Qwen Image (Supportive Code)

Introduction

This is a supportive repo for the Qwen Image model. This repo is created and managed by RenderFlowAI

The goal of this repo is to help people run image generation with Qwen Image model either via their own GPUs or via GPUs instances on cloud. We packages a Gradio app for the original Qwen Image model as well as some of its quantized variants, so that you can try this model on different hardware. You can either use the source code in this repo to build your own container image, or use our built ones to quickly start local or remote Qwen Image inference server with Gradio UI support.

If you see problems using the code in this repo, please don't hesitate to create issues to us, or search for guidance from the official Qwen Image repos:

• Official Qwen Image model weights: https://huggingface.co/Qwen/Qwen-Image
• Official Qwen Image Github repo: https://github.com/QwenLM/Qwen-Image

If you want to deploy this model on Runpod, feel free to use our referral link: https://runpod.io?ref=jc1efdrq

Sub-Directories

/full

This directory contain an implementation of a gradio app using the full original Qwen Image model. It is not quantized and therefore has the full capability of the model itself.

A pre-built docker image for this gradio app can be found at: renderflowai/qwen-image-20b:gradio-app-full. You can pull this image and start gradio app container with it directly.

If you want to build an image by yourself, you can run command

IMAGE_URL=$YOUR_OWN_IMAGE_URL
cd full
docker build -t $IMAGE_URL .

Upon starting of the container with this image, the Qwen Image model will be auto loaded before the gradio app is ready to serve request. The default port number of the gradio app is 7860. You can override this port number via environment variable GRADIO_PORT

This gradio requires at least 40GB VRAM, so the easiest way to get it work is to run it with Nvidia A100 GPU

/dfloat11_quantized

This directory contains an implementation of a gradio app using the quantized Qwen Image model. This is a DFloat11 losslessly compressed version of the original Qwen/Qwen-Image model. It is obtained from this DFloat11/Qwen-Image-DF11 repo on HuggingFace, so all credits go to its authors.

A pre-built docker image for this gradio app can be found at: renderflowai/qwen-image-20b:gradio-app-dfloat11-quantized. You can pull this image and start gradio app container with it directly.

If you want to build an image by yourself, you can run command

IMAGE_URL=$YOUR_OWN_IMAGE_URL
cd dfloat11_quantized
docker build -t $IMAGE_URL .

Similar to the full version gradio app, this gradio app is listening to port 7860 by default, and can be changed to another port by overriding the GRADIO_PORT environment variable.

3 other environment variables specifically related to this DFloat11 quantized model are:

• CPU_OFFLOAD: Default is true.

• CPU_OFFLOAD_BLOCKS: Default is 0, and can be tuned to different values If CPU_OFFLOAD is true

• PIN_MEMORY: Default is true. If false, this can further reduce the amount of GPU memory needed for inference, at the expense of run time.

Detailed explanation and usage of these environment variables can be found in the official repo: https://huggingface.co/DFloat11/Qwen-Image-DF11

More

We will push more Qwen Image related container images in our official Docker Hub repo: https://hub.docker.com/r/renderflowai/qwen-image-20b. Feel free to check it out

Disclaimer

RenderFlow AI is not affiliated with Alibaba, Tongyi Lab, or the teams that developed the Qwen-Image model. All rights to the Qwen-Image model and its associated trademarks belong to their respective owners.