| --- |
| base_model: |
| - circlestone-labs/Anima |
| license: other |
| license_name: circlestone-labs-non-commercial-license |
| license_link: LICENSE.md |
| tags: |
| - diffusion-single-file |
| - comfyui |
| --- |
| |
| <!-- <img src="montage.jpg"> --> |
|
|
| Kirazuri (Anima) 3.0 is a full fine-tune of the [Anima Base v1.0 model by CircleStone Labs](https://huggingface.co/circlestone-labs/Anima) focused on several goals: |
|
|
| - Learn new concepts/styles/characters past the base model dataset cutoff of 2025 September |
|
|
| - Enhance the model aesthetic guided by manually applied quality, aesthetic, and style tagging |
|
|
| - Improve rendering and understanding of fine-details through high-resolution training for 1024^2, 1280^2, and 1536^2 resolutions |
|
|
|
|
| # Versions |
|
|
| - Version 3.0 (Latest) |
| - For in-depth details of version 3.0 training and tooling, see: [Kirazuri (Anima) 3.0 Training Diary](https://github.com/motimalu/diffusion-training-configs/blob/main/diffusion-pipe/anima/notes/kirazuri3.0-notes.md) |
| - An int8 convrot quantization model anima-kirazuri-v3-int8convrot is also provided |
| - Created with the library https://github.com/silveroxides/convert_to_quant |
| - Observed a ~20% generation speed increase on Blackwell with minimal degradation |
|
|
| ## Training Details Summary |
|
|
| Trainer: diffusion-pipe commit b0aa4f1e03169f3280c8518d37570a448420f8be |
|
|
| Training device: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
|
|
| Total training time: ~10 days |
|
|
| Total samples seen(unbatched steps): ~2,550,000 |
|
|
| Training resolutions: |
|
|
| - 512^2 |
| - 768^2 |
| - 1024^2 |
| - 1280^2 |
| - 1536^2 |
|
|
| ### Stage 1 |
| - Samples seen(unbatched steps): ~2,000,000 |
| - Training time: ~125 hrs |
| - Learning Rate: 6e-6 |
| - Learning Rate Scheduler: Cosine |
| - LLM Adaptor Learning Rate: 8e-7 |
| - Precision: Mixed BF16 |
| - Optimizer: AdamW8bit with Kahan Summation |
| - Weight Decay: 0.01 |
| - Timestep Sampling Strategy: Logit-Normal |
|
|
| ## Stage 2 |
| - Samples seen(unbatched steps): ~550,000 |
| - Training time: ~118 hrs |
| - Learning Rate: 3e-6 |
| - Learning Rate Scheduler: Cosine |
| - LLM Adaptor Learning Rate: 0 |
| - Flux Shift: Enabled |
| - Multi-Scale Loss Weight: 0.5 |
| - Precision: Mixed BF16 |
| - Optimizer: AdamW8bit with Kahan Summation |
| - Weight Decay: 0.01 |
| - Timestep Sampling Strategy: Logit-Normal |
|
|
| ### Additional Features |
| Tag Dropout: 30% with protected first 8 tags |
| Tag Shuffle: Applied to last unprotected tags |
| Natural Language: Short and Long Caption variants |
|
|
| ### Changes from Kirazuri (Anima) v2.0 |
| Dataset includes recently curated 7,071 images increasing total size from 35,537 to 42,608 images |
| Dataset cutoff now of 2026/05/12. |
| Trained at 5 total resolutions in two-stage training |
| Stage 1 - 512^2, 768^2, 1024^2 |
| Stage 2 - 1024^2, 1280^2 1536^2 |
| Introduced cosine learning rate scheduler for smooth learning rate transition between training stages |
| Re-captioned full dataset for a second natural language captions variant with updated captioning script |
|
|
| ### Installing and running |
|
|
| Workflow: |
| <img src="example.png" width="200"> |
|
|
| Reference the [Anima Base 1](https://huggingface.co/circlestone-labs/Anima) instructions. The model is natively supported in ComfyUI. The above image contains a workflow; you can open it in ComfyUI or drag-and-drop to get the workflow. |
|
|
| Note: Most preview images on the model card additionally use the [comfyui-prompt-control](https://github.com/asagi4/comfyui-prompt-control) node for schedule prompting syntax to mix concepts i.e. `[word1|word2]` |
