Plus local VibeVoice voiceovers at 4x real-time, and INT8 Qwen encoders that run on 8GB VRAM.
The Motion Picture Association sent ByteDance a cease-and-desist calling Seedance systemic infringement, after a viral clip faked a Brad Pitt and Tom Cruise fight. Yet studios will not approve the tool while quietly tolerating it on a don't ask, don't tell basis. The risk is not performance but contested legal ground under a tool everyone uses anyway, so keep a fallback and expect the likeness rules to tighten.
Judge Alvin Hellerstein in New York denied Sony Music's bid to add roughly 30,000 recordings to its training-data suit against Udio, keeping the case at the original 333 works. That caps statutory damages near $49.95 million instead of a multi-billion figure. Sony is now the last major without an Udio license after Universal, Warner, and Merlin settled. If you sell Suno or Udio tracks, training-data law is still unsettled, not solved.
Aronofsky's studio Primordial Soup put out 'On This Day 1776,' a short AI history series made with Google Veo 3 for TIME's YouTube, to reviews branding it AI slop. The team paid SAG-AFTRA voice actors and a composer, so the human layer is real, but the Veo 3 visuals draw the unhinged reactions. The lesson: for narrative AI video, model quality is no longer the gate, continuity across a full scene is.
audio.cpp is a pure C++/ggml runtime with no Python dependency that now runs Microsoft's VibeVoice 1.5B for local text-to-speech. The number that matters: a 90-minute podcast renders in about 23 minutes, roughly 4x real-time and 2.8x faster than the Python reference. For anyone paying per minute for cloud voiceover, that moves the cost onto hardware you own. The catch is an early, single-maintainer project, so test before trusting it in production.
Community INT8 quantizations of the 2B, 4B, and 9B Qwen3.5 text encoders, as drop-in replacements for the BF16 versions in ComfyUI. The method is ConvRot, a Hadamard rotation that suppresses the weight outliers that wreck naive INT8 image quality, loaded with the stock CLIPLoader node. The payoff is large-encoder prompt fidelity on a consumer 8GB card instead of rented VRAM. The limit: you are trusting a community quant, so A/B against BF16 first.
The open-weights Krea 2 Turbo, an 8-step distilled text-to-image model, can often generate directly at 4k, which drops the separate upscale step some workflows bolt on. Community tests call it inconsistent though, hence 'often' rather than always, and native 4k means a heavier VRAM bill per image. The read: try 4k direct before wiring in an upscaler, but keep the upscale fallback for prompts where it smears.