SelfLift for MiniMax H3: shift most sampling steps to cheap low-res pass (6+2)
wjc_5 · reddit · 2026-09-19
The author updated his MiniMax H3 director-console workflow with SelfLift as the second sampling stage, moving most of the step budget into the cheap low-resolution pass: simple latent upscale (2 low + 6 high) → latent upscale model (4+4) → SelfLift (6+2).
The reason earlier versions needed many high-res steps: the upscaled latent was still noisy, so the high-res stage had to both adapt to the distribution shift and finish denoising. SelfLift instead takes the clean prediction from the end of the low-res pass, upscales it, re-adds the expected noise level, and rejoins the original trajectory — leaving high-res steps only for detail refinement.
Other points: a single high-res guide is resampled down for the low-res pass, keeping both stages spatially aligned and reducing seam offsets; the latent upscale model noticeably worsens color shift across stitched segments (the 2+6 scheme shifts less); a prompt workaround — starting each new segment with a short continuation of the previous shot that gets trimmed in editing — mitigates color drift; VAE repair params stay at 0 since this workflow uses the external H3 Upscaler. A full YouTube tutorial is provided.
More from Multimodal
- Experimental video mixes visual styles, generated with Seedance 2.5 on Adobe Firefly — umesh_ai · 2026-11-04
- Hailuo AI claims H3 scored perfect on language tests; Japanese creator praises its Japanese signage fidelity — Hailuo_AI · 2026-11-04
- YANAMI: 400K-dialog multimodal dataset shows models still fail at stateful multiturn interaction — Shengyu Wang · 2026-11-04
- Open-source Idol model turns a single image into 3D in 45s on 8GB VRAM — Promptmethus · 2026-11-04
- ComfyUI v0.44.0 released with new Patina texture nodes — Fcking_Chuck · 2026-11-04
- Krea 3 technical report released: a deep dive into data, architecture and training — bdsqlsz · 2026-11-04