Attribution and dependencies

Built with Qwen. Model weights and configurations derive from Qwen/Qwen-Image-2.1, revision 790c92633540aa0cb11d9abf19eb46d861714758. The original Qwen Research License is reproduced without changes in LICENSE, with the required notice in NOTICE. The license applies to the model and its quantized derivatives. This package does not grant additional model rights.

EliovpAI modified the weight storage to the balanced MXFP4 profile and provides the Python loader, single-frame cache handling, CLI, API and validation tools. The adapter code is distributed under LICENSES/Apache-2.0.txt. Dependency and model licenses remain separate.

Diffusers, Transformers, Accelerate and Safetensors retain their Apache-2.0 licenses and upstream notices. PyTorch and Torchvision retain their BSD-style licenses, NumPy its BSD license, and Pillow its HPND license. Preserve the notices included with installed distributions. No dependency source or native extension binary is bundled here.

AMD Quark supplied the original offline weight conversion using OCP E2M1 and E8M0 formats. It is not an inference dependency. The packed tensors were reused without requantization for this portable consumer. The decoder implements that format with eager framework operations and BF16 reconstruction.

The optional quality tools use OpenAI CLIP ViT-B/32, CLIPScore's scoring formula and Richard Zhang and collaborators' LPIPS AlexNet v0.1 model. Their model identities and hashes are in evaluation/metric-manifest.json. Metric weights are downloaded separately and retain their respective upstream notices.