Minimax-H3 - Best x2 Upscaler-Refiners

Описание: Minimax-H3 - Best x2 Upscaler-Refiners

I've found a H3 Upscaler that uses a 2 sampler pass and works best of all so far, but it is slow. I've also improved the LTX 2.5 (3mp upscaler!) with a small update. That is much faster workflow and results are pretty good. Both are usable and are what I will be working with moving forward.

This will probably be the last specifically Minimax H3 based video until some other model drops. Now I will go back to making content, and use the thing to see what it can do. So the next video will be a few days away, and probably be more about content and the results I achieve, as well as any tweaks to the workflows along the way.

The two workflows discussed in this video are:

"Minimax-H3\MBEDIT - MH3_r2v_DualSampler_v12.json"

"MBEDIT-v2v_LTX25_ResizeRefiner-w-SingleSampler_vrs6.json"

LINKS

Follow me on Patreon for free workflows, AI info, and other tools -   / aimakingmovies  

Latest Minimax H3 workflows - https://github.com/mdkberry/comfyui_w...

Latest LTX2.5 upscaler/refiner workflow - https://github.com/mdkberry/comfyui_w...

Lightx2v Lora that I use from Kijai - https://huggingface.co/Kijai/MiniMax-...
(theres been updates but I havent downloaded them yet)

Also "silveroxides" light2xv, but it needed a fix for the adaln_proj error, and the result is here https://huggingface.co/nynxz/H3_Loras...

Comfyui needs to use Cuda130 or above for this to work, and you need it updated to August 2026 commits (latest is best) - https://docs.comfy.org/installation/c...

Int8 models from here - https://huggingface.co/Comfy-Org/Mini...
(The official workflows are in the model card)

W4a8 is experimental new model type, you need to be updated on Comfyui but you can get it here https://huggingface.co/Kijai/MiniMax-...

Comfyui Kitchen Attention is part of Comfyui if you update to latest. I find it faster than Sage Attn on a 3060 RTX.

Official prompting guides:
https://huggingface.co/MiniMaxAI/Mini...
https://huggingface.co/MiniMaxAI/Mini...

Point your favourite LLM at one of the above links depending on your model you are using, and give it your prompt idea and it should sort it out.

For more details on speed-up loras see the last video at    • Minimax-H3 - Speeding It Up  
or the one before that for more on sigma nodes (currently bypassed but still in the workflow)

OTHER LINKS

AIMMS Storyboard Management software is available for trial or license from here - https://markdkberry.com/software/

My AI short films and music videos:    • AI Videos  

Books & Stageplays: https://www.amazon.com/stores/Mark-DK...

Music: https://markdkberry.bandcamp.com/

Website: https://markdkberry.com