Just trained first Qwen Image 2512 and it behaves like FLUX Dev model. With more training, it becomes more realistic with lesser noise. Here comparison of 240 vs 180 vs 120 epochs. 28 images used for training so respectively 6720 vs 5040 vs 3360 steps
Just trained first Qwen Image 2512 and it behaves like FLUX Dev model. With more training, it becomes more realistic with lesser noise. Here comparison of 240 vs 180 vs 120 epochs. 28 images used for training so respectively 6720 vs 5040 vs 3360 steps
Full tutorial : https://youtu.be/DPX3eBTuO_Y
Tested config : 200 epoch Tier1_29400_MB.toml
Imgsli full quality comparison : https://imgsli.com/NDM4NDEx/0/2
Comments (8)
These results are actually INSANE Dr!!
thanks a lot . i agree really good
Hey! Is SECourses_Musubi_Trainer_v26.zip the last version of the trainer app?
yes it is latest. it installs latest version configs are all valid and up to date. i trained old base model, 2509 edit, 2511 edit and 2512 new base model. all works perfect. will make a video hopefully about them
Thanks a lot for sharing, Furkan. Dont know if it's just me but i like 180 epochs as the best result. Find 240 slightly better expression but without any grain
true it is a choice. 180 is best of all
you keep burying recent posts you've made with all these rapid posts. I don't know if it's the crappy patreon interface or what. but all your linstalls are now replaced with training and face swapping posts... and this update. I thought you just released Comfy_UI_V71.zip update; can't find it or 70.zip
it is 100% crappy patreon interface. please use here : https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Patreon-Posts-Index.md comfyui is here : https://www.patreon.com/posts/105023709