Hugging Face's logo

Aikimi
/
anima-mixbit-6.44

Text-to-Image
Safetensors
Japanese
English
comfyui
anima
diffusion
quantized
mixed-precision
low-vram
Model card Files
xet
 Community

Anima MixBit 6.44

比較 Default Anima BF16 INT8 ConvRot MixBit 6.44 MixBit 省VRAM
必要ファイル合計 5.628 GB 3.768 GB 2.265 GB 2.265 GB・同一checkpoint
実行方式 通常CFG／VAE 通常CFG／VAE 通常CFG／VAE 逐次CFG／tiled VAE 192/64
GPU条件 RTX 3090で実測 RTX 3090で実測 RTX 3090で実測・対応 RTX 40／50系相当
CC 8.9または9.0以上
RTX 5080 Laptopで実測
RTX 3090は非対応
VRAM・モデルロード前 → 最大値 4,739 → 10,575 MiB 4,740 → 8,838 MiB 4,730 → 8,485 MiB 1,498 → 1,988 MiB
必要追加VRAM 5,836 MiB 4,098 MiB
BF16比 −29.8%		 3,755 MiB
BF16比 −35.7%		 490 MiB（3回最大）
ウォーム生成時間・中央値 5.401秒 7.165秒
BF16比 +32.7%		 7.452秒
BF16比 +38.0%		 13.187秒
BF16画像との一致度 基準 SSIM 0.5487
LPIPS 0.2461
CLIP 0.9567		 SSIM 0.3981
LPIPS 0.3372
CLIP 0.9421		 20件
SSIM 0.7040
LPIPS 0.3474
CLIP 0.9120
プロンプトCLIP類似度 0.3600 0.3654 0.3643 未測定
結論 品質基準・最速・最大 BF16に最も近い 通常モードで最小容量・最小追加VRAM
INT8比 +4.0%		 追加VRAM最大490 MiB
対応GPU限定・最も遅い

通常3方式はRTX 3090 24 GB、ComfyUI 0.30.0、PyTorch 2.11.0+cu130、512×512、30 steps、CFG 4.5、er_sde／simpleで測定しました。必要追加VRAMは「モデルロード前のnvidia-smi値 → 全実行中の最大値」の差です。

比較画像

Default Anima BF16、INT8 ConvRot、MixBit 6.44の同一条件アニメ生成比較

比較prompt
  1. Neon city:
masterpiece, best quality, year 2025, newest, highres, absurdres, anime screenshot, official art, safe, 1girl, solo, original character, adult woman, short pink hair, teal eyes, futuristic magical courier, black and cyan jacket, pleated skirt over leggings, glowing star wand, leaping between neon city rooftops at night, dramatic foreshortening, colorful billboards, crescent moon, speed lines, energetic expression, vibrant colors, clean lineart, detailed cel shading
  1. Mecha hangar:
masterpiece, best quality, year 2025, newest, highres, absurdres, anime screenshot, official art, safe, 1girl, solo, original character, adult woman, tan skin, black ponytail, amber eyes, orange and white pilot suit, inside a vast mecha hangar, towering white and navy humanoid robot behind her, helmet under one arm, maintenance crew silhouettes, low angle wide shot, morning light beams, intricate mechanical detail, confident grin, clean lineart, cel shading
  1. Magic library:
masterpiece, best quality, year 2025, newest, highres, absurdres, anime screenshot, official art, safe, 1girl, solo, original character, adult woman, long emerald hair, pointed ears, violet and black witch dress with gold embroidery, floating ancient library, open spellbook, circular magic sigils, papers and glowing runes swirling in the air, stained glass windows, candlelight, elegant pose, fantasy anime illustration, detailed background, clean lineart, cel shading

Negative:

worst quality, low quality, artist name, blurry, jpeg artifacts, chromatic aberration, photorealistic, realistic, 3d, monochrome, grayscale, simple background, malformed face, asymmetrical eyes, bad anatomy, bad hands, malformed hands, fused fingers, extra fingers, missing fingers, detached fingers, duplicated limbs, extra limbs, deformed, nsfw, explicit, watermark, signature, logo, text

