DiffSynth: Anima Inference

Prompt in, image out. The interesting part is what the model accepts: Anima is the anime base where 1girl, solo, artist_name, cherry blossoms, soft lighting and "a girl leaning against a fence, late afternoon, warm rim light on her hair" are the same sentence as far as the encoder is concerned. You can mix them - tags for the specific things Danbooru indexes well, prose for spatial relationships and mood.

Required: pipe (from the Anima Loader), prompt, seed, num_inference_steps (default 30), height, width (1024). One output: image → Save Image.

The settings that actually change output

• cfg_scale - default 4, and Anima's model card recommends 4-5 on the base checkpoint. This matters: Anima's base is not one of the guidance-distilled models where CFG 1 is the target, so your negative prompt does real work here. Load the Turbo LoRA or run a distilled variant and that flips to CFG 1, at which point the negative box is inert - ComfyUI won't even compute the unconditional pass, it just discards what you typed.
• num_inference_steps - 30 is the node default; the shipped template uses 50. Anima is slow per step, so pick deliberately: 30 for iteration, 50 when you've found the one.
• input_image + denoising_strength - the img2img path. Since there's no ControlNet for Anima, this is your only structural conditioning, and it's how you do the second pass: render at 1024, then feed the result back in at 1.25× size with denoising_strength around 0.3-0.4. That's the in-node equivalent of the latent-upscale trick Anima users lean on to fix its soft default output.
• sigma_shift - a flow-matching schedule shift, default 0. Change it if you're chasing a specific look; the templates never touch it.
• negative_prompt - worth a short, real list at CFG 4. Don't stack forty tokens of legacy SDXL boilerplate; Anima isn't an SDXL finetune and doesn't respond to it the way you'd expect.

Prompting, in one paragraph

Anima was trained with Pony-style score tags and the card's order is quality/meta/year/safety, then character count, character, series, artist, then general tags. masterpiece, best quality, score_7, safe is the recommended prefix. The community's better advice: drop the quality prefix the moment you're naming a specific artist or unusual style, because score tags pull output toward a generic polished look. And expect the first ten prompts to disappoint - Anima's default aesthetic is deliberately plain, and the fix is spending an hour finding artist tags you like, not adding more prompts. People who skip that step call the model bad, then reverse the verdict.

Install

# Manager: search "DiffSynth-ComfyUI" → Install → restart
pip install git+https://github.com/modelscope/DiffSynth-Studio.git
cd ComfyUI/custom_nodes && git clone https://github.com/modelscope/DiffSynth-ComfyUI.git

Fastest route to a working graph: Template tab → DiffSynth Anima workflow, which ships with this node, the Loader, the tokenizer ModelConfigs and a VRAM Limit already wired.

Troubleshooting

Everything is slow - that's expected, Anima runs high per-step compute for its size, so resist the urge to "fix" it by dropping to 8 steps; you'll get mush. Native resolution is 512-1536 in practice, and 2048 renders but holds together less consistently. If a multi-character prompt bleeds one design into another, that's a documented limitation: characters with similar traits collide (two witches with different hats). And if you're coming from a workflow that expects DiffSynth: Anima Inference to have a LoRA socket - it doesn't; LoRAs go on the pipe through the separate LoRA Clear / LoRA Load nodes before you reach this one.