anima / sdcpp / createImage

Anima is the anime model that finally lets you type sentences. It's a 2B DiT built on Cosmos-Predict2 with a Qwen text encoder, which means it takes natural language and Danbooru tags in the same prompt, understands multi-character scenes, and follows instructions in a way no Illustrious fine-tune can. This node runs it on Civitai's sdcpp fleet - so you get Anima's prompt adherence without downloading the checkpoint or waiting out its famously slow local iterations.

The defaults on this node are telling: steps 30, cfg 4 - exactly the model card's recommended config - and 1024 square. On a local card Anima is slow for its size (over a minute on a 4060); through the cloud you trade that wait for Buzz. The use case is obvious: an anime character or scene in your head, typed in plain language, out as pixels, from a machine with no GPU. Or testing an Anima LoRA before you commit to a local setup.

How it works

Standard pack plumbing: submit engine: sdcpp, ecosystem: anima, operation: createImage to the Orchestration API, long-poll, get a ComfyUI IMAGE back. The model is Anima base - a diffuser_model input exists for swapping, but you can leave it unwired and let the recipe use the default.

The inputs that matter

• prompt (required) - the payoff. Write sentences, mix in Danbooru tags, or both: "girl with silver hair sitting by a window, soft evening light, looking into the camera, 1girl" works exactly the way you'd hope.
• negative_prompt - worth using. Anima is trained with dataset tags, so the trick that actually works is putting deviantart in the negative - it's a literal training-data tag that steers away from that aesthetic.
• steps / cfg_scale - 30 / 4, the model card's own recommendation. Don't lower steps much; Anima wants them.
• loras - via a Civitai LoRA Selector. Anima's LoRA ecosystem is young but real, and it stacks multiple LoRAs well.
• diffuser_model - optional CIVITAI_AIR; leave empty for the base model, wire a selector to try a fine-tune.

Install

Pack install, same as every node in the pack: ComfyUI Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or git clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials: Civitai Auth node into api_config or CIVITAI_API_TOKEN (from civitai.com/user/account); headless installs need the env var.

The Anima-specific surprise is the plain default aesthetic. Anima's base is deliberately style-neutral - a bare prompt returns something technically correct and visually boring, and a lot of first-timers conclude the model is bad. That's not a bug; it's the design. Budget a prompt that names a style or artist, or accept the plain look.

And the standing billing note: every run costs Buzz, and Anima's 30-step default isn't the cheapest job. One image first, then decide if you want a batch.