Johnny-Z/Anima-Light-Lavender

English | 中文

TL;DR: A major boost to natural-language understanding at the 512-token scale; prompt with a long natural-language description by default (the image_description field). We recommend following the "Quick Start" guide below and assembling the structured caption with comfyui-zako-pe; if you don't want to write a long description, you can also provide just tags and let zako-pe expand them.

📌 Overview

Anima-Light-Lavender is a version obtained by post-training on top of Anima-Base v1.0 (anima-base-v1.0.safetensors) from the circlestone-labs/Anima series. Compared with the base model, this release focuses on improving the model's understanding of natural-language descriptions at the 512-token scale, and leverages higher-quality training data to improve the overall generation quality.

This release makes no architectural changes: the layer count, parameter count, and file layout are all identical to the base model, so you can swap in the upstream weights directly and plug it into existing workflows.

Item	Description
Base model	Anima-Base v1.0 (anima-base-v1.0.safetensors) from the circlestone-labs/Anima series, built on nvidia/Cosmos-Predict2-2B-Text2Image
Parameters / Architecture	Identical to the base model (~2B parameters, a 28-layer DiT); no added layers, no distillation
Task type	text-to-image: anime illustrations, characters, and stylized art
Text encoder	Qwen3-0.6B (qwen_3_06b_base.safetensors)
VAE	Qwen-Image VAE (qwen_image_vae.safetensors)
Weight files	anima-light-lavender.safetensors (single file, BF16); an MXFP8 version anima-light-lavender_mxfp8.safetensors is also available (single file, faster inference)
Training data	~1.7M Danbooru images annotated and filtered by an in-house data pipeline; all models used in the pipeline — annotation, filtering, preference judgment, and other steps — were trained independently by the author
Training data cutoff	End of November 2025
License	CircleStone Labs Non-Commercial License

🎯 Use Cases

Best for ✅

• Anime illustrations and character art: for generating images with a specific art style, composition, lighting, and atmosphere.
• Natural-language driven: for precise control over the image with a descriptive passage (rather than stacked tags).
• Drop-in upgrade: for users already running an Anima workflow — switch the checkpoint directly to improve generation quality.

Not suitable for ⛔

• Photo-realism: photorealistic photography is not the goal of this series; results are limited.
• Long-text rendering: hard to faithfully reproduce text on signs, subtitles, or long multi-line sentences.

🚀 Quick Start (ComfyUI)

The Anima series is natively supported in ComfyUI, and this release can directly replace the upstream weights.

1. Place the Model Files

File	Target directory
anima-light-lavender.safetensors	ComfyUI/models/diffusion_models
anima-light-lavender_mxfp8.safetensors	ComfyUI/models/diffusion_models
qwen_3_06b_base.safetensors	ComfyUI/models/text_encoders
qwen_image_vae.safetensors	ComfyUI/models/vae

Weights are provided in two versions, BF16 and MXFP8 — pick either one and place it into diffusion_models; the MXFP8 version speeds up inference.

(The last two files are the Qwen3-0.6B text encoder and Qwen-Image VAE used by upstream Anima; if you already run an Anima workflow, you can reuse them as-is.)

2. Install the Custom Nodes

Install the comfyui-zako-pe custom nodes: clone the repository into ComfyUI/custom_nodes/ and restart ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/aa0525/comfyui-zako-pe.git

If git is not installed, you can also click Code → Download ZIP on the repository page, extract it to ComfyUI/custom_nodes/, and restart ComfyUI for it to take effect.

The node pack provides Danbooru Caption JSON and Danbooru Prompt Extend (OpenAI): the former assembles the structured caption, while the latter calls an OpenAI-compatible service to expand a tag-style image_description into natural language. For the expansion service, we recommend the companion model zako-pe (ZAKO-V0.1): load its GGUF weights in LM Studio and start the local server (default address http://127.0.0.1:1234) to use it; see its model card for downloads and detailed usage.

