计划执行与检查节点
executing-plans
按既定实施计划完成工作，保持过程检查和反馈。
探索技能
Custom nodes (git clone into
效率与协作 · SKILL.md
anima-base
anima-base【效率与协作】Anime/illustration text-to-image (ANIMA 1.0, ~2B Cosmos DiT). Use for anime, manga, illustrated characters; accepts Danbooru tags + natural language; runs/trains on <6GB VRAM; includes anime inpainting via Anima-LLLite ControlNet
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
技能说明
GitHubname: anima-base description: Anime/illustration text-to-image (ANIMA 1.0, ~2B Cosmos DiT). Use for anime, manga, illustrated characters; accepts Danbooru tags + natural language; runs/trains on <6GB VRAM; includes anime inpainting via Anima-LLLite ControlNet globs:
- "**/*.json"
ANIMA 1.0 (Anima Base Ultra) Text-to-Image Workflows
Overview
Anima is a ~2B-parameter anime / illustration text-to-image base model from CircleStone Labs, made in collaboration with Comfy Org. It is not SDXL-lineage. The architecture is NVIDIA Cosmos-Predict2-2B-Text2Image (a DiT / flow model), trained on several million anime images plus ~800k non-anime artistic images. It suits anime, manga, and illustrated characters and styles, not realism.
Key traits:
- Accepts Danbooru-style tags and/or natural language in the same prompt.
- Very low VRAM. It generates and trains on <6GB VRAM and runs on any PC that can run SDXL/Illustrious.
- License: CircleStone Labs Non-Commercial License, with NVIDIA Open Model License terms on the weights and derivatives. Generated images are usable commercially per the model card. Verify the current license text before relying on this.
ComfyUI loads it with standard split-file loaders, not a single checkpoint:
| Component | Node | Model file | Folder | Notes |
|---|---|---|---|---|
| Diffusion model | UNETLoader | anima-base-v1.0.safetensors | models/diffusion_models/ | weight_dtype default; ~4GB fp |
| Text encoder | CLIPLoader | qwen_3_06b_base.safetensors | models/text_encoders/ | Qwen3-0.6B base; type": "stable_diffusion" in this pack |
| VAE | VAELoader | qwen_image_vae.safetensors | models/vae/ | Qwen-Image VAE (~254MB) |
Verified from the pack's workflow JSON:
CLIPLoaderwidget values are["qwen_3_06b_base.safetensors", "stable_diffusion", "default"]. The HF model card describes standard loaders; the exact CLIPtypestringstable_diffusionis what the Aitrepreneur "Anima Base Ultra" workflow ships. Use it as-is.
Installation
The "Anima Base Ultra" pack (by Aitrepreneur) installs custom nodes and downloads all models. Models are mirrored on https://huggingface.co/Aitrepreneur/FLX/resolve/main; the official source is https://huggingface.co/circlestone-labs/Anima.
Custom nodes (git clone into ComfyUI/custom_nodes/)
| Node pack | Repo | Used for |
|---|---|---|
| ComfyUI-Manager | https://github.com/ltdrdata/ComfyUI-Manager.git | management |
| ComfyUI-Impact-Pack | https://github.com/ltdrdata/ComfyUI-Impact-Pack | FaceDetailer / EditDetailerPipe |
| ComfyUI-Impact-Subpack | https://github.com/ltdrdata/ComfyUI-Impact-Subpack | UltralyticsDetectorProvider |
| rgthree-comfy | https://github.com/rgthree/rgthree-comfy | Power Lora Loader, Fast Groups, Any Switch |
| ComfyUI-KJNodes | https://github.com/kijai/ComfyUI-KJNodes | helpers |
| ComfyUI_UltimateSDUpscale | https://github.com/ssitu/ComfyUI_UltimateSDUpscale | tiled upscaling |
| ComfyUI_tinyterraNodes | https://github.com/TinyTerra/ComfyUI_tinyterraNodes | ttN seed |
| comfyui_controlnet_aux | https://github.com/Fannovel16/comfyui_controlnet_aux | DWPreprocessor, DepthAnythingV2 |
| ComfyUI-Anima-LLLite | https://github.com/kohya-ss/ComfyUI-Anima-LLLite | AnimaLLLiteApply_sdscripts (ControlNet + inpainting) |
Models (download URLs from the pack's .bat / .sh)
Base $HF = https://huggingface.co/Aitrepreneur/FLX/resolve/main, $YOLO11 = https://huggingface.co/Ultralytics/YOLO11/resolve/main. Append ?download=true.
