a different scope—and bubble detection is experimental on manga tools—usually irrelevant for packshots
Whole-video subtitle cleanup is "coming soon"; stills/frames only on the subtitles tool

Try it / feedback

If you maintain a product image pipeline, walk one authorized SKU through the single-image path at https://remove-text.com/, then try a 10-up batch on similar shots. Comments welcome—especially on how you gate marketplace policy checks relative to pixel cleanup.

From one reference still to a talking character clip

lyric li — Thu, 03 Sep 2026 01:29:57 +0000

I wanted a character who could survive more than one pretty still—something I could move, then make speak—without training a LoRA or wiring ComfyUI on a Tuesday night.

The answer that stuck for me is a short pipeline on Consistent Character AI:

Lock a reference still (text create or a photo you trust).
Optionally remix scenes with Create Consistent Character.
Pick one video tool by job: Animate Image, Motion Control, or Lip Sync.

This post is the how-to, with the credit numbers and limits taken from the public product pages—not from vibes.

Step 1 — Lock the still before you spend video credits

Video will not fix a weak face. Animate Image, Motion Control, and Lip Sync all start from a character image. If the still is wrong, you pay video rates to animate the wrong person.

Option A — Text first: use Create Character. Describe appearance, clothing, pose, style; pick Lite / Standard / Professional; generate until one still is cover-worthy.

Option B — You already have a face: open Create Consistent Character (Image-to-Image Standard). Upload the reference, describe only what should change (pose/scene/outfit/lighting), confirm the credit cost, generate, and compare the output to the reference. The page says results vary—review is part of the workflow.

Standard consistent images cost 2 credits on the pricing table. New accounts currently get 4 credits with no card required, so you can usually mint or remix a couple of stills before you touch video.

Step 2 — Pick the video tool by the motion job

The product’s motion-control blog frames the choice clearly:

Job	Tool	What you provide
Subtle / cinematic motion without an action clip	Animate Image (Seedance 1.5 Pro)	Start still (+ optional end frame), motion prompt
Specific full-body action (dance, gesture, walk)	Motion Control	Character image + reference motion video
Mouth matches dialogue; body mostly still	Lip Sync (Kling AI Avatar)	Clear front portrait + short audio

Common stack called out on that blog: generate character → motion control → lip sync for dialogue. You do not need every step every time. Match the tool to the beat.

Animate Image — still to cinematic clip

Open Animate Image.

Upload the start frame (your locked still).
Optionally add an end frame if you want a controlled transition or trajectory.
Describe the motion.
Set aspect ratio / resolution / duration (UI defaults observed in the scrape: 16:9, 720p, 8 seconds, Fixed Camera).
Toggle Generate Audio only if you want cinematic SFX/ambience—the page marks it as additional cost.
Generate and download.

Pricing lists video generation with a starting cost of 6 credits. Final cost can move with duration, resolution, audio, and mode—believe the on-screen meter.

Use this when you need ambient life in a poster-like still for social or a marketing mascot beat, and you do not have a motion-capture-style reference clip.

Motion Control — copy real action onto your character

Open Motion Control.

Upload the character image (photo or AI still).
Upload the reference video of the action you want.
Pick resolution (720p / 1080p).
Generate and download.

The tool page claims full-body motion sync, including fine hand/finger work for dance and gestures, and videos up to about 30 seconds per generation. The companion blog mentions reference clips in the roughly 3–30 second range and notes that matching composition helps; extreme clothing conflicts between still and motion reference hurt results.

Use this when the performance lives in the reference video and your character should do that.

Lip Sync — make the portrait talk

Open Lip Sync.

Upload a clear front-facing portrait (identity preservation is the selling point).
Add audio: MP3, WAV, AAC, or OGG; ≤15 seconds; FAQ also says under 10MB. Prefer clean voice.
Optionally add expression prompts.
Generate and download.

Powered by Kling AI Avatar per the page copy. Best for educational presenters, spokesperson clips, and social talking-avatar posts—not for dance choreography (use Motion Control for that).

A minimal end-to-end example

Here is the shortest path I actually run for a talking clip:

Create Character → keep one front-facing still.
(Optional) Create Consistent Character once if the lighting is wrong for video. Cost: 2 credits.
Lip Sync with a 10-second voiceover under 10MB.
If I later need a wave or walk cycle for the same character, I go back to the same still and run Motion Control with a short reference action clip—not with the lip-sync output as a new identity source.

That last habit matters. Video outputs are for publishing. The identity source of truth stays the still (or a carefully approved still remix).

If you need higher detail for a client pitch before motion, insert Create Pro (5 credits for 2K) between still and video. Same pipeline shape; different sharpness budget.

Credits, signup, and honest limits

Quick map from the public pricing table:

Text-to-image: 1
Standard consistent still: 2
Pro consistent 2K: 5 · Pro 4K / music: 10
Video: from 6 (then meter)

Monthly plans if you outgrow trial credits: Lite $9.90/100, Plus $19.90/300, Pro $39.90/600. Yearly claims save up to 50%. Subscription credits refresh each cycle and unused do not carry over; PAYG packs are described as valid for 1 year.

Limits worth budgeting around (welcome post + tool FAQs):

Extreme camera angles on stills are hard.
Big costume changes can soften facial consistency.
Video generation is still evolving; peak hours can slow runs.
Lip sync audio must fit the format/size/duration caps above.
Always review stills against the reference before you spend video credits.

Commercial use: read the Terms of Service. Support: support@consistentcharacterai.org.

Try it / tell me where it breaks

Pipeline in one line: still → choose Animate / Motion / Lip Sync → meter → review.