3. Load the Workflows

This model repository ships with two workflow files; drag them into the ComfyUI canvas to load them — no manual wiring needed:

Workflow file	Description
anima.json	Basic workflow: feeds the assembled structured caption directly into the positive prompt encoding; write image_description yourself as a long natural-language description
anima-pe.json	Adds a Danbooru Prompt Extend (OpenAI) node on top of the basic workflow: write only tags in image_description, and the expansion service rewrites them into natural language automatically

Both workflows load the MXFP8 weights (anima-light-lavender_mxfp8.safetensors) by default; if you only downloaded the BF16 version, switch to anima-light-lavender.safetensors in the Load Diffusion Model node.

4. Generation Parameters (Recommended)

Parameter	Recommended Value
Sampler	euler
Scheduler	simple
Steps	25
CFG	4.0
Resolution	Around 1280×1280 pixels
Negative prompt	Leave empty

✍️ Prompting

Structured Caption

This release is trained on structured captions, so it can understand text blocks with a fixed field order:

{
  "year": 2025,
  "preference_level": "best",
  "artist": [...],
  "copyright": [...],
  "character": [...],
  "image_description": "...",
  "extra_tags": [...]
}

Field	Purpose and usage
year	Year anchor (integer, default 2025)
preference_level	Preference level: normal / high / very_high / best (default best)
artist	Artist / style tags
copyright	Series / franchise tags
character	Character tags
image_description	Image content description: a natural-language description, as detailed as possible
extra_tags	Supplementary tags for the image content

🧠 Training Details

Data

• ~1.7M Danbooru images, all annotated and filtered by an in-house data pipeline; training data goes up to the end of November 2025.
• All models involved in the data pipeline were trained independently by the author: from image annotation and quality filtering to the preference judgment mentioned below, every model used was trained by the author rather than an off-the-shelf third-party model.
• Uses the structured caption shown above, in which image_description is pure natural language.
• preference_level tiers are determined by a preference judgment model trained on ~3M samples.

Training Dynamics

Item	Setting
Precision	Training runs in BF16 overall; matrix multiplications in the MLP and Attention use MXFP8 GEMM
Optimizer	Composite optimizer: 2D parameters of the MLP and Attention are updated by Muon (momentum 0.95, match_rms_adamw); the remaining parameters stay in BF16 and are updated by AdamW (β = 0.9 / 0.95, ε = 1e-8)
Batch size	1024
Learning rate	4e-5
Weight decay	1e-3
Gradient clipping	1.0

Training Focus

• Substantially enhances understanding of natural-language descriptions at the 512-token scale, making precise description — rather than rolling random tags — the default way to generate.
• The improved data quality also improves the aesthetics of generated images: even without adding quality words or filling in a negative prompt, the results are still clean and tidy enough.

🔌 Compatibility

• The architecture is fully identical to the upstream base: the layer count, parameter count, and file format are all the same, so the checkpoint can be replaced directly.
• Uses the same text encoder and VAE (Qwen3-0.6B text encoder, Qwen-Image VAE) — no workflow changes needed.
• Single-file weights: put them into models/diffusion_models and ComfyUI can load them.

⚠️ Limitations

• Not good at realism: the model targets anime illustrations and non-photorealistic art; the photorealistic photography style is out of its scope.
• Weak text rendering: single words and short phrases usually render correctly, but the rendering of long text is unreliable.

📜 License

The model weights and any derivatives are subject to the CircleStone Labs Non-Commercial License and are for non-commercial use only. In addition, this model is a Derivative Model of Cosmos-Predict2-2B-Text2Image and is therefore also subject to the corresponding terms of the NVIDIA Open Model License Agreement.

🙏 Acknowledgements

• Base model: Anima-Base v1.0 (anima-base-v1.0.safetensors) from the circlestone-labs/Anima series, built on nvidia/Cosmos-Predict2-2B-Text2Image.