Anima Turbo · Image Editing

Anima is the first anime model in years that made people actually switch stacks, and this node is the cheap way to point it at a photo you already have. One source image, one instruction, one strength slider - from about $0.012 an image, the cheapest image node in the pack.

The slider is the whole story. At 0 you're barely touching the source; at 1 you've thrown it away and the model is drawing from your prompt alone. The default of 0.8 is a sensible stated position - keep the composition, replace the style - and the practical range for photo-to-anime conversion sits roughly between 0.6 and 0.85. Below that you get a filter. Above it you lose the person.

Why Anima is worth caring about

Anima is CircleStone Labs' 2B DiT built on NVIDIA's Cosmos-Predict2 rather than SDXL, with a Qwen3 text encoder. That architecture choice is why it understands natural language as well as Danbooru tags - the CLIP encoders in every SDXL anime model were trained so hard on tags that they forgot how to read a sentence. Anima's didn't. It is also, per its weights, unfiltered, and this endpoint inherits that.

Turbo means a distilled variant: same model, far fewer steps, seconds instead of tens of seconds. In a local setup Turbo asked for tradeoffs - the community's own testing found the distilled path can flatten variety and soften quality. Rented per image at a couple of cents, though, that tradeoff is somebody else's problem, and you get to keep the prompt adherence that made Anima popular in the first place.

Inputs

• prompt - and the tooltip is the important part: describe the image you want back, not just the change. "Anime illustration of a woman in a red coat, 90s cel shading" is a working prompt. "Make it anime" is not.
• image_url - the single source. The output keeps its aspect ratio, so what you compose in is what you get out.
• strength - 0 to 1, default 0.8. How far to move away from the source. This is the field you'll actually be tuning.
• resolution - 1k (about a megapixel) or 1.5k (about 2.4MP). Aspect ratio is preserved either way.
• seed - fixed by default so a re-queue returns the cached result rather than billing again. Set control after generate to randomize when you want a different take at the same strength.

Outputs: images for Save Image or any downstream node, and task id, a STRING naming the request in SpicyAPI's logs.

Install

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

Restart, add a key from spicyapi.ai/console/keys in Settings > SpicyAPI - or export SPICY_API_KEY if there's no browser in front of the machine - and restart again so your account's catalogue loads. ComfyUI Manager: search SpicyAPI. Nothing else to install; no weights, no Python packages, no GPU needed, which is the entire reason a 2B DiT node can sit in a workflow on a laptop.

Cost and troubleshooting

About $0.012 per image, shown on the node and quoted before the run. At that price the failure mode isn't your wallet, it's impatience: variations are cheap, so iterate on strength rather than agonising over the first result. Settings > SpicyAPI > Max cost per run still exists if you want the belt and braces. As everywhere in this pack, cancelling stops your wait and not the server-side job.

Errors: No SpicyAPI key is set, 40310 for an unverified email, 40004 when the options you chose aren't a served combination (change what the message names; nothing charged), and request_id ... messages to hand to support. A model missing from your menu needs a ComfyUI restart.