This is a MiniMax H3 workflow focusing on a few problems I've been trying to fix: the oily/plastic look from acceleration LoRAs, blurry distant faces, and quality loss in high-motion scenes. I also added a two-stage latent upscale workflow, so you can generate a quick low-res preview first, find a good seed, and then refine it to around 1080p. This saves quite a bit of time compared to doing the full-resolution generation every time. The workflow also uses a merged H3 model, so text-to-video, image-to-video, first/last frame, and multi-reference generation can all be handled in basically the same workflow.
For more details, please check the note within the workflow.
If you don't have the hardware to run this workflow locally, you could try this online (Free 48GB GPU）.
Links:
Cinematic Version:
https://www.runninghub.ai/post/2101654857613963265
V1:
https://www.runninghub.ai/post/2094673232666742786
Check the tutorial video for instructions before using this workflow:
Cinematic Version:
V1:
Description
Fix the overly white skin tone and loss of details in the V1 version. No more weird artifacts in the last few frames. no more broken sound in the first stage.
FAQ
Comments (3)
It’s an excellent workflow; I’ve already tried it out after watching your YouTube guide. It’s very versatile and delivers great quality. It would be perfect if you could integrate a way to merge the various clips. I had recently discovered the "Minimax Director" workflow, which offers that feature, but it lacks the highly effective upscaler found in your workflow. For me, the ideal setup would combine the final quality of your workflow with the convenience of being able to merge clips. I hope you’ll consider this!
Thanks also for your guide on LLMs—that was excellent too.
模型及lora下载地址发一下呗
Interesting but I get errors when using Loras: ERROR lora diffusion_model.blocks.14.adaln_proj.linear.weight shape '[96768, 8]' is invalid for input of size 260112384