MiniMax H3: Seamless Long Videos, No Color Shift — SelfLift Dual Sampling & Sharper Latent Upscaler
I updated my MiniMax H3 workflow for stitched long videos in ComfyUI. In my tests the join is now hard to spot even when I point it out, and the color stays consistent across segments. Three changes did most of the work: a small noise change on the context reference, a color transfer step, and a sharper community-finetuned latent upscaler.
What changed
Second stage uses a community finetune of the H3 latent upscaler (h3_upscaler_lms_v0.1) — visibly sharper in my tests.
Context mode now runs 4 low-res steps + 4 high-res steps at lowres scale 0.65 (~1.5x upscale). Fewer low-res steps and a smaller upscale ratio made the washed-out look and the color shift much less noticeable.
The reference frames right before the seam get 0.1 noise, and the last frame stays clean — this removed the flicker at the join.
Color Transfer (target_frame): match the new segment against the previous segment's last frame, then apply the transform to the whole new segment so the color stays stable end to end.
The new int8 VAE (ComfyUI 0.38.0+) roughly halves encode/decode time per ComfyUI's blog post.
The context reference is now a single switch — off, it behaves like a regular multi-reference generation.
Notes from testing
Sticking with the two acceleration LoRAs: 4-step v1.2 and 8-step 768. The recent 2-step LoRAs were basically unusable in my tests.
If a segment goes pale, lower the low-res step count and the upscale ratio first.
Links
Online workflow: https://www.runninghub.ai/post/2105882473606447106?inviteCode=rh-v1495
Sharper latent upscaler: https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI/tree/main/latent_upscaler
This workflow is super easy to use—I’ve uploaded a detailed tutorial to YouTube, so just follow the video along with this workflow to recreate the effect; please make sure to watch the full tutorial before starting to avoid common mistakes, and feel free to leave a comment if you have any questions!