ANIMA 3.8B INT8 ConvRot for Forge Neo

Unofficial community quantization. This repository contains weight-only conversions of lylogummy/Anima-3.8B. These files were not trained by Aikimi and are not official or endorsed releases from LyloGummy, CircleStone Labs, NVIDIA, or Forge Neo.

This release keeps the ANIMA 3.8B architecture and converts only the 520 main attention and MLP matrices in the 52-block diffusion transformer to Forge-native int8_tensorwise + ConvRot. AdaLN, embeddings, boundary layers, normalization layers, and the v1.1 Semantic Connector v2 remain BF16.

Files

File	Role	Size	SHA-256
`Anima-3.8B-v1.1-int8-convrot.safetensors`	Recommended v1.1 bundle with Semantic Connector v2	5,543,364,574 bytes	`640ef9bc91900cef01e9f7fb49390b89e27d197dfc7adeb2abe776dcdaf54914`
`Anima-3.8B-int8-convrot.safetensors`	Legacy v1 52-block checkpoint	4,238,326,342 bytes	`ab869391139d4aa17bab2559ac4640a6813c95f4278e58cc8a2f1564f4fbb18c`
`Anima-3.8B-expanded_adapter.safetensors`	Unmodified legacy v1 Qwen3.5 adapter	88,131,712 bytes	`f9851ac4668ce069f7be7cf99755335c98879b463f3d486aaa731083978f0d71`

The v1.1 conversion is 37.1% smaller than its 8,809,227,318-byte BF16 source. The legacy v1 conversion is 43.5% smaller than its 7,504,189,974-byte BF16 source.

This is not a full-INT8 bundle. In v1.1, all 190 Semantic Connector v2 tensors remain BF16 because Forge loads the connector as a sampling-time module.

Required external files

The text encoders and VAE are separate inference components and are not included here. Download them from the upstream repositories:

Verify their expected hashes in conversion_manifest.json.

Forge Neo installation

Use a Forge Neo build with Forge-native ConvRot support and the current ANIMA 3.8B Forge extension.

models/
├── Stable-diffusion/
│   └── Anima-3.8B-v1.1-int8-convrot.safetensors
├── text_encoder/
│   ├── qwen35_4b.safetensors
│   └── qwen_3_06b_base.safetensors
└── VAE/
    └── qwen_image_vae.safetensors

For v1.1, select the checkpoint, native Qwen encoder, and Qwen Image VAE. The extension detects the bundled Semantic Connector v2 automatically. The trained connector strength is fixed at 1.0; do not select the legacy adapter.

Recommended starting point:

Resolution: 832x1216 or another approximately 1MP size
Sampler / Scheduler: Res Multistep + Beta
Steps: 28-50
CFG: 4-7
Shift: 3.0
Diffusion in Low Bits: Automatic

Four-step smoke tests produced an almost black image in the verified environment. Use the recommended 28 or more steps before treating an output as a model failure.

Legacy v1

Place the legacy checkpoint in models/Stable-diffusion and Anima-3.8B-expanded_adapter.safetensors in models/text_encoder. Enable the ANIMA 3.8B panel, select the adapter, and use strength 1.0.

Conversion provenance

The conversion used the open implementation below:

Converter: tools/convert_anima38_int8_convrot.py
Converter commit: 631fa18e508b99aa65f8c90d3c2bdcf1d33ba19c
ConvRot group size: 256
Quantized weights: 520
Runtime used for verification: Forge Neo 2.28, PyTorch 2.11.0+cu130, Comfy-Kitchen 0.2.31, RTX 3090 24GB

The complete source revisions, source hashes, output hashes, tensor counts, preserved scopes, and dependency hashes are recorded in conversion_manifest.json. Published checksums are also available in CHECKSUMS.sha256.

No-LoRA comparison

The following four pairs use identical positive and negative prompts, seeds, resolution, sampler, scheduler, CFG, and Shift. No prompt LoRA or standard LoRA was used. Only the checkpoint-specific Qwen3.5 connector path differs.

In this RTX 3090 run, warm API wall time averaged 24.05 seconds per changed prompt for legacy v1 and 31.99 seconds for v1.1. These values include prompt conditioning and are not universal performance claims. The outputs are not pixel-identical and should be judged visually for each use case.

Scope and limitations

Verified in Aikimi Neo / Forge Neo only. This Forge-native .comfy_quant ConvRot checkpoint has not been validated in standard ComfyUI, Diffusers, or AUTOMATIC1111.
Qwen3.5 4B, native Qwen3 0.6B, and the Qwen Image VAE are still required.
Quantization changes numerical output and may change composition, style, hands, text, identity, and prompt adherence.
The legacy v1 and v1.1 connector paths are different architectures; the size and speed comparison is not a pure quantizer benchmark.
The repository is download-only and does not provide a hosted inference API.

License and notices

The model weights and derivatives remain subject to the CircleStone Labs Non-Commercial License and the applicable NVIDIA Open Model License Agreement. Free download from this repository does not grant additional commercial model-use rights.

Read NOTICE.md before redistribution. It includes the required attribution, modification statement, non-endorsement statement, and upstream sources.

日本語案内

これはANIMA 3.8Bの学習済み重みを使った、Forge Neo向けの非公式INT8 ConvRot変換です。追加学習やfine-tuningは行っていません。推奨版はAnima-3.8B-v1.1-int8-convrot.safetensorsで、DiT主要520行列だけをINT8化し、Semantic Connector v2はBF16のまま保持しています。

Qwen3.5 4B、標準Qwen3 0.6B、Qwen Image VAEは含まれていません。上記の公式リンクから別途取得してください。旧v1版を使う場合だけ、同梱するAnima-3.8B-expanded_adapter.safetensorsが必要です。

利用・再配布前に、LICENSE-CIRCLESTONE.md、LICENSE-NVIDIA-OPEN-MODEL.md、NOTICE.mdを確認してください。このrepoを無料公開しても、元ライセンスの非商用条件が解除されるわけではありません。

Acknowledgements

Downloads last month: 267

Model tree for Aikimi/Anima-3.8B-INT8-ConvRot-neo

Base model

lylogummy/Anima-3.8B

Quantized

(1)

this model