Repolode
Submit
tazmannner

tazmannner/ComfyUI-H3-RelayHires

ComfyUI nodes that render MiniMax H3 video at 1920x1088 by upscaling a 0.5 MP take and redrawing it, aiming for about half the time of a 1 MP upscale route.

View on GitHub
Stars
54
Forks
5
Open issues
0
Last commit
4 days ago

What is it?

H3 Relay Hires is a set of ComfyUI nodes for MiniMax H3 video generation, aimed at 2D anime. The method renders the take at 960x544 to decide layout, motion, acting and audio, then upscales the decoded frames to 1920x1088 with AnimeSR, a small anime upscaler trained in the project, and runs one DMAD denoising step to redraw detail while keeping the audio pinned to the first pass.

The author reports that on an RTX 5090 a 5-second clip finishes in about 116-121 seconds, compared with roughly 240 seconds for a 1 MP render with latent upscale and refinement and about 500 seconds for a native 1920 render. An optional H3 Fast Tail Sampler shortens the first pass further, and a Pika H3 Flex LoRA can replace that pass for maximum speed at the cost of a different take.

The repository includes example ComfyUI workflows that add face detection, de-rope motion repair and hold-frame handling, a Sol-H3 VAE option for a brightness flicker, measurement scripts for line width and flicker, and training scripts for the AnimeSR model.

Key features

  • 01Renders the first pass at 0.5 MP, then upscales to 1920x1088 with a redraw step.
  • 02Keeps the preview's take and soundtrack in the final 1920 result.
  • 03Includes AnimeSR, a small anime upscaler that can be trained on your own clips.
  • 04Optional H3 Fast Tail Sampler reduces first-pass time without changing the take.
  • 05Accepts any ComfyUI upscale model as an alternative to AnimeSR for non-anime looks.

Use cases

  • Producing 1080p-class anime-style H3 clips on a single GPU with less waiting per render.
  • Previewing a take at low resolution and then finishing the chosen take at 1920 without changing it.
  • Comparing upscalers or training a custom anime upscaler for a specific art style.

Things to know

  • Timings were measured on an RTX 5090, so renders on weaker GPUs will take considerably longer.
  • Very fast action can come out as motion blur rather than crisp poses.
  • Ink lines remain somewhat thicker than in a native 1920 render.

In collections

AI video & audio toolsAI image generators & Stable Diffusion tools

Topics

#comfyui#video-generation#upscaling#anime#minimax-h3#stable-diffusion#image-super-resolution

FAQ

What is ComfyUI-H3-RelayHires?
ComfyUI nodes that render MiniMax H3 video at 1920x1088 by upscaling a 0.5 MP take and redrawing it, aiming for about half the time of a 1 MP upscale route. H3 Relay Hires is a set of ComfyUI nodes for MiniMax H3 video generation, aimed at 2D anime. The method renders the take at 960x544 to decide layout, motion, acting and audio, then upscales the decoded frames to 1920x1088 with AnimeSR, a small anime upscaler trained in the project, and runs one DMAD denoising step to redraw detail while keeping the audio pinned to the first pass.
Is ComfyUI-H3-RelayHires free to use? What license does it have?
ComfyUI-H3-RelayHires is open source under the MIT license. Check the license terms before using it in commercial products.
Is ComfyUI-H3-RelayHires still maintained?
Actively maintained. The last commit to ComfyUI-H3-RelayHires was on Nov 3, 2026.
What can I use ComfyUI-H3-RelayHires for?
Producing 1080p-class anime-style H3 clips on a single GPU with less waiting per render. Previewing a take at low resolution and then finishing the chosen take at 1920 without changing it. Comparing upscalers or training a custom anime upscaler for a specific art style.
Where can I get ComfyUI-H3-RelayHires?
The source code and installation instructions are on GitHub: https://github.com/tazmannner/ComfyUI-H3-RelayHires

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects