Simple is easier, but less quality options when running.
T5XXL FP16 offloaded to CPU + Q8_0 is superior to the Fp8 method. same quality at FP16, less memory.
Confusing for people that were running this a week before this "combined checkpoint" was released.
I get endless comments from people who downloaded this version, when community did lots of research with kijai's wrapper to find the best combination of models and settings.
all details and links in thread.
https://github.com/kijai/ComfyUI-MochiWrapper/issues/26
Also - you can save .latent to disk and the load .latent from disk for the VAE spatial tiling Decoder, so there is no difference in speed.
it's nice to have a combined checkpoint for casual users - but they will become frustrated when they see the loss in quality with their outputs.
Thanks. I'll update my Donut-Mochi workflow pack to add a new option for casual users who prefer the "Straight FP8" setup used here./
examples :
V8 included "true image to video: https://civitai.com/models/886896
V7 Gallery with pseudo i2v and solid t2v: https://civitai.com/posts/8613066
Is it possible to run in colab ( I don't know coding)