<h1 id="akanezora-anima">Akanezora-Anima</h1> Akanezora is a full Anima DiT fine-tune trained entirely on a single RTX 3060 with 12GB of VRAM using Aozora. It is released both as a usable model and as proof that full Anima DiT fine-tuning can be done on consumer hardware.If you're interested in fine-tuning your own model, the training code is available here: [<a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training">Aozora_SDXL_Trainer</a>]<h3 id="quantized-versions">Quantized Versions</h3><ul><li>FP8: Recommended for RTX 40-series or newer GPUs with native FP8 support. Saves VRAM with near-base quality; older GPUs may use slower FP8 emulation.</li></ul><ul><li>INT8: Recommended for RTX 20-series or newer. Uses ComfyUI <code>int8_tensorwise</code> loading and requires ComfyUI v0.27.0+.</li></ul><ul><li>GGUF Q8_0 / Q5_1: Broad low-VRAM GPU support through [<a target="_blank" rel="ugc" href="https://github.com/city96/ComfyUI-GGUF">ComfyUI-GGUF</a>]. Use its Unet Loader (GGUF) node. Q8_0 is most faithful; Q5_1 is smaller but changes seeds more. </li></ul>For non gguf quants Use <code>torch.compile</code> for additional speed if supported, If a model wont load first make sure comfyui is updated to the latest version v0.27.0+. for gguf you are may need to install gguf into comfy via <code>python -m pip install --upgrade gguf</code> Unlike most mixed-precision quants that trim a few layers for modest savings, these are fully quantized across every supported layer and then QAT-repaired via retraining to hold quality near base — took extra weeks, but worth it i think. Converting Anima to GGUF was a hassle because ComfyUI uses a custom quantized loader and naming format, while GGUF doesn’t officially support Anima and only recognizes its underlying Cosmos architecture. Some shenanigans were needed to make everything compatible with base ComfyUI and the ComfyUI-GGUF addon.Since I think everyone deserves access to lower-VRAM models, here is the complete open-source code for converting Anima models to GGUF or other quantized formats like FP8. [<a target="_blank" rel="ugc" href="https://github.com/Hysocs/Aozora_SDXL_Training/blob/main/tools/convert_anima_to_quants.py">Convert_Anima_to_quants</a>] - Enjoy all 4100 lines of nonsensical experimental code that does indeed work<h3></h3><h2 id="version-0.65b-preview">Version 0.65b Preview</h2>A higher-resolution retrain split from the original Akanezora run using the new Anima Aesthetic-v1.0 release as a shortcut to allow the model to adapt to 1536 res faster, using a stripped-down dataset of 11093 hand curated images at up to 1536px.Improves composition at higher resolutions, Danbooru-tag response, NSFW quality, and unwanted-text suppression. As an early preview, occasional hallucinations remain and some outputs may resemble the base model. this higher res retrain now allows highres fix to work properly so i recommend using it <h3 id="training">Training</h3><ul><li>Base: Anima Aesthetic-v1.0</li><li>Resolution: Up to 1536px, aspect-ratio bucketed</li><li>GPU: RTX 3060 12 GB</li><li>Precision: BF16</li><li>Batch: 1 × 4 gradient accumulation</li><li>Optimizer: Raven/AdamW, momentum offloading</li><li>Learning rate: 3e-6 peak, custom schedule</li><li>Objective: Custom Weighted flow-matching MSE</li><li>Sampling: Stratified Uniform timesteps</li><li>Frozen: <code>llm_adapter.*</code></li><li>Conditioning: Danbooru tags, 0.9–1.1 soft scaling</li><li>Memory: Approximately 11.4 GB using caching, checkpointing and SDPA</li></ul><h2 id="recommended-generation-settings">Recommended Generation Settings</h2>Sampler: ER_SDE Scheduler: BetaSteps: 15-50CFG: 3-5Negative Prompt: worst quality, low quality, lowres, score_1, score_2, score_3, blurry, jpeg artifacts Note: <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">You need to use <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders">qwen_3_06b_base.safetensors</a><span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px"> for text encoder, and <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae">qwen_image_vae.safetensors</a><span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px"> for VAE. <h2 id="model-transparency-notice"> Model Transparency Notice</h2>For transparency, this release includes the training setup and links back to the open-source trainer/code used to create it. This is a full fine-tune checkpoint with no LoRA, LoKR, LyCORIS, or model merge operations applied.This model:<ul><li>Training started with unmodified base weights</li><li>Zero merge operations applied</li><li>No LoRA adapters — ever</li><li>Full end-to-end training, no sublayer freezing besides the required (llm_adapter)</li></ul><h2 id="notes">Notes</h2>Feedback is welcome, especially on prompt following, anatomy, hands, style consistency, repeated patterns, overfitting, and behavior without heavy negative prompts. <h2 id="license">License</h2>This model follows the license of its base model, Anima. Review and comply with the base model terms before using or redistributing.

Akanezora - Multi-Precision + GGUF

<ul><li>Following the 50% checkpoint, I have transitioned to a more optimized configuration. I determined that the default Anima fine-tuning parameters previously utilized by the Anima base, were suboptimal in my opinion. I have since pivoted to a more primitive SDXL-derivative timestep and loss formula. Early results are promising, showing improved seed uniqueness and composition while successfully retaining the core Anima knowledge base. Additionally, I pruned the dataset from 36k to 15k images to mitigate prompt and Style similarity bleeding.</li></ul>

qwen_image_vae.safetensors

akanezoraMultiPrecision_v055B.safetensors

akanezoraMultiPrecision_v055B_txt.safetensors

v0.55:B

v0.65:B

v0.55:B (FP8 & INT8)

v0.5:A

Mixed-precision post-training quantization for Anima DiT.FP8 version: calibrated FP8 E4M3 quantization. Large compatible Linear/matrix weights were quantized to FP8 E4M3 with per-layer scaling after no-grad Anima DiT forward-pass calibration on representative dataset samples.INT8 version: calibrated tensorwise INT8 quantization. Large compatible Linear/matrix weights were quantized to INT8 using per-output-channel weight scales after the same no-grad Anima DiT calibration pass.For both versions, norms, biases, embeddings, positional embedding buffers, final layers, and calibration-sensitive layers remain BF16. Files are saved with ComfyUI mixed-precision quantization metadata, including weight_scale, comfy_quant, and quantization metadata entries for native mixed-precision loading.

Files with SHA256: 5c2e36c18d4c5d85fe930b876c71b9dd7007e29dde774083ce62452b1c15ab09