CivArchive
    Akanezora - Multi-Precision + GGUF - v0.65:B
    NSFW
    Preview 136570488
    Preview 136532733
    Preview 136535102
    Preview 136535348
    Preview 136531573
    Preview 136531574
    Preview 136536331
    Preview 136535583
    Preview 136536848
    Preview 136536073
    Preview 136538032
    Preview 136537594
    Preview 136538763

    Akanezora-Anima


    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: [Aozora_SDXL_Trainer]

    Quantized Versions

    • 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.

    • INT8: Recommended for RTX 20-series or newer. Uses ComfyUI int8_tensorwise loading and requires ComfyUI v0.27.0+.

    • GGUF Q8_0 / Q5_1: Broad low-VRAM GPU support through [ComfyUI-GGUF]. Use its Unet Loader (GGUF) node. Q8_0 is most faithful; Q5_1 is smaller but changes seeds more.

    For non gguf quants Use torch.compile 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 python -m pip install --upgrade gguf


    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. [Convert_Anima_to_quants] - Enjoy all 4100 lines of nonsensical experimental code that does indeed work

    Version 0.65b Preview

    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

    Training

    • Base: Anima Aesthetic-v1.0

    • Resolution: Up to 1536px, aspect-ratio bucketed

    • GPU: RTX 3060 12 GB

    • Precision: BF16

    • Batch: 1 × 4 gradient accumulation

    • Optimizer: Raven/AdamW, momentum offloading

    • Learning rate: 3e-6 peak, custom schedule

    • Objective: Custom Weighted flow-matching MSE

    • Sampling: Stratified Uniform timesteps

    • Frozen: llm_adapter.*

    • Conditioning: Danbooru tags, 0.9–1.1 soft scaling

    • Memory: Approximately 11.4 GB using caching, checkpointing and SDPA

    Sampler: ER_SDE
    Scheduler: Beta

    Steps: 15-50

    CFG: 3-5

    Negative Prompt: worst quality, low quality, lowres, score_1, score_2, score_3, blurry, jpeg artifacts
    Note: You need to use qwen_3_06b_base.safetensors for text encoder, and qwen_image_vae.safetensors for VAE.


    Model Transparency Notice

    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:

    • Training started with unmodified base weights

    • Zero merge operations applied

    • No LoRA adapters — ever

    • Full end-to-end training, no sublayer freezing besides the required (llm_adapter)

    Notes

    Feedback is welcome, especially on prompt following, anatomy, hands, style consistency, repeated patterns, overfitting, and behavior without heavy negative prompts.

    License

    This model follows the license of its base model, Anima. Review and comply with the base model terms before using or redistributing.

    Description

    FAQ

    Checkpoint
    Anima
    by Hysocs
    Download (Beta) View on CivitAI
    anime
    base model
    artists
    nsfw
    base
    full tune

    Details

    Downloads
    139
    Platform
    CivitAI
    Platform Status
    Available
    Created
    7/12/2026
    Updated
    7/15/2026
    Deleted
    -

    Files

    qwen_image_vae.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (172 mirrors)
    CivitAI (104 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (1 mirrors)
    ModelScope CN (5 mirrors)

    akanezoraMultiPrecision_v065B_txt.gguf

    Size:
    609.82 MB
    SHA256:
    9465e63a22add5354d9bb4b99e90117043c7124007664907259bd16d043bb031

    Mirrors

    HuggingFace (4 mirrors)
    CivitAI (1 mirrors)

    akanezoraMultiPrecision_v065B.safetensors

    Size:
    3.89 GB
    SHA256:
    104cca1eff5e0c4fafdf5c98d3c866bd057f01d4a7d724fdf1cca9c7efc11b36

    Mirrors

    CivitAI (1 mirrors)

    akanezoraMultiPrecision_v065B.safetensors

    Size:
    2.00 GB
    SHA256:
    a1cf3b2e591f51737581f012d3b31f8acce7737e6b498dc10e4d817687da31fa

    Mirrors

    CivitAI (1 mirrors)

    akanezoraMultiPrecision_v065B_txt.safetensors

    Size:
    1.11 GB
    SHA256:
    cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba

    Mirrors

    HuggingFace (81 mirrors)
    CivitAI (96 mirrors)
    ModelScope CN (2 mirrors)

    akanezoraMultiPrecision_v065B.gguf

    Size:
    1.46 GB
    SHA256:
    9218812174ec4cc838da858b2720aaab3617cc101aed205754e69257e3f25e96

    Mirrors

    CivitAI (1 mirrors)

    akanezoraMultiPrecision_v065B.safetensors

    Size:
    2.00 GB
    SHA256:
    86312704e94b57a320ba2455b7f5c90a333fc067fac91f78806e95821a3664d8

    Mirrors

    CivitAI (1 mirrors)

    akanezoraMultiPrecision_v065B.gguf

    Size:
    2.07 GB
    SHA256:
    77818b2e90a8349e6f3d7b197e4f2d9e376ff9d58be3c25b6c30378ad03f507b

    Mirrors

    CivitAI (1 mirrors)