導入方法

共通

  1. Hugging FaceのFiles and versionsから次をダウンロードします。

  2. custom nodeを次の構成でコピーします。

    ComfyUI/custom_nodes/AnimaMixBit/
ComfyUI/custom_nodes/AnimaVRAMFusion/

  3. ComfyUIのPython環境へ依存関係を入れます。

    通常版ComfyUI:

    python -m pip install -r .\custom_nodes\AnimaMixBit\requirements.txt

    Windows portable版:

    .\python_embeded\python.exe -s -m pip install -r .\ComfyUI\custom_nodes\AnimaMixBit\requirements.txt

  4. modelをComfyUI/models/diffusion_models/anima-mixbit-6.44-proportional-vram.safetensorsへ置きます。外部text encoder、VAE、LoRAは不要です。

1. 通常モード

UI workflowのanima_mixbit_6_44_standard.jsonをComfyUIへドラッグして読み込みます。

  • loader: AnimaMixedLowBitLoader
  • VAE: VAEDecode
  • RTX 3090で動作確認済み

通常版ComfyUI:

python main.py

Windows portable版:

.\python_embeded\python.exe -s .\ComfyUI\main.py --windows-standalone-build

2. 省VRAMモード

UI workflowのanima_mixbit_6_44_low_vram.jsonをComfyUIへドラッグして読み込みます。

  • loader: AnimaVRAMProportionalLoader
  • VAE: VAEDecodeTiled、tile 192／overlap 64
  • Compute Capability 8.9、または9.0以上
  • RTX 3090は非対応

通常版ComfyUI:

python main.py --reserve-vram 11.0 --vram-headroom 5.0 --disable-smart-memory --use-pytorch-cross-attention

Windows portable版:

.\python_embeded\python.exe -s .\ComfyUI\main.py --windows-standalone-build --reserve-vram 11.0 --vram-headroom 5.0 --disable-smart-memory --use-pytorch-cross-attention

ライセンス

このmodelはAnima Aestheticの非公式な量子化派生物です。CircleStone Labsによる公式配布・承認・検証品ではありません。model weightsは非商用ライセンスです。

MIT Licenseはmodel weightsには適用されません。利用・再配布前にNOTICEと第三者component一覧を確認してください。

English

Anima MixBit 6.44

Comparison Default Anima BF16 INT8 ConvRot MixBit 6.44 MixBit low-VRAM
Required files 5.628 GB 3.768 GB 2.265 GB 2.265 GB, same checkpoint
Runtime regular CFG/VAE regular CFG/VAE regular CFG/VAE sequential CFG/tiled VAE 192/64
GPU requirement Measured on RTX 3090 Measured on RTX 3090 Measured and supported on RTX 3090 RTX 40/50-series equivalent
CC 8.9 or 9.0+
Measured on RTX 5080 Laptop
RTX 3090 unsupported
VRAM, before model load → maximum peak 4,739 → 10,575 MiB 4,740 → 8,838 MiB 4,730 → 8,485 MiB 1,498 → 1,988 MiB
Required additional VRAM 5,836 MiB 4,098 MiB
−29.8% vs BF16		 3,755 MiB
−35.7% vs BF16		 490 MiB (maximum of 3 runs)
Median warm generation 5.401 s 7.165 s
+32.7% vs BF16		 7.452 s
+38.0% vs BF16		 13.187 s
Image agreement vs BF16 Reference SSIM 0.5487
LPIPS 0.2461
CLIP 0.9567		 SSIM 0.3981
LPIPS 0.3372
CLIP 0.9421		 20 cases
SSIM 0.7040
LPIPS 0.3474
CLIP 0.9120
Prompt CLIP similarity 0.3600 0.3654 0.3643 Not measured
Result Quality reference, fastest, largest Closest to BF16 Smallest files and additional VRAM in standard mode
+4.0% vs INT8		 490 MiB maximum additional VRAM
Restricted GPU support and slowest

The three standard modes were measured on an RTX 3090 24 GB with ComfyUI 0.30.0 and PyTorch 2.11.0+cu130 at 512×512, 30 steps, CFG 4.5, and er_sde/simple. Required additional VRAM is the difference between the nvidia-smi value before model loading and the maximum peak across all runs.

