Filmmaker renders a 35-second AI tracking shot with Singularity, patched SageAttention workflow
illumnat · reddit · 2026-10-19
An indie filmmaker who started with Disco Diffusion shares a 35-second walking/tracking AI shot (a sci-fi scene of a technician walking through a server corridor), pushing past the typical 15-second generation limit.
Workflow details: rendered at 1.6 MP, 20 steps with Singularity; two 5-panel reference images (left/right profile, 3/4 views, front) lock the character's face and body separately; rendered on RunPod's RTX Pro 6000 (96 GB VRAM, $2.49/hr) using a custom patched SageAttention — stock SageAttention causes "brown snow" artifacts at certain sizes; finished in DaVinci Resolve with AI Superscale 4X, color correction, noise reduction, and film grain. He also describes a two-stage setup: prep locally on an RTX 3070 Ti, render on cloud GPUs.
More from Multimodal
- Dreamy AI illustrations: furry-textured forests and snowy blue nights — OVolosin82152 · 2026-10-19
- ByteDance's SeedAudio 1.5 Brings Prompt-to-Cinematic-Audio to CapCut PC — FellMentKE · 2026-10-19
- "Light": a cinematic set of AI-generated city night scenes — gizakdag · 2026-10-19
- A reusable 'Cybernetic Neuro Nexus' AI image prompt template with color variables — LudovicCreator · 2026-10-19
- MOSS-Image launches: open-source 5B DiT image generation foundation model — bdsqlsz · 2026-10-19
- GIMP 3 plugin adds local Generative Fill, Remove and Expand via ComfyUI — Speedeee_ · 2026-10-19