hmm maybe i forgot to delete the download text from the link
hmm maybe i forgot to delete the download text from the link
F\SwarmUI\Models\diffusion_models
Jcan be fixed? or we need to wait?
XRenamed one of the safetensor files inside of \SwarmUI\Models\diffusion_models and delted the "?download" and it worked
Fwait testing
Fok working gonna update now. i hope nothing else broken :d
Ftesting
Ffixed
Fplease run update file
Fit should fix
Jthanks!!
JI'm ending a video, I'll update when it ends
Fsure
JSame prompt but with 14B model (FP8)
Fwhat a difference
Fi hope nothing broken please test :d
1Hey, did I read correctly that with the new Flux Kohya training I can use the 48Gb preset with a 5090 ok?
JWAN 2.1 14B Image-to-Video 480P FP8 working fine now. Thanks!!
Fi hope results accurate
Flet me know
Fit fits into vram :d
1I thought that was case . Just testing a fresh install. It started ok but seems to have froze and not calculating it/s
Fit works. make sure that you vram usage is below than 2 gb
Fit uses like 28 gb
Falso these are when all images 1024x1024
Fif you have different res it impacts
1images are fine, Will check to free up memory. Just noticed I wasn't using onboard GPU for monitor.
1crippled it
Fye using too much vram
Fyou are using 48GB_GPU_28200MB_6.3_second_it_Tier_1.json right?
148GB_GPU_Quality_Tier_1_40000MB_9.3_Second_IT_Trains_T5_and_T5_Attention for Lora
1sorry, it's late. it's 40Gb needed for that.
Fit literally says 40 gb 
1I'll drop it to the Tier_3
1long day.
Jsomething is wrong with img to video 14b FP8. After 2-3 frames goes full pink image
J39 minutes to generate it with teacache and FP8
JDoing a second video. I'll post the result here
JWAN 2.1 14B Image-to-Video 480P FP8 is broken
JAnd the negative prompt box is always empty when you start the app
Jshould have the default or saved profile prompts, like previous versions
T@Dr. Furkan Gözükara I just come back and thank you for:
1. the Dreambooth training for 5090 is working ! (In windows)
2. the MAssed Cloud got 300G version for me to train my models !! U are a star ! @Dr. Furkan Gözükara
1. the Dreambooth training for 5090 is working ! (In windows)
2. the MAssed Cloud got 300G version for me to train my models !! U are a star ! @Dr. Furkan Gözükara
Jhey brothers, is there a way to make social banners without letters/fonts errors? with stable diffusion?
JI think the latest DiffSynth-Studio fix for FP8 on WAN 2.1 14B Image-to-Video 480P has broken the FP16 quality too. I'm testing with previous commit before the FP8 fix
JConfirmed
J@Dr. Furkan Gözükara Please revert (or fix) the latest DiffSynth update. It breaks colours and quality in FP16, and in FP8 the video turns completely pink. I think it only affects WAN 2.1 14B Image-to-Video 480P model
P@JaviEM i do noit have that iissue .. strange..
I do have a issue so running video to video
RuntimeError: The size of tensor a (36) must match the size of tensor b (13) at non-singleton dimension 2
I do have a issue so running video to video
RuntimeError: The size of tensor a (36) must match the size of tensor b (13) at non-singleton dimension 2
JI reverted to previous DiffSynth commit and it's working fine again, so it's a problem from the latest commit
POk that is weird .. for me it seems to work fine on rtx 3090
Fthe offical repo authors made some fixes