Issue #36 · Monday, September 14, 2026 · 7 stories

🎵 Suno Says Licensed Music, Its Own v6 Announcement Only Says Developed With

ElevenLabs published its caps, YuE2 published its VRAM, and the 11x speed-up wants eight GPUs.

Suno's v6 announcement runs 3,469 words and never once says trained on. It says developed with, which is not pedantry but the entire legal question, and the same announcement leaves Universal, Sony and GEMA unsettled while its artist payment mechanism is still unbuilt. The pattern repeats all week: ElevenLabs published its download caps, an open music model publishes the VRAM it needs alongside the licence that stops you selling the output, and NVIDIA's 4.5 times faster depends on a module that has not shipped. Even the solo-business teardown at the end plays fair and flags its own revenue figure as self-reported. The claim is always in the headline, the condition is one document down, and this week it is worth the click.

NEWS

Suno v6 Claims Licensed Music, But Never Says 'Trained On' · 18 min read

Suno's v6 announcement runs 3,469 words but the word 'trained' appears zero times, replaced by 'developed with.' Only Warner's press release commits to 'licensed' in writing, and that bundles a settlement, asset sale, and a future opt-in payment mechanism. The artist-payment feature hasn't shipped, the deal doesn't cover Universal, Sony, GEMA, or the Isbell/Lowery lawsuit, and Suno acknowledged scraping YouTube audio the day before launch. The gap between 'developed with' and 'trained on' is exactly where licensing ambiguity lives.

ElevenMusic v2.5 Is the New Default, and the Real Limit Is Your Download Cap · 3 min read

ElevenLabs shipped Music v2.5 on 11 September and made it the default for prompted and reference generation in ElevenMusic, with the model also live in Studio, Flows and the API. The pitch is fuller arrangements and more natural instruments, widest on vocal-led and acoustic genres, backed by a blind test across 47,885 prompt pairs. For anyone scoring video, the constraint is not the audio, it is the download cap: Free gets five lossless downloads a day, Pro gets 400 a month, and downloads are blocked outright on any track built on another artist's song. The useful part is the licensing: what you make is yours on every plan including Free, and the permissions attach to the track, so downgrading or cancelling does not retroactively strip rights to work you already made.

RTX Spark's Half-Frame Trick Cuts Video Render Time 4.5x · 7 min read

NVIDIA's RTX Spark pipeline renders a 2-second 1080p clip in about 1:25 by generating at 30 FPS in NVFP4 precision, then using offline AI frame generation and upscaling to hit 60 FPS. That's 4.5x faster than full-frame generation, but the frame-gen module isn't out yet, and fast motion or fine textures can still produce artifacts. The efficiency gain may also become deadline pressure rather than creative headroom, a concern for animators and short-form creators.

ANALYSIS

Solo AI Headshot Business Hits $300k/Month by Staying Boring · 3 min read

Danny Postma's HeadshotPro reports roughly $300k/month in revenue (unverified) by selling one-time packs ($29-$59) instead of subscriptions, and building an affiliate program that generates $50k+/month. The case study shows how narrowing the promise to professional LinkedIn headshots and pricing to buying behavior drove growth.

TOOLS

An Open Music Model Says It Matches Suno, and Hands You the Score · 9 min read

YuE2-3B is an open music model from the Multimodal Art Projection consortium that claims song quality competitive with Suno v5 and v6 and takes the top open-model average on WildSongBench. It builds a full song with vocals from lyrics and a style prompt, then lets you shape melody and chords through an editable score, which is precisely the control a closed platform will not give you. Resource numbers are published rather than implied: a 3.6-minute song in 71 seconds on an RTX 4090, roughly 11GB peak VRAM, one song at a time on a 24GB card. The catch is the licence. It ships CC BY-NC 4.0, so treat it as a prototyping and learning tool, not something you put under paid client work.

A Three-Step LoRA Makes MiniMax H3 Stream, If You Have Eight GPUs · 6 min read

Alibaba's TaoLive AIGC team open-sourced TaoMate-H3, a three-step LoRA that turns MiniMax H3 into a streaming audio-and-video runtime, claiming 11.45 times faster pure-DiT time than stock H3 with synchronised audio at 480p, 768p and 1080p. Portrait resolutions are native, 768x1376 and 1088x1920 included, so you are not cropping down from landscape. Read the run command before you clear your evening: the reference example spans eight GPUs across eight devices, and the released adapter is a step-3000 EMA checkpoint at rank 128. On one consumer card this is a read-the-code release, not a drop-in speed-up. Community ComfyUI conversions are appearing, which is where it becomes usable for the rest of us.

Reference Images Replace LoRA for Consistent Anime Characters · 9 min read

PixAI's Tsubaki 3 model lets creators upload a reference image and reuse that character across poses, outfits, and scenes without training a custom LoRA. The tutorial covers four workflows: outfit transfer, multi-character replacement with @-tagged references, pose transfer from sketches, and converting 2D designs to figure-style renders. The key is tagging reference images inline in the prompt.