To run the INT6 W6A8 version, the ComfyUI version must be <a target="_blank" rel="ugc" href="https://docs.comfy.org/changelog#v0-38-0">v0.38.0</a> or higher.Source: <a target="_blank" rel="ugc" href="https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/tree/main/v23">Phr00t/Qwen-Image-Edit-Rapid-AIO</a><hr /><ol><li>Source Precision: The author did not provide an FP16 version, so the INT8 and INT6 versions were converted from the FP8 version.</li><li>Model Extraction: I only extracted the Diffusion Model part from the AIO version, and removed the Text Encoders and VAE parts.</li><li>Personal Recommendation: I personally use the INT8 Convrot version frequently, and I often use it for editing tasks.</li><li>About INT6 W6A8: W6A8 is a bit slower than INT8 Convrot (e.g., 43s vs. 39s), but the quality loss is acceptable and it has a smaller file size.</li><li>About INT4/W4A8: Compared to the INT8 and W6A8 versions, the details of the INT4 and W4A8 versions are too poor, making the INT4 version almost unusable. I do not recommend trying any INT4/INT4 Mixed/INT4 W4A8 version of Qwen Image.</li><li>Compared with Qwen Image 2.1: Although Qwen 1's details or textures are weaker than Qwen Image 2.1 and Qwen 1 is much slower, Qwen 1's prompt adherence is better.</li><li>Issues: For various issues, embedded LoRAs, optimal parameter settings, etc., please check the original author's page. The greasy skin has been mentioned by other users on the original repo's Community page, and it is highly likely an issue with the embedded LoRAs.</li></ol>Disclaimer: This model belongs to <a target="_blank" rel="ugc" href="https://huggingface.co/Phr00t">Phr00t</a>, I have not made any modifications to the original model myself, and I am only sharing the quantized version.

Files with SHA256: 6b7f8ac60335e041c151698e1d7777278e561f3ecb54c9b2a981694444f3c604