Anima Base

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 (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 is best for anime, manga, and illustrated characters/styles (not realism).

Key traits:

• Accepts Danbooru-style tags AND/OR natural language in the same prompt.
• Very low VRAM — generates (and trains) on <6GB VRAM; runs on any PC that can run SDXL/Illustrious.
• License: CircleStone Labs Non-Commercial License (with NVIDIA Open Model License terms on the weights/derivatives). Generated images are usable commercially per the model card — verify the current license text before relying on this.

Loaded in ComfyUI with standard split-file loaders (not a single checkpoint):

Verified from the pack's workflow JSON: CLIPLoader widget values are ["qwen_3_06b_base.safetensors", "stable_diffusion", "default"]. (The HF model card describes standard loaders; the exact CLIP type string stable_diffusion is 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/)

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.

The DWPreprocessor/DepthAnythingV2 aux models (dw-ll_ucoco_384_bs5.torchscript.pt, yolox_l.onnx, depth_anything_v2_vitl.pth) are auto-fetched by comfyui_controlnet_aux 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)

Applied 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 }
}

Other LoRAs the pack ships (toggle on/off 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 (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]; output: patched MODEL.

{
  "class_type": "AnimaLLLiteApply",
  "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
  }
}

Settings

The base model and the turbo-LoRA path want different settings:

Sampler character (from model card): er_sde = neutral style, flat colors, sharp lines; euler_ancestral = softer/thinner lines; dpmpp_2m_sde_gpu = similar with more variety. Optional beta57 scheduler for painterly looks.

Resolutions

The base model supports 512²–1536². The pack recommends these to avoid distortion:

Prompt Style

Anima accepts Danbooru tags + 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.

Structure: quality tags → subject/count tags → optional @artist name → anime style tags → 2–4 natural-language sentences describing subject, outfit, pose, composition, background, lighting, 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/).

Negative prompt (recommended):

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" }}
}

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: load an image with a painted mask, VAEEncode it, apply SetLatentNoiseMask, patch the model with the inpainting LLLite (fed the same image + mask), then sample. The mask region is regenerated from the prompt while 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", "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
  }},
  "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" }}
}

Other LLLite ControlNets (same AnimaLLLiteApply node, swap lllite_name, feed a preprocessed control image; mask can be a full-white/blank mask when not inpainting):

• anima-lllite-pose-1.safetensors ← DWPreprocessor (OpenPose)
• anima-lllite-depth-1.safetensors ← DepthAnythingV2Preprocessor
• anima-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/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 → generates in <6GB VRAM (runs anywhere SDXL/Illustrious runs).
• 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) — would need a GGUF loader node (e.g. ComfyUI-GGUF), not included in this pack. Unverified against this workflow.

Troubleshooting

1. Weird/distorted images → use a recommended resolution (1024x1024, 896x1152, 832x1216, 768x1344, 640x1536).
2. Turbo result looks washed/flat → that's turbo at CFG 1; for max quality switch to base mode (drop turbo LoRA, 30–50 steps, CFG 4–5).
3. AnimaLLLiteApply missing → install ComfyUI-Anima-LLLite; it is not a standard ControlNet node.
4. CLIP loads but output is garbage → confirm CLIPLoader type is stable_diffusion and the file is qwen_3_06b_base.safetensors (the Qwen3-0.6B base, not the chat/edit Qwen models).
5. Inpainting ignores the mask → ensure both SetLatentNoiseMask AND the inpainting AnimaLLLiteApply receive the painted mask; encode the source image with VAEEncode (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.