Kirazuri (Anima) 3.0 is a full fine-tune of the
Anima Base v1.0 model by CircleStone Labs 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
- 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
Reference the
Anima Base 1 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 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
