English | δΈζ
> **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](https://huggingface.co/Johnny-Z/zako-pe) expand them. ## π 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](https://huggingface.co/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](https://huggingface.co/circlestone-labs/Anima) series, built on [nvidia/Cosmos-Predict2-2B-Text2Image](https://huggingface.co/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 β annotation, filtering, preference judgment, and other steps β were trained independently by the author** | | Training data cutoff | End of November 2025 | | License | CircleStone Labs Non-Commercial License | ## π― Use Cases **Best for β ** - **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 β switch the checkpoint directly to improve generation quality. **Not suitable for β** - **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. ## π 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** β 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](https://github.com/aa0525/comfyui-zako-pe) custom nodes: clone the repository into `ComfyUI/custom_nodes/` and restart ComfyUI: ```bash cd ComfyUI/custom_nodes git clone https://github.com/aa0525/comfyui-zako-pe.git ``` If git is not installed, you can also click **Code β 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](https://huggingface.co/Johnny-Z/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 β 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 1280Γ1280 pixels | | Negative prompt | **Leave empty** | ## βοΈ Prompting ### Structured Caption This release is trained on structured captions, so it can understand text blocks with a fixed field order: ```text { "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 | ## π§ 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** (Ξ² = 0.9 / 0.95, Ξ΅ = 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 β rather than rolling random tags β 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. ## π 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) β no workflow changes needed. - Single-file weights: put them into `models/diffusion_models` and ComfyUI can load them. ## β οΈ 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. ## π License The model weights and any derivatives are subject to the [CircleStone Labs Non-Commercial License](https://huggingface.co/circlestone-labs/Anima/resolve/main/LICENSE.md) 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](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). ## π Acknowledgements - Base model: Anima-Base v1.0 (`anima-base-v1.0.safetensors`) from the [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima) series, built on [nvidia/Cosmos-Predict2-2B-Text2Image](https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image).