Local AI video stopped being a hobby project this week. MiniMax H3 does 2K with native stereo audio on a 3060, under open weights, with ComfyUI support on day one, and Kroma cuts Krea 2 to eight steps on the same machine. The awkward part is that the same seven days handed you three new reasons to think about where that output goes: an EU labeling mandate that is now in force, a Snapchat payout cut, and a German court telling Suno that fair use does not cover it.
NEWS
The EU AI Act requires companies to label AI-generated content that appears authentic, effective August 2, 2026. New AI systems must comply immediately; existing systems get a four-month grace period. Non-compliance fines reach 3% of total gross revenue. Exemptions exist for personal content and clearly artistic or satirical works. Creators publishing AI-generated video, images, or audio in the EU must ensure proper labeling or face penalties.
Snapchat updated its Spotlight program to exclude fully AI-generated videos from rewards. The platform adjusted its recommendation system to prioritize human-made content, though AI editing tools remain allowed. This follows similar moves by LinkedIn, Substack, Meta, and YouTube, signaling a broader industry shift away from low-quality AI slop. For creators on Snapchat, earning distribution and rewards now requires original human storytelling.
A German court found Suno's AI music models copied and reproduced copyrighted songs like 'Rasputin' and 'Mambo No. 5' in generated outputs. The ruling grants GEMA injunctive relief and damages, rejecting Suno's fair use defense. Suno plans to appeal, but the decision sets a European precedent that AI music services need explicit licenses, which could mean service restrictions or licensing changes for users.
ByteDance released Dreamina Seedance 2.5 globally. It generates native 30-second video in a single pass, with a beta mode for 3-minute clips. Creators can edit specific frames without regenerating the whole clip. Plugins for Maya and Blender let teams integrate video generation directly into 3D production pipelines for previs and pitch work. The model accepts up to 50 multimodal references and structural R2V guidance. Pricing is unpublished, the long video mode is beta, and availability is regionally staged. The native shot length and localized editing directly reduce the assembly and retry overhead that has blocked professional adoption.
MiniMax released its third-gen video model H3 with open weights and native stereo audio, and ComfyUI added support the same day. The model accepts text, images, video, or audio inputs and generates up to 15 seconds of 2K video with synchronized stereo sound in a single pass. It can run locally on a 3060 GPU.
xAI added multi-image reference (up to 7 images), voice consistency, text-to-video, and native 1080p output to Imagine Video 1.5. The update is rolling out to Grok subscribers, starting with SuperGrok Heavy and Plus in the US. Developers can access image references, text-to-video, and 1080p via the API, while voice references require a request.
TECHNICAL
A new LoRA fine-tune called Kroma lets Krea 2 users generate images in 8 steps using the Turbo checkpoint, packaged as a single safetensors file for ComfyUI. It includes full-weight deltas for normalization and modulation tensors, applied automatically by the standard LoraLoader. The creator recommends Turbo with 8 steps, CFG 0-1, and shift 1.15 for best results.
TOOLS
Fish Audio has made the S2.1 Pro API free under a fair use policy with no hard character cap, and pushed the window from July 24 out to August 31. One model covers 83 languages at roughly 90ms time to first audio (about 70ms on a single request), with voice cloning and word-level control over emotion, intonation and pacing, and the company put it in front of a $52M seed round announced on July 28. Read the terms before you wire it into anything that ships: no SLA, no latency guarantee, requests may be used to improve the model, and products over $1M in annual revenue are asked to make contact first. The comparison Fish Audio picks is ElevenLabs' free tier at roughly 6 to 10 minutes a month, which is the honest read on why this is worth prototyping with. Free here means free to test, not free to depend on.