Qwen-Image-2.1-Uncensored-Omni-Quantized: Running Qwen-Image 2.1 on 4GB VRAM

Quantized transformer weights created by merging ChatR Labs' rank-128 Omni adapter into Qwen-Image-2.1. It performs text-based image generation and editing of existing images. The text encoder and VAE must be obtained from a separate repository, as this repository contains only the transformer.

It offers various quantization formats, including GGUF, FP8, INT8, and NVFP4. It supports both ComfyUI and Diffusers, with the NVFP4 build specifically operating at a peak VRAM level of around 4GB on an RTX 5090. The significantly reduced size compared to BF16 enables inference even in low-spec GPU environments.

Prompt adherence for adult content increased by 4.6 points compared to the base model. Performance on general SFW prompts shows no significant difference from the base model, and anatomical accuracy has not improved. A setting of 35 steps and CFG 5.0 is recommended, and enabling the KV cache is mandatory.

It is suitable for developers who want to bypass the filtering of the existing Qwen-Image-2.1 for more unrestricted image generation. It provides ComfyUI workflows, making it easy to build node-based pipelines. When using Diffusers, dependencies on the bitsandbytes or torchao libraries may occur.