| This custom node is entirely optional but required to exactly recreate the outputs in ComfyUI. |
|
|
| The model files go in their respective folders inside your model directory: |
| - anima-kirazuri-v3.safetensors (this model) goes in ComfyUI/models/diffusion_models |
| - qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders |
| - qwen_image_vae.safetensors goes in ComfyUI/models/vae (this is the Qwen-Image VAE, you might already have it) |
| |
| ## Generation Settings |
| - Trained in mixed resolutions for the majority of training, and finished with dedicated high resolution training. |
| - Previews are generated mostly at 1280^2 e.g. 1520x1040 or 1536^2 e.g. 1248x1824 resolutions. |
| - 30-50 steps, CFG 4-5. |
| - Same samplers as recommended for the base model work, I like to use: |
| - er_sde: the recommended default for 30-50 steps. |
| - sa_solver_pece: can converge with good detail in 15-20 steps. |
|
|
| ## Prompting |
|
|
| Like the base model, this model is trained on booru-style tags, natural language captions, and combinations of tags and captions. |
|
|
| ### Tag order |
| [quality/meta/safety tags] [character] [series] [artist] [1girl/1boy/1other etc] [general tags] |
|
|
| Mostly the same order as the base model, only the [1girl/1boy/other etc] groups position is towards the end in this models dataset. |
|
|
| [quality/meta/safety tags] [character] [series] [artist] tag groups are also not shuffled, so their order may have some influence on generations. |
|
|
| ### Quality and Aesthetic tags |
|
|
| Human score based: `masterpiece`, `best quality`, `very aesthetic`, `aesthetic` |
|
|
| The `very aesthetic` and `aesthetic` tags are where this model diverges from the base, with the intent these can be used to guide the model toward a different aesthetic - a kind of house model bias. |
|
|
| ## Meta tags |
|
|
| absurdres, official art, etc |
|
|
| ## Styles |
|
|
| painterly, chiaroscuro, ligne claire, flat color, no lineart, blending, etc |
|
|
| traditional media, oil painting \(medium\), watercolor \(medium\), etc |
|
|
| [Optional] ComfyUI-Autocomplete-Plus prompt input assistance |
|
|
| An optional file danbooru_tags_kirazuri_3.txt is included with the version 3.0 model details. |
| |
| This file contains metadata that is derived from public sources for prompt assistance only, and is intended to be used with the ComfyUI-Autocomplete-Plus extension. |
| |
| Rename the file to danbooru_tags_kirazuri_3.csv and place it in your ComfyUI/custom_nodes/comfyui-autocomplete-plus/data directory. |
| |
| ## Known Limitations & Issues: |
| |
| Some concept bleeding and instability is noticeable when using short prompts, especially tag-only prompts. |
| |
| Longer tag strings and natural language prompts describing the image in detail should help with this. |
| |
| This reflects how the model was trained with a combination of natural language and tags. |
| |
| ## Recognitions |
| |
| - Thanks to CircleStone Labs for the Anima base models. |
| - Thanks to tdrussell of CircleStone Labs for the diffusion-pipe trainer. |
| - Thanks to bluvoll for support using their fork of diffusion-pipe. |
| - Thanks to narugo1992 and the deepghs team for open-sourcing various training sets, image processing tools, and models. |
| |
| ## License |
| |
| This model is a derivative work licensed under the CircleStone Labs Non-Commercial License. |
| |
| See the base model for details of the CircleStone Labs Non-Commercial License. |
| |
| Built on NVIDIA Cosmos |
