For everyone: sheets, GIFs and lettering
Sprite sheets, talking-avatar mouth shapes, GIFs from a single still, and text that sits on a surface instead of floating over it.
All of it works on free engines. None of it needs a card.
Styles
36 looks in six families — clay, paper, drawn, anime, photographic, graphic, tabletop — each with notes on which engines can really do it. Star the ones you use and they come first in the style library and in the wizard.
Two house styles:
- Skilitsa — a polished 3D cartoon in happy, saturated colour, with a small
skilitsa.commark worked into the scene. The mark is text, so this style is limited to models that can spell, and those cost money. - Skilitsa Plain — the same look with no mark. Runs on any free engine.
Sheets
A sheet is a set of frames that belong together — same character, same light, same style, one thing changed per frame.
Every frame gets its own seed and an explicit “change ONLY this” instruction, which is what stops the model quietly redesigning your character between frames.
| Sheet | What it is for |
|---|---|
| Walk cycle | contact · down · pass · up, on both feet |
| Action set | idle, run, jump, attack, hurt — whatever your game needs |
| Turnaround | the same character from every angle |
| Visemes | mouth shapes for a talking avatar — see below |
| Expressions | one face, many moods |
| Avatar eyes and brows | blinks and eyebrow positions, with the mouth held still |
Frames arrive as separate rows, so you can redo any single one without touching the rest.
Sheets work best on an engine that accepts a seed and a reference picture — your own machine or a Google model. OVHcloud takes no seed, so frames drift apart there.
Talking avatars: the mouth shapes
A face does not need a different mouth for all 26 letters. Animation has used about ten shapes — visemes — for a century:
| Shape | Sounds |
|---|---|
| rest | silence |
| A · I | “father”, “I” |
| E | “bed” |
| O | “go” |
| U · W · Q | “you” |
| M · B · P | lips pressed shut |
| F · V | teeth on the lower lip |
| L | tongue behind the top teeth |
| C D K N S T | teeth nearly together |
| TH | tongue between the teeth |
Generate the sheet once and you can make that character say anything. Type a line of dialogue and the app works out the shape order and timing for you.
Mouths alone do not make a face look alive, which is what the eyes and brows sheets are for.
GIFs
From a picture you already made. Any finished row can become a GIF with camera motion — push in, pull out, pan, drift, tilt — computed from the single still. No second generation, so it costs nothing even on a paid engine.
From a description. Your text model turns what you describe into a motion plan, and the plan drives the frames.
Encoding is local, so the GIF never leaves your machine.
Why camera motion rather than new frames? Asking an image model for twenty frames of the same scene gets you twenty slightly different scenes. Moving a virtual camera over one real image gives coherent motion — instant and free.
Text that sits on the surface
Drop a text layer on any picture and drag all four corners independently. It is a real projective warp, not a skew, so text lands on a signboard, a wall or a book cover and looks painted there.
- Real fonts. Spelled correctly every time.
- It shrinks to fit.
- Find a quiet spot — free, instant, no model.
- Ask a model where it goes — “put it on the hanging sign” — and a vision model returns the sign’s four corners for you to nudge.
No vision model set up? The quiet-spot finder covers you. Nothing breaks.
→ Setting up a free vision model — one Mistral key, no card.
Vectors
Vectors are code, not pictures, so an image model cannot make them. A code model can: SVG icons, SVG illustrations, and Lottie animations.
Everything generated is sanitised before it is shown or saved: scripts, event handlers, embedded objects and external references are removed.
Codestral is free and made for this.
Practical notes
Checking a picture properly. Click a row, then its picture, to see it full size. A thumbnail hides a sixth finger.
A consistent character across many pictures. Fix the seed, keep the description word-for-word identical, change one thing at a time. A sheet does this for you.
Words inside a picture. Use a Nano Banana model, which can spell, or add the words with the Letterer, which always can.
Aspect ratios. Set per row: 16:9 video, 9:16 shorts, 1:1 square. Cloudflare and OVHcloud ignore it and return their own size.
Bulk. The whole point. Sixty thumbnails in sixty rows, named and sorted, with a CSV of what happened.