MiniMax H3 seamless long-video workflow: dual sampling, color transfer and sharper latent upscaler kill visible seams
wjc_5 · reddit · 2026-10-06
The author updated their ComfyUI workflow for stitching long MiniMax H3 videos, reporting seams are now nearly undetectable with stable color across segments. Key changes:
- Community-finetuned H3 latent upscaler (h3upscalerlmsv0.1) for visibly sharper output
- Context mode now runs 4 low-res + 4 high-res steps at lowres scale 0.65 (1.5x) — fewer steps and smaller upscale cut washout and color shift
- 0.1 noise on frames just before the seam with a clean last frame removes flicker
- Color Transfer matches each new segment to the previous segment's last frame, then applies the transform to the whole segment
- int8 VAE (ComfyUI 0.38.0+) roughly halves encode/decode time; context reference is now a single switch
Notes: stick with the 4-step v1.2 and 8-step 768 acceleration LoRAs (recent 2-step ones were unusable); if a segment goes pale, lower low-res steps and upscale ratio first. Full tutorial on YouTube.
More from Multimodal
- Qwen Image 2.1 Uncensored MCP: a full image pipeline in 12GB VRAM — Delicious-Farmer-234 · 2026-10-06
- 8 prompts turn Claude Opus 5.5 in Claude Code into an explainer video studio — alex_verem · 2026-10-06
- MetaCanvas lands NeurIPS: MLLMs plan directly in diffusion latent space — mohitban47 · 2026-10-06
- Ming Image 0.1 in ComfyUI: free AI model tested for text-heavy design work — pixaromadesign · 2026-10-06
- One-word Midjourney v8.2 prompt: 'Diaphanous' renders sheer translucency — tisch_eins · 2026-10-06
- Reddit User Shares AI-Reimagined 'Princess and Dragon' Short Video — Throwaway350750 · 2026-10-06