Netflix's Eyeline Labs open-sourced DiffHDR, an Apache 2.0 LoRA that converts clipped 8-bit video into real HDR radiance you can re-expose. Free weights, free base model, one 75 GB catch.
Alibaba open-sourced a 7B image model with native RGBA transparency, and the community had it in 4.19 GB within a day. Free to run, but non-commercial only.
Tencent AuK is a free MIT-licensed 1.5B AI voice model for cloning, TTS and speech editing. Free demo, weights, VRAM numbers — and the license catch.
Doop v0.6.0 is a free, open-source AI design canvas. Self-host in one command, connect Claude Code over MCP, and watch agents design live beside you.
VC-Attention claims a 3.58x AI video speedup but ships no kernel. SageAttention in ComfyUI is free, Apache-2.0 and installs in one pip line. Here is the setup.
MuLa Labs released MuLaCover's weights and code on September 16. It turns any vocal-forward reference track into a new cover from your lyrics — free, ungated, and with one serious licensing catch.
Marigold V2 is free and Apache-2.0: depth, normals and albedo from one image, now native in ComfyUI. Specs, VRAM, exact commands and the real catch.
Gradium Voice Design turns a 500-character prompt into five brand-new AI voices in seconds, free on every plan — plus the open-weights local fallback.
Viggle-Animate swaps a video character from one repainted frame in 26 seconds. Free weights, free demo, and a license that shuts out the US, UK and EU.
Give your AI campaign brief a fulfillment owner, a realistic response expectation, and a recovery plan before the first buyer answers your call to action.