Comparison image

Matched anime generations from Default Anima BF16, INT8 ConvRot, and MixBit 6.44

Comparison prompt
  1. Neon city:
masterpiece, best quality, year 2025, newest, highres, absurdres, anime screenshot, official art, safe, 1girl, solo, original character, adult woman, short pink hair, teal eyes, futuristic magical courier, black and cyan jacket, pleated skirt over leggings, glowing star wand, leaping between neon city rooftops at night, dramatic foreshortening, colorful billboards, crescent moon, speed lines, energetic expression, vibrant colors, clean lineart, detailed cel shading
  1. Mecha hangar:
masterpiece, best quality, year 2025, newest, highres, absurdres, anime screenshot, official art, safe, 1girl, solo, original character, adult woman, tan skin, black ponytail, amber eyes, orange and white pilot suit, inside a vast mecha hangar, towering white and navy humanoid robot behind her, helmet under one arm, maintenance crew silhouettes, low angle wide shot, morning light beams, intricate mechanical detail, confident grin, clean lineart, cel shading
  1. Magic library:
masterpiece, best quality, year 2025, newest, highres, absurdres, anime screenshot, official art, safe, 1girl, solo, original character, adult woman, long emerald hair, pointed ears, violet and black witch dress with gold embroidery, floating ancient library, open spellbook, circular magic sigils, papers and glowing runes swirling in the air, stained glass windows, candlelight, elegant pose, fantasy anime illustration, detailed background, clean lineart, cel shading

Negative:

worst quality, low quality, artist name, blurry, jpeg artifacts, chromatic aberration, photorealistic, realistic, 3d, monochrome, grayscale, simple background, malformed face, asymmetrical eyes, bad anatomy, bad hands, malformed hands, fused fingers, extra fingers, missing fingers, detached fingers, duplicated limbs, extra limbs, deformed, nsfw, explicit, watermark, signature, logo, text

Installation

Common setup

  1. Download the following from Hugging Face Files and versions.

  2. Copy the custom nodes into this layout:

    ComfyUI/custom_nodes/AnimaMixBit/
ComfyUI/custom_nodes/AnimaVRAMFusion/

  3. Install dependencies in the ComfyUI Python environment.

    Standard ComfyUI:

    python -m pip install -r .\custom_nodes\AnimaMixBit\requirements.txt

    Windows portable:

    .\python_embeded\python.exe -s -m pip install -r .\ComfyUI\custom_nodes\AnimaMixBit\requirements.txt

  4. Place the model at ComfyUI/models/diffusion_models/anima-mixbit-6.44-proportional-vram.safetensors. No external text encoder, VAE, or LoRA is required.

1. Standard mode

Drag anima_mixbit_6_44_standard.json into ComfyUI.

  • Loader: AnimaMixedLowBitLoader
  • VAE: VAEDecode
  • Validated on an RTX 3090

Standard ComfyUI:

python main.py

Windows portable:

.\python_embeded\python.exe -s .\ComfyUI\main.py --windows-standalone-build

2. Low-VRAM mode

Drag anima_mixbit_6_44_low_vram.json into ComfyUI.

  • Loader: AnimaVRAMProportionalLoader
  • VAE: VAEDecodeTiled, tile 192/overlap 64
  • Compute capability 8.9 or 9.0+
  • RTX 3090 unsupported

Standard ComfyUI:

python main.py --reserve-vram 11.0 --vram-headroom 5.0 --disable-smart-memory --use-pytorch-cross-attention

Windows portable:

.\python_embeded\python.exe -s .\ComfyUI\main.py --windows-standalone-build --reserve-vram 11.0 --vram-headroom 5.0 --disable-smart-memory --use-pytorch-cross-attention

License

This model is an unofficial quantized derivative of Anima Aesthetic. It is not an official, endorsed, or validated CircleStone Labs release. The model weights are under a non-commercial license.

The MIT licenses do not apply to the model weights. Review the NOTICE and third-party component list before use or redistribution.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
Text-to-Image
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Aikimi/anima-mixbit-6.44

Base model

nvidia/Cosmos-Predict2-2B-Text2Image
Finetuned
circlestone-labs/Anima
Quantized
(32)
this model