Running MiniMax H3 on 24GB? One workaround pipeline for the 1080p second-pass OOM
newaccount47 · reddit · 2026-10-31
A Redditor reports that true 1MP two-pass generation with MiniMax H3 ref2va (1344x768 pass 1 into 1080p pass 2) crashes with a CUDA illegal memory access on a 24GB 4090, using Singularity v1.3 int8 at 141 frames—pass 1 runs 43 s/it with weight streaming.
Working fallback: generate at 1152x640, apply a 1.5x learned latent upscale, 2 steps with the ref2v turbo LoRA at 0.2 denoise, then RTX VSR up to 1080p—about 9 minutes per 6-second clip. The author asks whether anyone has a better route to true 1MP + 1080p on 24GB cards.
More from Multimodal
- Director runs 'vibe directing' with Claude Code: per-shot approvals, $0.82 clips — DavidmComfort · 2026-10-31
- Ostris trains MOSS Image with Primal VAE: 4K at 1080p compute cost — Crazy-Repeat-2006 · 2026-10-31
- Creative director's wordless AI short: 119 shots distilled from hundreds of generations — Superb-Meeting7108 · 2026-10-31
- Moiré Pattern Enigma prompt: overlapping lines create shifting AI art — LudovicCreator · 2026-10-31
- AI video of the week: Jim Cramer must save America — mark_k · 2026-10-31
- Zed AI demos cinematic portrait generation with film-grade motion blur — azed_ai · 2026-10-31