<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>Use <code>torch.compile node</code> for faster generation if supported. Speedup depends on GPU architecture, PyTorch/ComfyUI support, and whether the quantized op is native or emulated.FP8 E4M3: Best on GPUs with native FP8 Tensor Cores (RTX 40-series/Ada or newer, Hopper). Without native support, FP8 emulation may cut VRAM but rarely speeds things up.Saved as mixed-precision FP8 E4M3 with activation-aware layer selection, high-impact Linear layers quantized, sensitive layers kept BF16. Seeds won't match the base model exactly; expect minor quality loss and more seed variance.INT8: Uses ComfyUI's native <code>int8_tensorwise</code> mixed-precision loading (Requires comfyui version V0.27.0 and up), more numerically conservative than FP8, still saves VRAM with possible speed gains on supported hardware.Saved with calibrated tensorwise INT8 (per-output-channel weight scales), same activation-aware selection, with norms/biases/embeddings/positional buffers/final layers kept BF16. Closer to BF16 output than FP8, but seed matching still isn't guaranteed.<h3 id="branch-a-vs-b"> Branch A vs B</h3>Branch A: follows the recommended Anima tuning setup with bell-weighted loss and non-uniform timestep sampling. Branch B: uses my experimental setup with uniform timestep sampling and uniform loss weighting, giving the model more even exposure across the full noise range. In my testing, this improves visual style and composition, but is slower and harder to tune. <h2 id="version-0.55b-preview">Version 0.55b Preview</h2>This is a 55% training checkpoint continued from the 0.5a checkpoint, trained on 15k images using experimental Branch B settings. The dataset consists of 50% Danbooru-tagged images with generated natural language and 50% hand-tagged / natural language-captioned images. During training, text conditioning was split into 90% tags and 10% natural language prompts.While this release is functional, it should be considered a work in progress rather than a finished product. It is being shared early to demonstrate the viability of the training method and to showcase the Aozora trainer’s ability to fine-tune Anima DiT models on low-VRAM hardware. Pros:<ul><li>Reduces unwanted text generation by around 70%, reducing the need for heavy negative prompts.</li><li>More responsive to Danbooru-style tags.</li><li>Slightly more dynamic seed variation due to soft conditioning.</li><li>More SDXL-inclined generation style.</li><li>Better overall composition and prompt feel across varied prompts.</li><li>Improved NSFW output quality.</li></ul>Cons:<ul><li>Some seeds may still closely resemble the base model.</li><li>Lighting effects often need to be prompted directly.</li><li>Still early in training and may hallucinate content.</li></ul><h2 id="0.55b-training-settings">0.55b Training Settings</h2>Base Model: Akanezora V0.5a Training Hours: Unknown (Power went out so it took 2x longer, estimated around 50 hours) GPU Used: NVIDIA GeForce RTX 3060 (12 GB) | Driver version: 32.0.15.9636VRAM Usage: ~11.4GBMixed Precision: bfloat16Batch Size: 1Gradient Accumulation: 4Learning Rate: 6e-6 Timestep: Uniform Loss: UniformOptimizer: Raven[AdamW float32 variant with offloading] | (betas:0.9, 0.999 | eps:1e-08 | Weight Decay: 0.01| Debias: 1.0)Max Train Steps: 201010 (Completed:115282)Current Checkpoint: ~55% through planned trainingTrainable Parameters: - (P: 1,956,405,248 | P Frozen: 6.44% [llm_adapter.*])Soft text cond: ( 0.75 - 1.25)Dataset Size: 15164 Training Resolution: 1152x1152 (Aspect Ratio Bucketed: 864x1536 to 1536x864)VRAM saving techniques: (Momentum offloading, bfp16 mixed precision, pre-caching VAE and text encoders, Gradient Checkpointing)<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

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.

akanezora_v055BFP8INT8.safetensors

v0.55:B (FP8 & INT8)

v0.55:B

v0.5:A

<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>

Files with SHA256: 223b4daee86fff01c8e0b5014bd509fde571100ef75034d3b101e267aa233b70