Neo-Image-2.1-W4A8

Neo Image 2.1 W4A8

Built with Qwen. This is an unofficial, weight-only W4A8 conversion of Qwen Image 2.1 for Aikimi Studio Neo / Forge Neo. It is not trained, approved, or released by the Qwen team. The original weights and this conversion are subject to the Qwen Research License; use and redistribution are limited to non-commercial research or evaluation under that license. Read NOTICE.md for the required attribution and change notice.

Contents

This repository contains a 4.20 GB diffusion transformer and a 7.56 GB Qwen3-VL text encoder, both converted from upstream revision b3179ad355be050328e483a9dfdd9e60cd62adfa. The transformer has 224 W4A8 linear layers; the language component has 252. The remaining layers retain their source precision. The quantizer uses Comfy Kitchen's AsymW4A8Int8Layout: 4-bit packed weights, 8-bit activation computation, group size 16, ConvRot group size 256, and FP8 group scales. It requires the matching CUDA kernels.

transformer/ and text_encoder/ contain the saved packed tensors and their w4a8.json maps. release_manifest.json records source revision, conversion recipe, exact runtime versions, file sizes, and SHA-256 hashes. The VAE, processor, scheduler, and unconverted source files are not included.

Use in Forge Neo

The current Neo loader needs the pinned official BF16 model installed first. This release avoids the local W4A8 conversion step; it does not currently remove the need to download and retain the official BF16 files. Install Neo's Qwen environment with aikimi-qwen-image21-setup.bat, then run from the Neo root in PowerShell:

& .\venv\Scripts\hf.exe download Aikimi/Neo-Image-2.1-W4A8 --local-dir .\tmp\neo-image21-w4a8
& .\models\Qwen-Image-2.1\worker-env\Scripts\python.exe -X utf8 .\tmp\neo-image21-w4a8\install.py install --precision w4a8 --neo-root .

Select W4A8 in the Qwen Image 2.1 tab. Neo checks the source revision and runtime versions before installing this release and verifies every downloaded file against release_manifest.json. A subsequent load should report quantized_cache status hit for both components. The installer does not replace an existing saved model. These files are not a standard Diffusers from_pretrained() repository; the W4A8 loader lives in Neo.

Verification and limits

The saved W4A8 components were reloaded in another process and used for an actual 1024×1024 image edit on an RTX 3090. A separate new-image comparison with INT8 is documented with images and measurements; it is one trial, not a general speed or image-quality guarantee. Persistence and edit validation and the W4A8 implementation details are available in Neo. Quantization can change composition and fine details. CUDA with BF16 support and the versions recorded in the manifest are required by this release's importer.

日本語: Neo向けの非公式変換です。公式のQwen Image 2.1一式を先に導入してから、上記の2コマンドで変換済み重みを取り込みます。再量子化を省けますが、現行のNeoでは元のBF16ファイルも必要です。非商用の研究・評価用途に限る元ライセンスと変更・帰属表示を確認してください。