Anima Turbo LoRA · Text to Image

Anima Turbo is a 2B anime DiT that runs on a 6GB card if you want it locally, so why pay $0.018 a picture to rent it? Because of the LoRA slot. This node lets you attach up to three of your own Anima weights per request and stack them, which turns "which of my styles do I need loaded today" into a dropdown decision instead of a swap-and-reload.

It's also the least fussy way to try Anima before committing to it. No 6GB+ of weights, no Qwen3 encoder folder to get right, no sampler discoveries. Wire a prompt and go. If Anima's prompt comprehension is what you're after, this is the cheap audition.

Prompting it properly

Anima's whole personality is that Danbooru tags and natural language both work, and they can be mixed in the same prompt. The model card's order is quality tags, character count, character, series, artist, then general tags - and the community's loudest advice is to use real Danbooru tags rather than approximations, because the model knows the tagging system in detail. masterpiece, best quality, score_7, safe is the recommended prefix if you're prompting sparsely, but drop it the moment you're specifying an artist or an unusual style: it narrows the distribution toward generic polished anime and eats your variation. score_8 and score_9 in particular pull toward furry territory, which surprises people.

The loras input is on this node as an optional field, and LoRA trigger words belong at the front of the prompt - the tooltip says so, and it's true of every trigger-word model.

Inputs and outputs

• prompt - Danbooru-style tags, natural language, or both. Trigger words first.
• aspect_ratio - fifteen options, 1:1 by default.
• resolution - 1k (about one megapixel) or 1.5k (about 2.4; that's Anima's native training scale, so it's the one to take when detail matters).
• seed - fixed by default, so re-queueing returns the cached image instead of a second charge. Switch control after generate to randomize for a new draw, which is what you usually want when sampling a style.
• loras (optional) - up to three Anima-trained .safetensors files, each with its own strength, supplied by SpicyAPI LoRA nodes chained together. LoRA nodes take a direct download URL that opens without a login, and a scale from 0 to 4.

Out comes images, plus task id, a STRING that records which request this was.

Install

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

Restart, paste a key from spicyapi.ai/console/keys into Settings > SpicyAPI (or export SPICY_API_KEY), restart once more so the node list is built from your account. ComfyUI Manager: search SpicyAPI. No Python dependencies and no weights ship with the pack; it's a thin client, so CPU-only machines are fine.

What it costs, and what to expect from old weights

About $0.018 per image. Cheap enough that the sensible approach is to burn four or five cents comparing prompt phrasings rather than reasoning about it. Set Settings > SpicyAPI > Max cost per run if you want a ceiling, and remember the pack forces seeds to fixed at node creation because every run is metered.

Expect older Anima LoRAs to be unpredictable. The model went preview 1 → 2 → 3 → base 1.0 quickly, and preview 2 weights got noticeably worse on preview 3, with the preview 3 → base jump contested: generic concepts survive, precise character and style detail often don't. Lower the scale and see what happens before you conclude the LoRA is broken.

Errors are the pack-standard four: No SpicyAPI key is set; 40310 for an unverified email; 40004 when your option combination has no deployment, where the message names the field and nothing is charged; and request_id ... errors, which are for support. A private LoRA URL gets rejected rather than ignored. And if a node or model is missing, restart ComfyUI.