Simple IMG2IMG workflow using QWEN IMAGE (qwen_image_fp8_e4m3fn).
Results are pretty solid.
Use a denoise value between 0.5–0.9.
0.5 = more similarity to the original image
Higher values = more creative deviation
If you're going above 0.5 denoise, you’ll likely need more refined prompts. ChatGPT or another AI with image analysis capabilities can help with that.
Tested on:
GPU: 4060Ti (16GB VRAM)
RAM: 32GB
Image: 768x1344, 30 steps
Generation time: ~4 minutes
Quality is very good
You can experiment with ModelSamplingAura
use shift between 0.5 - 3.0
P.S. Only use this if you’re sure Qwen Image is installed and running properly on your machine. I’m not tech-savvy, so I won’t be able to help troubleshoot.
Description
v1
FAQ
Comments (8)
Smth must be wrong, the workflow does not load, and qwen is installed and working, maybe post one of the images without the stitching (just a normal result image) so I can drag that in comfyUI, OR rightclick in an empty space (no node), below you will see "workflow image" choose export and regular PNG, NOT the one with the (no embedded workflow) that defeats the purpose, and post that in here
The workflow file loads fine for everyone else — if you drag the JSON into ComfyUI, it works. There’s also an image in the gallery with the workflow embedded; just drag that in. If it still doesn’t load, the problem’s on your end (broken install, outdated build, or missing nodes).
gunjinokanrei not outdated more like specialized to work with a 5080 rtx, hot tip before you buy a 50XX wait a bit. You might be right though of it being on my end but I updated everything comfyUI related yesterday, tested the regular qwen image and I have no problems loading other workflows, I will try the image.
gunjinokanrei the image works
serget2 glad to hear it!
Thank you for your workflow, it was very helpful in getting me started.
Thank you especially for DENOISE... it is reversed !!!
Normally, it is towards 1 = greater similarity to the original image...
but you are right, it's the opposite. 👍👍👍
my DENOISE first test ... 0 to 0.7 = original image
I saw the test specs and it reminded me of how anemic the gaming cards are for A.I. I mean, it it works, it works, but dang. those things are really expensive and great for games but they don't do so well on A.I. Not even the blackwell game cards do great. The new blackwell commercial cards are fantastic though.