And SAM 3 tracks objects for pennies, while a Krea 2 LoRA finally locks a face across your edits.
Spotify removed over 75 million AI-generated tracks in the past year, targeting low-quality spam, fake uploads, and royalty manipulation. New rules ban unauthorized AI voice clones and improve detection. For creators using AI music tools, this signals tighter enforcement but continued allowance of legitimate AI-assisted music with proper rights and transparency.
TikTok is testing an opt-in tool that lets creators identify and flag AI generated content using their face without consent. Creators verify their identity via Jumio, then TikTok scans for potential matches and allows reporting of impersonating accounts. The test is limited to a small group of US creators. This gives creators a platform level protection against unauthorized deepfakes, but the verification requirement and limited rollout mean it is not yet a universal solution.
The AI image generation field has split into four distinct lanes. GPT Image 2 leads on reasoning and text-heavy layouts. Nano Banana Pro wins on realism and style matching. Seedream 5.0 Pro offers layer separation and sketch editing. Ideogram 4 provides open weight typography and design. No single model dominates all tasks. Choose based on what you are building, not the brand.
Meta released SAM 3 on fal.ai, a unified video segmentation model that tracks objects across frames using text, points, boxes, or masks. At $0.005 per 16 frames (roughly $0.31 per 1,000 frames), it undercuts frame-by-frame image segmentation. Features include adjustable detection thresholds, per-frame bounding box export, and optional mask application for content moderation, sports highlights, and medical video analysis.
An unofficial Krea 2 Raw fine-tune for instruction-based, identity-preserving image editing: re-stage a person in a new scene while holding the face, plus head and face swaps, object add and remove, and two-input scene-plus-person compositions. The v1.2 weights ship as safetensors with reduced-rank r128 and r64 variants for lower-VRAM cards. It is trained on RAW and meant to run on the distilled 8-step Turbo for inference. 425 likes and five Spaces already use it, so the adoption is real, not a demo.