| Folder | File | Source |
|---|---|---|
diffusion_models/ | anima-base-v1.0.safetensors | $HF |
text_encoders/ | qwen_3_06b_base.safetensors | $HF |
vae/ | qwen_image_vae.safetensors | $HF |
controlnet/ | anima-lllite-inpainting-v1.safetensors | $HF |
controlnet/ | anima-lllite-depth-1.safetensors | $HF |
controlnet/ | anima-lllite-lineart-1.safetensors | $HF |
controlnet/ | anima-lllite-pose-1.safetensors | $HF |
controlnet/ | anima-lllite-any-test-like-1-step2000.safetensors | $HF |
loras/ | anima-turbo-lora-v0.1.safetensors | $HF |
loras/ | anima-highres-aesthetic-boost.safetensors | $HF |
loras/ | anima-preview-3-masterpieces-v5.safetensors | $HF |
loras/ | anima_p3_rdbt_v0.29.b.122.safetensors | $HF |
upscale_models/ | 4x_foolhardy_Remacri.pth, 4x-ClearRealityV1.pth | $HF |
ultralytics/bbox/ | face_yolov9c.pt, hand_yolov9c.pt, Eyeful_v2-Paired.pt | $HF |
ultralytics/segm/ | ntd11_anime_nsfw_segm_v5-variant1.pt | $HF |
ultralytics/segm/ | yolo11m-seg.pt | $YOLO11 |
sams/ | sam_vit_b_01ec64.pth | $HF |
comfyui_controlnet_aux fetches the DWPreprocessor/DepthAnythingV2 aux models (dw-ll_ucoco_384_bs5.torchscript.pt, yolox_l.onnx, depth_anything_v2_vitl.pth) on first use.
Key Nodes
Loaders
{
"1": { "class_type": "UNETLoader", "inputs": { "unet_name": "anima-base-v1.0.safetensors", "weight_dtype": "default" }},
"2": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_3_06b_base.safetensors", "type": "stable_diffusion", "device": "default" }},
"3": { "class_type": "VAELoader", "inputs": { "vae_name": "qwen_image_vae.safetensors" }}
}
Anima Turbo LoRA (the shipped default — 12-step fast mode)
The pack applies it with rgthree Power Lora Loader. The plain ComfyUI equivalent is LoraLoaderModelOnly:
{
"class_type": "LoraLoaderModelOnly",
"inputs": { "model": ["1", 0], "lora_name": "anima-turbo-lora-v0.1.safetensors", "strength_model": 1.0 }
}
The pack ships three other LoRAs you can toggle in Power Lora Loader: anima-highres-aesthetic-boost, anima-preview-3-masterpieces-v5, anima_p3_rdbt_v0.29.b.122. In the non-turbo groups these three are enabled and the turbo LoRA is off; in the turbo groups only the turbo LoRA is on.
AnimaLLLiteApply_sdscripts (ControlNet + inpainting — from ComfyUI-Anima-LLLite)
Patches the MODEL. Anima uses LLLite-style control, not standard ControlNetApply conditioning. Inputs: model, image, mask; widget order [lllite_name, strength, start_percent, end_percent, preserve_wrapper]; output: patched MODEL. ComfyUI core now owns the old ID AnimaLLLiteApply (different signature: a MODEL_PATCH from ModelPatchLoader, no mask), so this pack uses the kohya-ss node ID AnimaLLLiteApply_sdscripts.
{
"class_type": "AnimaLLLiteApply_sdscripts",
"inputs": {
"model": ["<model>", 0],
"image": ["<control_or_source_image>", 0],
"mask": ["<mask>", 0],
"lllite_name": "anima-lllite-pose-1.safetensors",
"strength": 1.0, "start_percent": 0.0, "end_percent": 1.0,
"preserve_wrapper": true
}
}
Settings
The base model and the turbo-LoRA path want different settings:
| Mode | Steps | CFG | Sampler | Scheduler | Denoise | Notes |
|---|---|---|---|---|---|---|
| Base (no turbo LoRA) | 30–50 | 4–5 | er_sde | simple | 1.0 | Author-recommended for the base model |
| Turbo LoRA (shipped default) | 12 | 1.0 | er_sde | simple | 1.0 | anima-turbo-lora-v0.1 enabled |
| Upscale pass (UltimateSDUpscale) | 12 | 1.0 | er_sde | simple | 0.28 | 4x_foolhardy_Remacri.pth, scale 2x |
Sampler character, from the model card: er_sde gives a neutral style, flat colors, sharp lines; euler_ancestral gives softer, thinner lines; dpmpp_2m_sde_gpu is similar with more variety. The optional beta57 scheduler gives painterly looks.
