Its super super fast!! thank you! :)
hi! i just wanna thank you for making the "DasiwaWAN22I2V14B" NVFP4 new models they work extremely fast on my 5090! for 720p it takes only 10s for a step! :) the output was amazing!!!!
I also tested it, and the speed is indeed very fast.
However, the I2V image quality and character consistency seem to drop significantly.
Is there any way to improve this?
I also tested it, and the speed is indeed very fast.
However, the I2V image quality and character consistency seem to drop significantly.Is there any way to improve this?
i had no issues at all, everything works very well, try using this workflow https://civitai.com/models/1847730?modelVersionId=2264611
I also tested it, and the speed is indeed very fast.
However, the I2V image quality and character consistency seem to drop significantly.Is there any way to improve this?
It is like the rogue quant of gguf's an other guy made, the quality is absolute trash in comparison... But well stealing without any credit is the way to go for some.
There is no way to get the same quality, because they dequant and requant my model, loosing significant quality and understanding.
Just make the same video with fixed settings and you will see ... All what this does is destroying my model's quality and reputation...
I also tested it, and the speed is indeed very fast.
However, the I2V image quality and character consistency seem to drop significantly.Is there any way to improve this?
It is like the rogue quant of gguf's an other guy made, the quality is absolute trash in comparison... But well stealing without any credit is the way to go for some.
There is no way to get the same quality, because they dequant and requant my model, loosing significant quality and understanding.
Just make the same video with fixed settings and you will see ... All what this does is destroying my model's quality and reputation...
My apologies, I have deleted the models.
For me personally the output quality was close to the gguf q4, but faster. Also I released my quantization script 2 days ago. https://github.com/gitmylo/ModelConverter
I think it would really be appreciated if you could make official nvfp4 quants of the models. (for anyone else, you can use the script I made, but what darksidewalker said is true, requanting a model is going to lose quality. Additionally my script currently doesn't do calibration, so no input scale is used, which also loses accuracy.)
Thank you, sorry if I was harsh, but maybe we can find a solution and cooperate to make it happen with satisfying quality. Than I consider to release one official version, so everyone is happy.👍