Attempt to build sage attention 2.2+ it speeds up MiniMax significantly.
Tested on rtx5070ti.
Diffusion model:
Text encoder:
VAEs:
Turbo LoRA:
Ollama LLM (gemma 4 12B it) for prompt helper:
You can add as many references as needed..
DOWN BELOW IS OLD VERSION CONTENT.
DISABLE comfyui_smznodes if you have this extension!!!
(It breaks any LTXV workflow.)
Diffuser model (pick one):
Text encoders (both):
VAEs:
Latent upscaler:
LoRA (distilled LoRA is needed, do not disable):
Enjoy!
Description
Added ollama prompt helper, tuned optimization/quality.
FAQ
Comments (12)
I missing 2 node
-Solattnpatch
-spectrumapplyminimaxh3
It's just efficiency nodes you can bypass and it ll work, if you want it just get it's repos from git by cloning it to comfyui/custom_nodes as instructed in the note of workflow
In the PARAMS / MAIN node group there is a sigma shift node with some example values. How did you get them, is there a documentation for other values like 0.2 MP?
You can experiment with going lower (to like 10), but based on latest findings and documentation, minimax H3 doesn't work like flux.. Meaning you should just stick to 12 video 3 audio, if audio is broken due to turbo LoRA you should try 12 video 6 audio shift.
@blinkdotleh thank you
Hey bro, thx a lot for the workflow ! Could you tell me how and where i could add a lora node ?
Use normal load Lora and connect it right after load diffusion model/unet loader.. If you want to try the turbo LoRA, find the original author on huggingface, you ll need to follow his instructions, but in my testing turbo LoRA underperformed
The upsampling module performs extremely poorly. It consumes more memory than the minimax h3 in generating videos. The additional time consumed is already sufficient for the minimax h3 to generate a high-resolution video natively. This workflow has successfully wasted my time.
Ltxv upscaling has even worse quality, traditional esrgan flickers, ultimate SD flickers as well and needs an SD model, rtx VSR is harder to install and worse quality.. Original mh3 upscale is pay walled (closed source). This is the best local option I found.
Assuming this is true, high resolution Minimax also can start introducing weirdness to the generation, adding strange details and body horror. Not in all kinds of video concepts, but in some kinds. Still, I'd want to do SeedVR2 as a separate action, instead of in the same workflow, so you can save generation time by doing tests and then only upscaling the exact clip that worked out. That would certainly save generation time.
@Jellai just bypass upscale node, it does exactly what u want.