Results are better if we use native nodes than the turbo nodes
I tried multiple generations and I found that using native nodes produce much better results than turbo nodes. I used the script from https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/discussions/6 to convert these weights to use with native Lora loader.
Currently running some tests to check If I'm doing something wrong with turbo nodes.
Will post the results later, I have potato PC.
UPDATE: I ran multiple tests and found that setting lora strength to 0.8 on turbo lora loader node by Larryvrh did came close to the results of my native setup, I'm still noticing some issues but I will continue testing to check if I can find the sweet spot on turbo loras to match my native setup which I actually like more.
Are you talking about the lora loader node?
I'm still running tests though yeah I'm talking about using only native ksampler select and Beta + Native lora loder working better for me with best results
dit video+audio models won't work good with Euler/Simple or Normal samplers, with minimax dpmpp_2m and sgm_uniform is the absolute best sampler/scheduler. therefore fp16/fp8 will give u the best combo of speed and quality, try 8step and shift 8.0, for conversation or minimal movements scenes use 5 or 6 step, i just did some testing on this lora, still dpmpp_2m is the fastest. in minimax both video and audio share 50 blocks, there are no separated video blocks, and lora by design is limited and will kill the audio part , good at some video, bad at everything else or make it sound like old tts models
Let me try this combination too and share my views, will be using bf16 SORRY I FORGOT AND USED INT_8_CONVROT HERE . And also I'm not looking for fastest, I'm trying to find the best combo.
Update
Dpmpp_2m + Sgm_Uniform gave me good results tho still morphing is worse than Euler + Beta,
Euler + Beta :-
=====================================
Dpmpp_2m + Sgm_Uniform :-
============= VIDEOS BELOW ===============
Euler + Beta :-
Dpmpp_2m + Sgm_Uniform :-
Both took same time and book looks good, but Euler + Beta morphing is not that worse
input image: screenshot of your video. not bad for 544x960 at 98 sec, with no 4step lora
High res image to test more: Please try different combos anyone who can.
High res image to test more: Please try different combos anyone who can.
Thanks, I was so fixated on speed that I totally missed the morphing issue. maybe the fp16 vae is the issue. I’ll try a full vae, and if I manage to find the reason, I’ll post it here :) thanks for mentioning it I didn’t notice before.
Thanks, I was so fixated on speed that I totally missed the morphing issue. maybe the fp16 vae is the issue. I’ll try a full vae, and if I manage to find the reason, I’ll post it here :) thanks for mentioning it I didn’t notice before.
It's fine, and yes please if you found the culprit do let me know as well, tho video vae is fp16 right? Int8 is the new light weight
Thanks, I was so fixated on speed that I totally missed the morphing issue. maybe the fp16 vae is the issue. I’ll try a full vae, and if I manage to find the reason, I’ll post it here :) thanks for mentioning it I didn’t notice before.
It's fine, and yes please if you found the culprit do let me know as well, tho video vae is fp16 right? Int8 is the new light weight
based on experience, I think BF16 for the vae is the best option else, fp32. int8 for detailed video, especially for action full scenarios, won’t perform good at the long run you'll end up doing more re-runs
based on experience, I think BF16 for the vae is the best option else, fp32. int8 for detailed video, especially for action full scenarios, won’t perform good at the long run you'll end up doing more re-runs
Where did you find fp32 VAE for video? fp32 for audio, for video max is fp16
https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae
based on experience, I think BF16 for the vae is the best option else, fp32. int8 for detailed video, especially for action full scenarios, won’t perform good at the long run you'll end up doing more re-runs
Where did you find fp32 VAE for video? fp32 for audio, for video max is fp16
here it is: https://huggingface.co/MiniMaxAI/MiniMax-H3/tree/main/FL2VA/video_vae/source
it is still there, hmm , i think the issue is the duration + prompt. 24 frames, 5 second and that jump = lots of pixel movement in small time window. the model tries to make it all happens in that duration 24 frames i think it is not enough for having all frames looks perfect, not enough pixel and frames for smoother more well defined images par frame, i'll test longer duration and frame interpolation, will see