Release of H3 Infinite Continuation Suite for ComfyUI: Create infinite length videos in consistently High Quality using Keyframes in FFLF-Mode (fl2v-Checkpoint)
The above video consists of 7 individual Minimax H3 clips generated in First-Frame-Last-Frame Mode, stitched together automatically without manual editing, upscaling or other post-processing.
Today I decided to release my experimental H3 Infinite Continuation Suite together with a set of workflows to make it easy to get started in ComfyUI.
The original idea was to combine the higher visual quality and keyframe control of H3's First Frame / Last Frame mode with the continuation capabilities of the Reference mode.
After quite a lot of experimenting, the output quality has reached a point where I hope some of you might find the nodes and workflows useful as well.
The example video was generated entirely with the included workflows at 736 × 1280, using 15 steps and no Turbo LoRA. I did cut a few seconds of nonsense speech from the very end because I was too lazy to regenerate the last clip. :D
How to get started
Install Herrgotts H3 Infinite Continuation Suite through the ComfyUI Manager.
Download the included workflows from GitHub.
Start with the `01_Start` workflow and provide your First Frame + Last Frame.
For every additional segment, use `02_Continue` and provide a new Last Frame for where you want the next clip to end.
Repeat for as many clips as you want.
When you're done, use `04_Stitch_Saved_Chain` to automatically combine the separately generated clips into the final video.
If you prefer to generate multiple chained clips in one workflow, use the included 3-Clip workflow. It contains the full continuation setup and is structured so you can extend it with additional clips without rebuilding the whole graph from scratch.
What the nodes handle automatically
- carrying motion and native audio into the next clip
- detecting and removing the frozen tail H3 often creates near the final keyframe
- choosing a suitable handover point between generations
- keeping the video and audio aligned
- smoothing the visual and audio transitions
- saving the individual clips so longer chains can be stitched afterwards without keeping everything in memory (no OOM, hopefully)
For the video above I used the default/recommended settings:
- Balanced Auto Handover
- 22 context frames
- Safe Tail Bridge: 2 frames
- Video crossfade: 4 frames
- Audio de-click: 15 ms
There are still occasional tiny brightness differences around some boundaries, but at this point I personally find them pretty difficult to notice during normal playback.
The pack is still experimental, especially when it comes to very long chains, different hardware configurations and prompt behavior. So if you try it, I'd be very interested in seeing your results and hearing what works or doesn't work for you.
GitHub: https://github.com/HerrgottMargott/Herrgotts-H3-Infinite-Continuation-Suite
ComfyUI Manager: search for `Herrgotts H3 Infinite Continuation Suite` or use "missing custom nodes" in one of the example Workflows.