Resolutions
The base model supports 512² to 1536². The pack recommends these to avoid distortion:
| Aspect | Resolution |
|---|---|
| 1:1 | 1024x1024 |
| 3:4 | 896x1152 |
| 5:8 | 832x1216 |
| 9:16 | 768x1344 |
| 9:21 | 640x1536 |
Prompt Style
Anima accepts Danbooru tags and natural language together. The pack's recommended formula:
masterpiece, best quality, score_7, safe, highres, official art,
1girl, solo,
@artist name,
clean lineart, detailed eyes, soft shading,
A young anime woman with long silver hair and blue eyes stands in a rainy neon city at night.
She wears a black futuristic jacket with glowing blue details. Medium close-up, wet pavement
reflections, soft background blur, cinematic lighting.
The order is quality tags, then subject/count tags, then an optional @artist name, then anime style tags, then 2 to 4 natural-language sentences describing subject, outfit, pose, composition, background, lighting, and mood. Use lowercase tags with spaces (not underscores), except score tags like score_7. Artist tags use @artist name; browse names at the community Anima Style Explorer (https://thetacursed.github.io/Anima-Style-Explorer/).
Recommended negative prompt:
worst quality, low quality, score_1, score_2, score_3, artist name, bad anatomy, bad hands,
missing fingers, extra fingers, extra arms, extra legs, duplicate, twins, text, watermark,
signature, simple background
Unlike Flux/Qwen, Anima does use a real negative prompt via a second CLIPTextEncode (CFG > 1 in base mode).
Complete Workflow: Text-to-Image (Turbo, 12-step)
{
"1": { "class_type": "UNETLoader", "inputs": { "unet_name": "anima-base-v1.0.safetensors", "weight_dtype": "default" }},
"2": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_3_06b_base.safetensors", "type": "stable_diffusion", "device": "default" }},
"3": { "class_type": "VAELoader", "inputs": { "vae_name": "qwen_image_vae.safetensors" }},
"4": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "anima-turbo-lora-v0.1.safetensors", "strength_model": 1.0 }},
"5": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["2", 0], "text": "masterpiece, best quality, score_7, safe, highres, official art, 1girl, solo, clean lineart, detailed eyes, soft shading,\n\nA young anime woman with long silver hair and blue eyes stands in a rainy neon city at night, cinematic lighting." }},
"6": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["2", 0], "text": "worst quality, low quality, score_1, score_2, score_3, bad anatomy, bad hands, extra fingers, text, watermark, signature, simple background" }},
"7": { "class_type": "EmptyLatentImage", "inputs": { "width": 896, "height": 1152, "batch_size": 1 }},
"8": { "class_type": "KSampler", "inputs": {
"model": ["4", 0],
"positive": ["5", 0],
"negative": ["6", 0],
"latent_image": ["7", 0],
"seed": 42, "steps": 12, "cfg": 1, "sampler_name": "er_sde", "scheduler": "simple", "denoise": 1
}},
"9": { "class_type": "VAEDecode", "inputs": { "samples": ["8", 0], "vae": ["3", 0] }},
"10": { "class_type": "SaveImage", "inputs": { "images": ["9", 0], "filename_prefix": "anima" }}
}
For the base-quality variant (no turbo), drop node 4 (feed ["1", 0] into KSampler), set steps: 30, cfg: 4.5. Optionally enable the three quality LoRAs (anima-highres-aesthetic-boost, anima-preview-3-masterpieces-v5, anima_p3_rdbt_v0.29.b.122) by chaining LoraLoaderModelOnly nodes.
Complete Workflow: Anime Inpainting (Anima-LLLite ControlNet)
The pack's "INPAINTING CONTROLNET" group loads an image with a painted mask, VAEEncodes it, applies SetLatentNoiseMask, patches the model with the inpainting LLLite (fed the same image and mask), then samples. The mask region is regenerated from the prompt and the rest is preserved.
{
"1": { "class_type": "UNETLoader", "inputs": { "unet_name": "anima-base-v1.0.safetensors", "weight_dtype": "default" }},
"2": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_3_06b_base.safetensors", "type": "stable_diffusion", "device": "default" }},
"3": { "class_type": "VAELoader", "inputs": { "vae_name": "qwen_image_vae.safetensors" }},
"4": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "anima-turbo-lora-v0.1.safetensors", "strength_model": 1.0 }},
"5": { "class_type": "LoadImage", "inputs": { "image": "<masked_image.png>" }},
"6": { "class_type": "AnimaLLLiteApply_sdscripts", "inputs": {
"model": ["4", 0], "image": ["5", 0], "mask": ["5", 1],
"lllite_name": "anima-lllite-inpainting-v1.safetensors",
"strength": 1.0, "start_percent": 0.0, "end_percent": 1.0,
"preserve_wrapper": true
}},
"7": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["2", 0], "text": "<what to paint into the masked area>" }},
"8": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["2", 0], "text": "worst quality, low quality, bad anatomy, text, watermark" }},
"9": { "class_type": "VAEEncode", "inputs": { "pixels": ["5", 0], "vae": ["3", 0] }},
"10": { "class_type": "SetLatentNoiseMask", "inputs": { "samples": ["9", 0], "mask": ["5", 1] }},
"11": { "class_type": "KSampler", "inputs": {
"model": ["6", 0],
"positive": ["7", 0],
"negative": ["8", 0],
"latent_image": ["10", 0],
"seed": 42, "steps": 12, "cfg": 1, "sampler_name": "er_sde", "scheduler": "simple", "denoise": 1
}},
"12": { "class_type": "VAEDecode", "inputs": { "samples": ["11", 0], "vae": ["3", 0] }},
"13": { "class_type": "SaveImage", "inputs": { "images": ["12", 0], "filename_prefix": "anima_inpaint" }}
}
The other LLLite ControlNets use the same AnimaLLLiteApply_sdscripts node. Swap lllite_name and feed a preprocessed control image; the mask can be a full-white/blank mask when not inpainting:
anima-lllite-pose-1.safetensors←DWPreprocessor(OpenPose)anima-lllite-depth-1.safetensors←DepthAnythingV2Preprocessoranima-lllite-lineart-1.safetensors/anima-lllite-any-test-like-1-step2000.safetensors← lineart / generic control
Upscaling (optional)
The pack upscales with UltimateSDUpscale (4x_foolhardy_Remacri.pth, 2x, denoise 0.28, 12 steps, er_sde/simple) and refines faces, hands, and eyes with Impact-Pack FaceDetailer driven by UltralyticsDetectorProvider (face_yolov9c.pt, hand_yolov9c.pt, Eyeful_v2-Paired.pt) + SAM (sam_vit_b_01ec64.pth).
VRAM
- Anima is ~2B params, so it generates in <6GB VRAM and runs anywhere SDXL/Illustrious runs.
- The text encoder (Qwen3-0.6B) and VAE are both small.
- A GGUF quantized build exists for even lower memory (
Abiray/Anima-base-v1.0-GGUF). It needs a GGUF loader node (e.g. ComfyUI-GGUF), which this pack does not include. Unverified against this workflow.
Troubleshooting
- Weird/distorted images. Use a recommended resolution (1024x1024, 896x1152, 832x1216, 768x1344, 640x1536).
- Turbo result looks washed/flat. That's turbo at CFG 1. For max quality switch to base mode (drop turbo LoRA, 30 to 50 steps, CFG 4 to 5).
AnimaLLLiteApply_sdscriptsmissing. InstallComfyUI-Anima-LLLite; it is not a standard ControlNet node. Do not add coreAnimaLLLiteApply— that ID now belongs to ComfyUI and has a different input signature.- CLIP loads but output is garbage. Confirm
CLIPLoadertypeisstable_diffusionand the file isqwen_3_06b_base.safetensors(the Qwen3-0.6B base, not the chat/edit Qwen models). - Inpainting ignores the mask. Ensure both
SetLatentNoiseMaskand the inpaintingAnimaLLLiteApply_sdscriptsreceive the painted mask, and encode the source image withVAEEncode. Denoise 1.0 is fine because the noise mask preserves unmasked pixels.
Training custom LoRAs
To train your own Anima LoRA (character/style) on <6GB VRAM, use the Citron Anima LoRA Trainer; see the anima-lora-trainer skill. Trained .safetensors LoRAs drop into models/loras/ and load via Power Lora Loader / LoraLoaderModelOnly exactly like the bundled LoRAs above.
Sources
- Official: model weights at https://huggingface.co/circlestone-labs/Anima; ComfyUI-Anima-LLLite README documents the node ID
AnimaLLLiteApply_sdscriptsafter the coreAnimaLLLiteApplycollision. No vendor prompting guide cited. - Empirical: tag-order / @artist prompting and sampler wiring from working graphs; Anima Style Explorer is community, not vendor docs.