Anima Turbo · Text to Image

Four widgets, about $0.012 an image, no filter. Anima Turbo is CircleStone Labs' distilled 2B anime model, and this node is the shortest path to it - a prompt, an aspect ratio, a resolution, a seed.

Anima is the anime base that got people off the SDXL stack. Not because it draws prettier faces, but because it reads: it runs on Cosmos-Predict2 with a Qwen3 text encoder, so natural language works and Danbooru tags work, in the same prompt, without the tag-trained CLIP encoders that make Illustrious and NoobAI treat a sentence as a pile of lexical debris. If your prompts have ever been "tags good, prose bad" or vice versa, this model is the one that stops caring.

Turbo is the distilled version: far fewer steps, seconds instead of tens of seconds. Locally that distillation has a reputation for flattening variety, which is the honest tax. Here it's a per-image rental, so the tax is roughly a cent.

Using it well

The single best thing you can do is learn real Danbooru tags - Anima responds to them more precisely than to approximations of them. The model card's prefix is masterpiece, best quality, score_7, safe with a negative side that a local setup would use; there's no negative-prompt field on this node, so lean on positive phrasing instead. And the community's most repeated tip is to drop that quality prefix when you're already naming an artist or a distinctive style - it steers toward generic polished anime and quietly reduces the variation you came for. score_9/score_8 skew furry, which is a known surprise rather than a mystery: those tags come from a classifier trained partly on e621.

If a prompt comes back sparse, that's the model working as designed. It follows what you state and omits what you don't; it is not a model that fills in generously. Budget ten minutes reading tag documentation and it stops feeling broken.

Inputs and outputs

• prompt - tags, sentences or both.
• aspect_ratio - fifteen ratios, 1:1 default. Cropping for a wide shot is a common cause of duplicated characters, since a square frame leaves space the model wants to fill.
• resolution - 1k or 1.5k. 1.5k is Anima's native training scale, so take it when anatomy matters.
• seed - fixed by default. Flip control after generate to randomize to get a new result on each run; that's how you'll want it while exploring.

Outputs: images, and task id, a plain STRING naming the SpicyAPI request.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, open Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys - or set SPICY_API_KEY on a headless box - and restart again so your account's model list loads. ComfyUI Manager: search SpicyAPI. Nothing installs: no packages, no weights, no GPU required.

Costs, and the errors you'll actually meet

About $0.012 per image, rate on the node, quoted before each run. That's cheap enough to be genuinely experimental with, and it's worth setting Settings > SpicyAPI > Max cost per run anyway, since a batch of tweaked copies adds up faster than people expect. Cancelling a run stops your wait but not the upstream job, which finishes and is billed if it succeeds.

Error messages map cleanly: No SpicyAPI key is set (key missing or mistyped), 40310 (unverified account email), 40004 (no deployment serves that option combination - change the field it names, nothing charged), and anything ending request_id ..., which is what support needs. Missing nodes mean restart ComfyUI so the catalogue refreshes from your account.