Start with the 4 signup credits at https://consistentcharacterai.org/ if you are new. If you have a cleaner still→talk stack (or a failure mode I should warn about), leave it in the comments—I read them.

How I turn flat AI images into PSD-ready layer stacks

lyric li — Wed, 02 Sep 2026 07:59:03 +0000

Flat AI exports are fine until you open them in Photoshop or Canva and remember: one raster plate. No subject layer. No background plate. No “move the mug without tearing a hole.”

This is the handoff path I use with Image to Layers — flat JPG/PNG/WebP → Photoshop-compatible PSD or a PNG ZIP I can stage in Canva. Steps, credit math, and limits. No miracle claims.

The contract (read before you burn credits)

What you get

Independent RGBA raster layers, aligned on one canvas
A Photoshop-compatible .psd (Image to PSD path: up to 6 layers) and/or downloadable transparent PNGs
Optional ZIP of PNGs for non-Photoshop reuse
For Canva: those PNGs as normal image uploads (move / resize / hide / crop)

What you do not get

Original editable text / fonts
Vectors or smart objects from a “lost” design file
Adjustment layers, original masks, effects, or edit history
A live connection to your Canva account, or native Canva text/vector objects

Site wording is consistent across Image to PSD and Qwen pages: this is generative decomposition, not forensic recovery of a source PSD.

Upload limits: JPG / PNG / WebP ≤ 10 MB. Runs in the browser.

Path A — Direct Image to PSD

Use when Photoshop is the destination and you want a .psd out.

1. Open https://imagetolayers.org/image-to-psd
2. Upload flat JPG / PNG / WebP (≤ 10 MB)
3. Let AI build independent raster layers (up to 6)
4. Download .psd
5. Optional: grab ZIP of PNGs for Figma / other tools

Page guidance I follow:

Match layer count to composition (2–3 simple; 4–6 multi-object)
Prefer clear visual boundaries; treat hidden areas as reconstruction
After export: inspect the stack, rename/organize, refine edges non-destructively, save a working master

Credit reminder on that page: 2–6 layers → 20–60 credits (pricing rule: 1 output layer = 10 credits). Failed layer credits return automatically.

Path B — Agent Layers → PNG stack → save as PSD yourself

Use when you want a named layer plan before spend, then you’ll assemble the PSD in Photoshop (or any editor that eats transparent PNGs).

1. Open https://imagetolayers.org/ (Agent Layers)
2. Upload supported image
3. Choose Agent Layers, set count in 2–6
4. Describe the stack back-to-front (incomplete OK — Agent fills assumptions)
5. Review the named layer plan BEFORE generate
6. Generate (~1 minute per home copy)
7. Download aligned full-canvas transparent PNGs
8. Drag PNGs into Photoshop in stacking order → File → Save as PSD

Homepage FAQ says the same PSD idea plainly: export PNGs named in stacking order; drag into Photoshop (etc.) to save as PSD.

Agent extras from the tool page: composite preview + full-size aligned layers; sign-in required to restore unfinished Agent tasks.

This is my default when the flat image is an AI photo / product / portrait and I care about arguing with the plan before credits leave.

Path C — PNG ZIP → Canva as image elements

Use when the team lives in Canva, not Photoshop.

1. Run a split (Agent or dedicated Canva-editable tools)
2. Download ZIP of independent PNGs
3. Upload PNGs into Canva as normal images
4. Arrange / resize / hide pieces on the artboard

Tools on the site:

Image to Canva Editable
PNG to Canva Editable (logos, stickers, transparent assets)

Explicit non-features: does not connect to your Canva account; does not create native Canva text or vectors. Raster lettering still needs Canva text or official brand files if you need real editability.

Upstream tip: flat AI image → layers

If your source is still a prompt, the site also ships an AI Photo Generator (six models listed on that page) with cross-links into layer tools. My habit: generate flat → pick the cleanest frame → Path A or B. Do not expect the layer step to invent typography you can retype.

Gallery line I keep in mind: nothing was cut out by hand; hidden areas behind the subject get filled so the background can stand alone. That is why a PSD stack helps motion / parallax / layout — not because it resurrected a designer’s original file.

Credits checklist (before a batch)

From pricing:

Item	Site fact
Cost unit	1 output layer = 10 credits
Agent / PSD band	2–6 layers → 20–60 credits
Failures	Failed layer credits returned
Signup	20 credits = one successful 2-layer Agent; expire in 7 days; public; no commercial license
Paid packs	Starter $9.9 / Creator $39 / Studio $99 — private + commercial; credits never expire

If the PSD might ship to a client, buy a pack first. Signup results can be featured; that is on the homepage and pricing page.

VIP options (Qwen 1–10 with PNG/ZIP/PSD export; Magic auto 2–17) exist when you need higher exact counts or automatic depth — out of scope for this handoff post, but they sit on the same credit pool.

Cleanup notes (where I still touch pixels)

Site blog / tool warnings: glass, smoke, fine hair, motion blur, heavy overlap, tiny text, and very complex scenes are weak. Soft edges may span layers. I always:

Toggle layer visibility in Photoshop
Check occlusion fills on the background plate
Soft-mask hair / glass by hand if the client will zoom
Rebuild real type in the destination app when the “text” is just pixels

Try it / feedback

Pick one flat AI export you already dislike opening as a single plate. Run Image to PSD or Agent Layers on https://imagetolayers.org/, open the .psd or PNGs in your editor, and drop a comment with where the edges still need a pass. I read those — the weak-material list is more useful when it is attached to a real file.