AI Emote MakerThe same face in every emote.

Ask any image model for “my character waving” and then “my character laughing” and you get two characters who could be related. Different nose, different hoodie, slightly different hair. For one image that’s fine. For an emote set that sits side by side in a chat picker, it looks like a template pack.

The fix isn’t a better prompt. It’s a reference that never changes. Before any emote, we draw a character sheet — front, three-quarter and back views plus six expressions — and you confirm it. Every later request sends that sheet and your original pictures as references, with a fixed instruction to keep face, hair, outfit, colors and line style identical.

Image models don’t output GIFs. So each animated emote is drawn as a 4×2 sprite sheet — eight frames of one loop in one image — at 16:9. Your browser slices it into frames, finds the character in each cell, lines them up on the head and the bottom edge so nothing jitters, adds a white outline, and encodes the GIF.

Doing the slicing locally keeps the server cheap and the result honest: what you preview is exactly what you download. Measured on our side, one sheet takes about 20 seconds to draw.

• PNGTuber states (idle, talking, blink, blink-talking) on one 1024 canvas.
• Sub badges for six tiers and bits badges, readable at 18 px.
• Channel points icons in a theme you pick.
• Avatar, offline screen, banner and six panels, with your channel name drawn as real text.
• Holiday outfits that become their own versions of the character.

Because the character is stored, you can come back next month, make three more emotes, and they still match.

Every prompt is assembled on our server from a recipe: the kind of asset, the grid, your preset, the consistency sentence, and at most one short piece of text from you. That’s deliberate. Free-form prompts make results drift off-model, and they are how people try to slip in someone else’s character or something a platform would reject.

The short text you can add is filtered too: it keeps plain letters and punctuation, and refuses a list of obvious problems and prompt-injection phrases. It’s not a perfect filter — the model has its own — but it keeps the tool focused on your character.

In practice the presets cover what chat uses. When they don’t, “Custom move” takes 60 characters, which is plenty for “spinning a basketball on one finger” or “typing furiously on a tiny keyboard”.

Generating the art is maybe a third of the job. The rest is mechanical and easy to get wrong at 1 a.m.: cutting frames out of a sheet, lining them up so the head doesn’t jitter, adding an outline, choosing the first frame for the thumbnail, encoding three sizes, keeping each under its byte limit, and counting flashes. All of that runs in your browser after the model is done, and the check panel shows the result per platform before you download.

Consistency is good, not perfect. Small accessories, text on clothing and hands are the usual drift points. Simple designs hold up best: a clear silhouette, two or three colors, no tiny logos. If your sheet looks off, redo it before you lock — that’s the cheapest place to fix anything.

One person or one character per picture. Group photos and busy screenshots give the model too many choices.