Geiru Context Suite test: a 2-minute no-degradation AI shot, rendered locally in ~2 hours
illumnat · reddit · 2026-11-07
A traditional film editor shares a first test of the open-source Geiru Context Suite, producing a 120-second, 2,880-frame shot with virtually no quality degradation.
Workflow & settings
- Model: MiniMax H3 Singularity v2 INT8, plus a custom character LoRA trained with Fizwig on 99 ChatGPT-ImageGen-generated images (strength 1.0)
- 1376×768 @ 24fps, resmultistep sampler, 20 steps, Sage Attention, dynamic VRAM disabled
- Render took 2h9m on a rented RTX PRO 6000 (96GB VRAM)
Issue found: two hard jump cuts at 35.6s and 44.1s, diagnosed via ChatGPT sub-agents as landing exactly on generation-window handoffs — only 2 of 13 handoffs show conspicuous cuts. Root cause hypothesis: the 20-local/20-total-steps setting finishes each window independently with no joint refinement across overlaps, so adjacent windows disagree on framing and pose.
Next test: reserve two steps at transitions for joint refinement while matching everything else — a testable fix. The author is very happy with the result and plans to keep using the workflow.
More from Multimodal
- AI-generated short film clip sparks debate: 'all it needs is a good script to win awards' — Flashy-Impact5757 · 2026-11-07
- Moss INT8 SFT is enough: a simplified workflow beats fancy settings for photorealism — YentaMagenta · 2026-11-07
- Creepy AI-generated 'Delirium tremens' video of two ghoul-like drinkers goes viral — Tall-Key2859 · 2026-11-07
- ComfyUI-freefield puts Freefield's 117 workflows in a sidebar with one-click model downloads — neuregex · 2026-11-07
- Endless Sky: a continuous first-person street flythrough showcasing long-horizon video generation — MikirahMuse · 2026-11-07
- Local GLM video gen on dual RTX Pro 6000: user produces winter scene with sound in ~20 minutes — HankYeomans · 2026-11-07