LynnReal-Omni: one open model generates, edits and restores video
The 28B Flash weights are already on Hugging Face, rendering 540p in 377ms on one H100 — plus an OpenAI SDK breaking change and a self-host break-even tool.
Copy markdown
One model for the whole visual pipeline
LynnReal-Omni (paper out Sept 14) does text-to-video, image- and reference-conditioned generation, structural control, editing, degraded-video restoration and long-video — plus joint audio — from a single 32B diffusion transformer, instead of stitching four separate tools together.
377ms for a 540p clip on one H100
The 27B "Flash" variant generates and decodes a 22-frame 540p clip in 377ms; the full 32B does it in 843ms, aided by a lightweight VAE decoder. That's near-real-time video off a single card.
The Flash weights are already on Hugging Face
stdstu123/LynnReal-Onmi-flash-beta-0.1 ships open: a 28B, 42-block joint video/audio DiT with INT8 (W8A8) quant and safetensors shards. Heads-up — it needs the repo's custom loader, not plain Diffusers.
Build this weekend: an agentic visual workflow
It accepts 3D renders and game recordings as inputs, so you can wire one endpoint that ingests a rough render, restores and edits it, and lays audio — the "agentic visual workflows" the paper is built around.
The catch: beta, 540p, self-scored
It's a beta checkpoint capped at 540p/22-frame clips on H100-class hardware, and the headline wins come from MSAVP, the authors' own 100-prompt benchmark. LICENSE files sit in the repo but are unstated — check terms before you ship.
Elsewhere: OpenAI's Python SDK 3.14 changes error codes
v3.14.0 (Sept 14) normalizes API error codes to strings, bounds vector-store file polling, and normalizes errors raised while reading streams. If your retry logic branches on numeric codes, audit it before upgrading.
Elsewhere: a break-even calculator for local vs. cloud
Sunk Cost, on the Hacker News front page today, tells you how long until a Mac, DGX Spark or Strix Halo box beats API pricing — a quick gut-check before buying hardware to self-host.