Quality Gap Between Qwen 2511 (4-steps) and 2509 Lightning (4-steps)
#7
by
David-PHR
- opened
Hello,
Thank you for all the work on the Qwen optimization.
I’m currently unable to achieve the same quality with Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors as I had with 2509. I attempted to reuse the same scheduler you previously shared for 2509 (Qwen-Image-Lightning), but the results still fall short.
Do you have any guidance or updates—such as a new scheduler—that could help improve the output?
Want to come to say the same thing. I found that
2511 + 2511 4 steps lora is cartoony.
2511 + "Qwen-Image-Edit-2509-Lightning-8steps-V1.0-bf16" is more real.
Best real config I found so far is 2511 + "Qwen-Image-Edit-2509-Lightning-8steps-V1.0-bf16" + 8steps + cfg2.5