Create Custom Emotes
in Seconds with AI

A few notes the tables never explain. Discord's emoji and sticker slots are genuinely different products, not two sizes of the same thing: an emoji is a 128px inline image with a 256 KiB cap and GIF animation, while a sticker is a 320px message-level image with a 512 KiB cap where animation means APNG and is capped at five seconds. Discord's API also accepts GIF and Lottie JSON for stickers, but Lottie uploads are restricted to Verified and Partnered servers, so APNG is the one that works for everybody.

YouTube is the odd one out twice over. It accepts anything from 48×48 up to 480×480 and gives you nearly a megabyte to play with, which is luxurious — but it does not animate. Upload a GIF and YouTube displays the first frame as a still. If animation is part of your set's identity, YouTube gets a static variant, and you should pick a frame that reads on its own.

Kick documents the slot counts — 60 channel emotes for any channel, plus 24 subscriber emotes once you're an Affiliate or Verified — but publishes no pixel size or byte cap. We've left those cells blank rather than repeat a number from a third-party blog. In practice, the square transparent PNG you made for Twitch is the safe thing to upload.

For the full Twitch breakdown — the three sizes, sub badges, follower emotes and the animated GIF budget — see the Twitch emote maker page. If you already have artwork and just need it cut to every platform's spec, the emote resizer does it in your browser for free.

Every spec above is really one constraint wearing different costumes. Twitch renders emotes in chat at 28×28. YouTube's own documentation notes that membership emoji show at 24×24 points on mobile. So the honest design target is roughly a 25-pixel square — about one line of this paragraph tall — and every other size you export is just a bigger copy of that.

Which produces a short list of rules that hold on every platform:

• One idea per emote. A character doing exactly one legible thing. Two things at 28px is zero things.
• Silhouette first. If the shape isn't recognisable as a solid black cutout, no amount of interior detail will save it. Squint at the 112px version; whatever you can still identify is what survives to 28.
• Contrast against both chat themes. A dark character with a dark outline vanishes in dark mode; a pale one vanishes in light mode. A thin rim in a contrasting colour fixes both at once, which is why so much good emote art has an outline that isn't black.
• No text. Five letters across 28 pixels is about five pixels per glyph. If the joke needs a word, it wants to be a Discord sticker at 320px, not an emote.
• Push the expression into the eyes and mouth. Oversize them past what looks right at full resolution. They will look correct at 28px, which is the only size that counts.

That last rule is the whole reason chibi proportions dominate emote art. It isn't a fashion — a big head and big eyes on a small body is the proportion that allocates the most pixels to the parts carrying meaning. If you don't have a strong reason to pick something else, pick that.

Describe a subject, not a scene. “Excited orange corgi with a tiny party hat” is a good emote prompt. “An orange corgi in a field at sunset looking excited” is a good wallpaper prompt and a bad emote: the field and the sunset are the first things destroyed by downsampling, and they eat the pixels the dog needed. Two to four identity anchors is the sweet spot — subject, one colour, one accessory, one expression. Past that the model starts averaging your instructions instead of following them.

Style isn't decoration, it's a file-size decision. A softly shaded 3D render and a flat pixel sprite of the same character are wildly different file sizes at the same dimensions, because PNG and GIF both compress flat colour areas well and gradients badly. That's why a 3D emote can land several times heavier than a bold-line emote of the same subject at the same 112×112 — and why the 3D one is the one that runs into Slack's 128 KB ceiling, or Twitch's tighter manual-upload limit, first. Choose the style with the platform in mind, not just the vibe.

Why transparent backgrounds are a second step. Image models don't emit an alpha channel. Ask one for “a corgi on a transparent background” and you'll usually get a corgi standing on a painted grey checkerboard, because a checkerboard is what “transparent” looks like in the training data. The approach that works is the opposite: generate the subject on a flat, solid background, then cut that background away as a separate pass. We append with a solid white background to every generation prompt and then run the result through a salient-object segmentation model, which identifies the subject rather than keying out a colour — so a white character on a white background still separates cleanly, which a chroma key could never do. Every emote you download from us has already been through it.

Static versus animated is one step versus three. A static emote is a single generation. An animated one is: generate the static, feed that exact image into a video model, then convert the video into a GIF small enough to upload. The middle step has to be image-to-video rather than text-to-video — text-to-video invents a fresh character every time, so you lose the one you just approved. The third step is where most animated emotes actually die, and it's covered below.

What image models are still bad at. Legible text inside the artwork — expect garbled letterforms. Exact brand hex codes — you'll get close, not exact, so correct it in an editor if it matters. And consistency across a set: re-running the same prompt gives you seven cousins, not seven expressions of one character. Generate one emote you actually like, then use that image as the reference for the rest of the set. That single habit is the difference between a set and a pile.