Quality worse with npfp4 model. Is it only good with fp16 fp8?
Quality worse with npfp4 model. Is it only good with fp16 fp8?
yeah sorry I tried as well doesn't seem to work as wella s the base model.
minimax_h3_ref2va_pruned_int8_convrot here weird glitches such as object disappears even if not prompted. any chance to fix it in next release?
I agree the lora need some fixes. Me personally noticed people glitching even if it's main subject in reference to video limbs become deformed and much more things that shouldn't happen, never had this issue without lora
Hello, gonna have a look to this comment, if you could send me any bad result you get ? But yes the model lora is made normally for fp16 or fp8 . Please send me on X or Linkedin , Lovis Odin/
Hello, gonna have a look to this comment, if you could send me any bad result you get ? But yes the model lora is made normally for fp16 or fp8 . Please send me on X or Linkedin , Lovis Odin/
I have same strange glitches but for me happens often when generating on lower resoltuion. I'm using mostly reference to video. I don't remember 1 megapixel when I had this issue but ~ 0.2 it happens 75% of the time. I believe that's what most people do: try in lower resolution, fix the prompt few times so finally I will get what I want then I switch to 1 megapixel for final generation, but with this lora I would be forced to do that in 1 megapixel all the time which takes few times more of compute time. To be more detailed: no sageattention and other tricks, just official comfy workflow on res_multistep, beta + switched to 32 steps (minimax recommends from 30 to 50 steps)
Please fix the stability. I hope we can hear back from you, thank you.
yeah sorry I tried as well doesn't seem to work as wella s the base model.
So the main issue are the combo of pruned models + turbo loras, that's why when it's on most of the time it does more bad than good. Yeah that would have to be expected. This lora would need a rework from the community to get proper results on limited hardware.
Yeah this model doesn't seem to work at all with ComfyUI int8. I swear it worked before so there must've been a ComfyUI that changed something on the lora loading side or I'm hallucinating.
Best workaround I've found is to use it at 0.5 strength, however I hope fal team will fix the stability someday.
Ok so much feedbacks, thanks you on that, will work on a new version this weekend that hopefully will fix your issues !