Open-source CLSS streams arbitrarily long audio+video on a 16GB GPU, cutting DiT evals by ~50%
nazgut · reddit · 2026-10-01
Reddit user nazgut shipped a major update to CLSS (Closed-Loop Streaming Synthesis), a ComfyUI toolkit that turns H3 — capped at 5–15s per pass — into an arbitrarily long streaming audio+video generator by overlapping chunks with drift control at the hand-off, no weight changes. It runs on a 16GB card: int8 DiT, a 5.5GB Qwen3-VL-4B text encoder replacing the 15.7GB 32B, 832×480, 10s chunk windows, tested on an RTX 3080 Laptop.
Key additions:
- Multi-scene prompts with proportional allocation, two-step crossfades, shared globaltext.
- R2V per-scene image/audio anchors (<Picture N>/<Audio N>); <Video k> video refs auto-resampled to 24fps with original soundtrack.
- i2v via first-frame keyframe pinning.
- Audio continuity kit: sample-exact seams, waveform-refresh continuation, seam pinning, export-only loudness anchor, loop guard re-rolling vamp takes, equal-power crossfades.
- Continue & re-edit from saved frames/audio, or rebuild single chunks in place.
- Per-chunk neural latent upscaling that never leaves VRAM.
- Spectrum hidden-state forecasting (adapted from the Spectrum paper): forecast steps skip the whole DiT stack, turning a 6-step turbo chunk into A A F A F A — 50% of DiT evals skipped.
- 16GB hardening (unload-all-models, CUDA expandable segments).
Repo: https://github.com/nazgut/ComfyUI-MiniMaxH3-CLSS
More from Multimodal
- AutoTrust AI debuts JEV-27B-VL, an open-weight multimodal decision model that acts on what it sees — alexcovo_eth · 2026-10-01
- Mixie3D launches: AI agents model, texture and light 3D scenes in parallel — gowthami_s · 2026-10-01
- A Qwen outpaint LoRA beats native outpainting on quality and consistency — linoy_tsaban · 2026-10-01
- Midjourney Artist Opens Her Entire Archive and Site mjpro.ai for Free — ciguleva · 2026-10-01
- Seeddance 2.5 used to create a 3D animated Louvre heist scene, prompt shared — azed_ai · 2026-10-01
- Orbit LoRA plus first/last frames freezes time in MiniMax video with 360 camera moves — AndrewJumpen · 2026-10-01