Comfy-Org

/

MiniMax-H3

like 732

Follow

Comfy Org 11.5k

As you said it's a wide shot, your character face takes much less space on screen than if It was a vertical aspect ratio for the same total pixel count, the more smaller the head will be on screen, the worse it will look.

I don't recommend the tip you gave, if you want better looking video, you can start low res to see if your prompt is on par with what you wan't, if it is, just restart the same gen at higher resolution.
Using upscalers that will re-encode to latents and decode again will destroy initial details, add artifacts, give worse temporal motion

Using 'EasyCache' node also degrades the quality of the final result.

UntMods ... Maybe you didn't read the post closely enough. I'm talking about fast iterations there. For instance, when you need to quickly put together a demo for a presentation. The issue that some other users might have noticed too is that render times grow exponentially as the resolution, video length, and the number of steps increase. Compared to LTX, iterating is quite a pain if you're stuck with limited, low-VRAM hardware.

So, this tip isn't for you, but it might save time for those who don't own an RTX 5090. And from what I've gathered, even people with an RTX 5090 aren't exactly thrilled with the render times for Minimax H3. Time will tell. It's open-source, and everyone can choose the workflow that suits them best—whether they use Wan, LTX, Minimax H3, or anything else.

Feel free to share your own samples and let us know what’s working for you and what isn’t, so others can learn from it too!

Cheers!

That's generally the case with all generative AI models, including image generation.

In VFX, there are a couple of common ways to deal with it:

• Increase the base resolution before generation, if your hardware can handle it. (Unfortunately, this isn't an option for everyone due to VRAM and compute limitations.)
• Repair only the affected region. In your NLE, crop the area where the quality has degraded, render that crop, then run it through your preferred workflow again using a ControlNet (such as Depth) with fixed start/end frames to preserve temporal consistency. Afterward, bring the result back into your NLE and blend it using feathered masks or rotoscoping.

If the degradation is relatively minor (not as severe as in your example), you can often get away with a light 0.15 denoise V2V pass instead of the more involved ControlNet + fixed start/end frame workflow.

Hope that helps!

Mike, I think the confusion is you lead with a rhetorical question to the community. You already knew the answer and were offering a solution.