<h1 id="kirazuri-(anima)">Kirazuri (Anima)</h1>Kirazuri (Anima) is a full fine-tune of the <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima/blob/main/split_files/diffusion_models/anima-base-v1.0.safetensors">Anima Base v1.0 model by CircleStone Labs</a> focused on several goals:<ul><li>Learn new concepts/styles/characters past the base model dataset cutoff of 2025 September</li><li>Enhance the model aesthetic guided by manually applied quality, aesthetic, and style tagging</li><li>Improve rendering and understanding of fine-details through high-resolution training for 1024^2, 1280^2, and 1536^2 resolutions</li></ul><h2 id="version-4.0-(latest)">Version 4.0 (Latest)</h2>For in-depth details of training and tooling, see:<ul><li><a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/blob/main/diffusion-pipe/anima/notes/kirazuri4.0-notes%20.md">Kirazuri (Anima) 4.0 Training Diary</a></li><li><a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/blob/main/diffusion-pipe/anima/notes/kirazuri3.0-notes.md">Kirazuri (Anima) 3.0 Training Diary</a></li></ul><h3 id="training-details-summary">Training Details Summary</h3>Trainer: <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a>Training device: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionTotal training time: ~10 daysTotal samples seen(unbatched steps): ~3,000,000Training resolutions:<ul><li>512^2</li><li>768^2</li><li>1024^2</li><li>1536^2</li></ul><h3 id="stage-1">Stage 1</h3><ul><li>Samples seen(unbatched steps): ~2,000,000</li><li>Training time: ~125 hrs</li><li>Learning Rate: 6e-6</li><li>Learning Rate Scheduler: Cosine</li><li>LLM Adaptor Learning Rate: 8e-7</li><li>Precision: Mixed BF16</li><li>Optimizer: AdamW8bit with Kahan Summation</li><li>Weight Decay: 0.01</li><li>Timestep Sampling Strategy: Logit-Normal</li><li>Training Resolutions: 512^2, 768^2, 1024^2</li></ul><h3 id="stage-2">Stage 2</h3><ul><li>Samples seen(unbatched steps): ~1,000,000</li><li>Training time: ~84 hrs</li><li>Learning Rate: 2e-6</li><li>Learning Rate Scheduler: Cosine</li><li>LLM Adaptor Learning Rate: 2e-7</li><li>Precision: Mixed BF16</li><li>Optimizer: AdamW8bit with Kahan Summation</li><li>Weight Decay: 0.01</li><li>Timestep Sampling Strategy: Logit-Normal</li><li>Training Resolutions: 512^2, 1024^2, 1536^2</li></ul><h3 id="additional-features">Additional Features</h3><ul><li>Masked Training</li><li>Tag Dropout: 30% with protected first 8 tags</li><li>Tag Shuffle: Applied to last unprotected tags</li><li>Natural Language: Short and Long Caption variants</li></ul><h2 id="changes-from-kirazuri-(anima)-v3.0">Changes from Kirazuri (Anima) v3.0</h2><ul><li>Dataset includes recently curated 2,450 images increasing total size from 42,608 to 45,058 images</li><li>Dataset cutoff now of 29/06/2026</li><li>Introduced Masked Training for images with simple backgrounds</li><li>Updated tags+caption variants structure</li></ul>Installing and runningWorkflow:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/bed46502-0073-4352-91f3-09bb1553bdcf/original=true/bed46502-0073-4352-91f3-09bb1553bdcf.jpeg" />Reference the <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima#installing-and-running">anima base instructions</a>. 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 custom <a target="_blank" rel="ugc" href="https://github.com/asagi4/comfyui-prompt-control">comfyui-prompt-control</a> node for schedule prompting syntax to mix concepts i.e. <code>[word1|word2]</code> 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:<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2495369">anima-kirazuri-v4.safetensors </a>(this model) goes in <code>ComfyUI/models/diffusion_models</code></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders">qwen_3_06b_base.safetensors</a> goes in <code>ComfyUI/models/text_encoders</code></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae">qwen_image_vae.safetensors</a> goes in <code>ComfyUI/models/vae</code> (this is the Qwen-Image VAE, you might already have it)</li></ul>QuantizationsInt8 supported with latest ComfyUI:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/motimalu/kirazuri-anima/blob/main/split_files/diffusion_models/anima-kirazuri-v4-int8-convrot.safetensors">anima-kirazuri-v4-int8-convrot.safetensors</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/motimalu/kirazuri-anima/blob/main/split_files/diffusion_models/anima-kirazuri-v3-int8convrot.safetensors">anima-kirazuri-v3-int8convrot.safetensors</a></li></ul>Distillations4 Step CFG 1 Turbo LoRA distilled with DP-DMD - diversity-preserved few-step distillation:<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2769422/kirazuri-anima-turbo?modelVersionId=3117703">kirazuri-anima-turbo</a></li></ul>Generation SettingsTrained 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:<ul><li>er_sde: the recommended default for 30-50 steps.</li><li>sa_solver_pece: can converge with good detail in 15-20 steps.</li></ul>[Optional] ComfyUI-Autocomplete-Plus prompt input assistanceAn optional file <code>danbooru_tags_kirazuri_4.txt</code> is included to provide prompt tag autocomplete.This file contains metadata that is derived from public sources for prompt assistance only, and is intended to be used with the <a target="_blank" rel="ugc" href="https://github.com/newtextdoc1111/ComfyUI-Autocomplete-Plus">ComfyUI-Autocomplete-Plus</a> extension.Rename the file to <code>danbooru_tags_kirazuri_4.csv</code> and place it in your <code>ComfyUI/custom_nodes/comfyui-autocomplete-plus/data</code> directory.You should see suggestions when writing comma separated tags in ComfyUI:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6fe1dc69-92ec-48ca-85fe-0168a8648875/original=true/6fe1dc69-92ec-48ca-85fe-0168a8648875.jpeg" />PromptingLike the base model, this model is trained on booru-style tags, natural language captions, and combinations of tags and captions.Tag order<code>[quality/meta/safety tags] [character] [series] [artist] [1girl/1boy/1other etc] [general tags]</code>Mostly the same order as the base model, only the <code>[1girl/1boy/other etc]</code> groups position is towards the end in this models dataset.<code>[quality/meta/safety tags] [character] [series] [artist]</code> tag groups are also not shuffled, so their order may have some influence on generations.Quality and Aesthetic tagsHuman score based: <code>masterpiece, best quality, very aesthetic, aesthetic</code>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 tagsabsurdres, official art, etcStylespainterly, chiaroscuro, ligne claire, flat color, no lineart, blending, etctraditional media, oil painting \(medium\), watercolor \(medium\), etcKnown Limitations &amp; 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<ul><li>Thanks to <a target="_blank" rel="ugc" href="https://civarchive.com/user/circlestone_labs">CircleStone Labs</a> for the Anima base models.</li><li>Thanks to tdrussell of CircleStone Labs for the <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a> trainer.</li><li>Thanks to bluvoll for support using their fork of diffusion-pipe.</li><li>Thanks to narugo1992 and the deepghs team for open-sourcing various training sets, image processing tools, and models.</li></ul>LicenseThis model is released under the same license as the base model.See the base model for details of the <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima/blob/main/LICENSE.md">CircleStone Labs Non-Commercial License.</a> Built on NVIDIA Cosmos

