Open-source Kinoloom turns an 8 GB GPU into a local AI film studio with seamless latent-space long takes
Dry-Acanthisitta-652 · reddit · 2026-10-26
A Redditor open-sourced Kinoloom, a Windows web app built on ComfyUI that runs a full local AI film studio on an 8 GB card (tested only on RTX 3070 + 64 GB RAM): any LLM writes the story, Qwen Image 2.1 generates the cast and first frames (38-78 s per still), and the pipeline renders, checks, and cuts the shots. MIT-licensed.
Key trick — latent-space joins: instead of starting each segment from the last frame (which causes visible jumps), Kinoloom passes each segment's ending into the next as latents and decodes the whole take in one pass. A 16-second LTX 2.5 take rendered in 4 segments took 28 minutes on the 3070; a 5-second clip takes 6 minutes warm.
MiniMax H3 is fully supported, but its license restricts use/output display in the EU, UK, South Korea and US — the EU-based author doesn't publish H3 clips; users elsewhere bring their own weights. Wan 2.2 (legacy) and five image models are also optional.
Solo project, tested on one machine only — feedback from other hardware welcome.
More from Infra
- GPU rental prices up ~80% this year as token demand climbs, argues analyst — JOBhakdi · 2026-10-26
- Musk warns of stranded compute; trader counters inference demand is near-infinite — markjeffrey · 2026-10-26
- $550 Dell R820 with 768 GB RAM: is an old monolithic server viable for LLM inference? — BobtheGodGamer · 2026-10-26
- Running DeepSeek Flash V4.1 IQ2_XS on Four R9700 GPUs: Real Numbers — poofph · 2026-10-26
- Kubernetes works for vLLM inference, but agent sandboxes needed Modal's custom stack at 1M+ concurrency — NielsRogge · 2026-10-26
- Anthropic reportedly offers Blackstone 10%+ yield on $9B junior debt for 1.5GW TPU deal — Beth_Kindig · 2026-10-26