Turns out there's an amazing t2i model hiding inside Qwen Image 2.1, and it just needed a bit of finetuning to fix the noise problem and shore up the detail. This is my take on it.
It's awesome at editing too. I recommend the workflow and associated comfy node that adds an improved sampler, as well as RES4LYF for the bong_tangent scheduler.
https://github.com/envy-ai/ComfyUI-DPMpp-2M-Sharp
https://github.com/ClownsharkBatwing/RES4LYF
NOTE: Model is re-uploading. Check back in 15 minutes!
Description
FAQ
Comments (9)
Is this the full 2.1 or the turbo? The showcase images look good, so I can't wait to try it out when it's done uploading.
One of the issues with Qwen 2.1 is the base CFG of 1 appears to be totally wrong, often showing clear CFG underbake artifacts similar to back in the SDXL days. What's your opinion on that?
I found out today that the under/overbaked look was, at least in my case, because I chose resolutions that were not divisible by 32. Whether CFG was 1 or 3+. So make sure it is correct res first. then if you still have issues, maybe some Lora you use causes the problem ?.
My opinion is that it's not a CFG-distilled model like they claimed, because negative prompts work well and raising the CFG above 1 improves it. I have no idea why they would do that -- maybe they tried to CFG-distill it and failed. I run it between 3 and 4. If you run it at 1, you're gonna have a bad time.
Just curious, is it INT8CR with what group size?
BF16 would be kinda lovely.
Preliminary testing: The image does not get blurry when the prompt exceeds 500 tokens, which is an improvement over the original version. When using the recommended DPMpp‑2M‑Sharp + bong_tangent combination, the results feel slightly inferior compared to res_multistep + bong_tangent. Outputs from DPMpp‑2M‑Sharp + bong_tangent tend to be softer, whereas res_multistep + bong_tangent delivers better sharpness and clarity.
That said, this model may produce rendering errors for Chinese text when using DPMpp‑2M‑Sharp. Most importantly, certain aesthetic LoRAs have poor compatibility with this model and will cause image corruption. Examples include: https://civitai.com/models/2972193/qwen-image-21-beautiful?modelVersionId=3368237 and https://civitai.com/models/2986100/qwen-image-21-glamour-realism-lora-16-gb-tested?modelVersionId=3385496
Additionally, I fully agree with the author’s viewpoint: QW21 is a highly promising model with plenty of creative potential. Its features for character reference, prompt adherence, and in‑image editing & correction hold great development value.
You also did "Qwen Image 2.1 Fix v2.0". Whatever it adds, it is that in this model too?