Kirazuri (Anima)

Version 1This is an experimental full finetune of the Anima Preview version 1 base.Total training dataset of 15,420 images curated with manual human quality and aesthetic ratings from 2025/07/03 to 2026/03/19, the model should have a fairly strong recency bias and be capable of generating many characters/concepts/styles prominent from that time period.Training Details<ul><li>Samples seen(unbatched steps): ~700,000</li><li>Training time: ~85 hrs</li><li>Learning Rate: 5e-6 (General Training) and 2e-6 (Aesthetic)</li><li>Text Encoder Learning Rate: 1e-6</li><li>Effective Batch size: 24 (General Training) (1 Batch Size) and 32 (Aesthetic) (16x2 Batch Size)</li><li>Precision: Mixed BF16</li><li>Optimizer: AdamW8bit with Kahan Summation</li><li>Weight Decay: 0.01</li><li>Timestep Sampling Strategy: Logit-Normal, Shift 3</li><li>Tag Dropout: 10%</li><li>Uncond Dropout: 10%</li><li>Tag Shuffle: True with keep first 8 tags</li></ul>Additional Features used:<ul><li>Protected Tags</li><li>Mixed Natural Language prompts at ratio:<ul><li>tags 50%, nl 10%, tags+nl 20%, nl+tags 20%</li></ul></li></ul>

Files with SHA256: a83699bc4897a3c1d52c4d638f537932d27e1fa53d08f82ce3aaccda749103f5