Creative AI Briefing September 10, 2026 morning edition 13 min read

The releases worth having this window hand back an editable intermediate instead of a finished render, so revising a generated result stops meaning rolling the dice again.

YuE2 SheetSage2 ID-V2V Adobe Premiere Krea 2 Turbo Wan2.2 music-gen audio-gen video-gen ai-editing creative-workflows open-weights

Creative AI Briefing: Thursday, September 10, 2026

Feed a three-minute recording to a transcriber, open the sheet music it returns as a text file, change four chords, and get back a jazz-funk cover of the same melody sung by a voice that has never existed. That pipeline shipped last night at 18:30 UTC, in six repositories posted eleven minutes apart. It is the clearest case of a pattern worth naming: YuE2 hands you a score, and Adobe's new Premiere tool drops generated clips onto your timeline as editable footage rather than as renders you accept or discard. Neither gives you a finished file and a re-roll button. Both give you a middle document you can open, change, and send back, which changes what "revise" costs.

New models

YuE2-3B is an open music model that claims to beat Suno v5, and it publishes the scoreboard. The Multimodal Art Projection group shipped it September 9 at 18:30 UTC alongside five companion repos: two VAEs, two MERT-v2 audio encoders, and SheetSage2, a transcriber. You give it a style prompt and lyrics, and it returns a complete song with vocals and accompaniment at 48 kHz stereo. On the group's own WildSongBench, 192 prompts, YuE2 best-of-8 posts a SongBench average of 6.9632 against Suno v5's 6.8721 and Suno v5.5's 6.7150. Read that as the authors' own evaluation of their own model against competitors they ran themselves, because that is what it is. The standard YuE2 number, 6.7316, sits below Suno v5, and the table shows it plainly rather than hiding it. Note what "standard" means here: the card says standard YuE2 already selects the lower-PER candidate from two. Neither published row is a single shot.

The speed table is the part that decides whether you use this. A 3.6-minute song takes 71.04 seconds on an RTX 4090, at 11.18 GiB peak VRAM. The card asks for a 24 GB GPU and 24 GB of available host RAM, no quantization, BF16 for the AR and NAR stages and FP32 for the VAE. Install is a wheel pulled straight from the Hub. That is a consumer card generating a full song faster than you can listen to a third of it.

Now the part launch coverage will skip. cot="full" makes the model write a melody-and-chord plan before it writes any audio, and pipe.plan() lets you stop there. song.save_artifacts() writes out score.abc, the tokens, the latents, the audio and the settings. Copy score.abc, reharmonize it in any text editor or hand it to an agent, pass it back as the abc= argument with the same seed, and the model renders your revision. The demo site walks one song, "The Last Train," through 9 steps and 14 versions from Mandarin pop to English jazz with a saxophone solo built on two statements of "Twinkle, Twinkle, Little Star."

The catch is the license, and it is not small. Weights are CC BY-NC 4.0, and so are SheetSage2, both MERT encoders and both VAEs. Every piece of this pipeline is non-commercial. You can learn from it, demo it, publish about it, and use it for anything you are not paid for. You cannot put a YuE2 track under a client's ad. Suno's output, whatever you think of the quality gap, comes with commercial terms. That is the trade this release actually offers (weights, demo site, code, 6.5k stars).

Image

A distillation writeup with no files in it is the most useful image-model document this week. lvladikov/Krea2-Turbo-Distill-2step-LoRA went up September 9 at 20:25 UTC and says at the top that nothing has shipped. What it publishes instead is why the obvious approach failed. The author spent 9,000 samples doing progressive distillation from a working 4-step adapter down to 2 steps, watched every metric stop moving, and got pictures with doubled contours on faces and crowds averaged into translucent overlaps. The stated reason is worth quoting because it generalizes past this model: a regression loss asks the student to land on the teacher's specific image, and when a two-step jump leaves several plausible images, the answer minimizing squared error is their average, and the average of two sharp images is a blurred one.

Switching to distribution matching, the DMD2 family, moved further in 1,000 samples than 20,000 samples of the old recipe had, and it did so while every latent distance to the teacher rose. The card reports those distances and says outright they are recorded but never used to keep or stop a run. Anyone who has ever picked a checkpoint by a number should read the "How it is judged" section. Training runs on one RTX 3090 at about 160 samples an hour against the 4-step recipe's 470. Licensed under the Krea 2 Community License (writeup).

Smaller: the WebNN organization published its own webnn/Z-Image-Turbo ONNX conversion September 10 at 02:50 UTC, one day after the community lwanming/Z-Image-Turbo-opt covered here yesterday, both Apache 2.0 and both targeting the browser through ONNX Runtime Web. Microsoft's live demo still runs with nothing installed. And ldov/ideogram-4-gguf (September 8) quantizes Ideogram 4 into eight GGUF grades for ComfyUI and stable-diffusion.cpp, splitting conditional and unconditional diffusion so you can spend bits where they matter and pairing them with a Qwen3-VL-8B text encoder. The base weights carry Ideogram's non-commercial model agreement, and the quantized repo mixes in unrelated adult LoRAs, so treat it as a source for the diffusion files and not as a curated collection.

Video

A mirror is the only new video weights on the Hub this morning, and it is worth saying so. gubernac/ID-V2V appeared September 10 at 07:08 UTC and looks like a launch. It is not. It is a personal reupload of Eyeline-Labs/ID-V2V, whose own createdAt is July 15 and whose GitHub repo was created July 14 and last pushed July 29. The underlying work, a Netflix and Eyeline Labs paper accepted to SIGGRAPH Asia 2026, is roughly eight weeks old. Go to the official repo, which has the Space and the likes, not the copy. The work itself is excellent and gets the full treatment in the workflows section below, framed for what it is: something already on the shelf that most people missed (official weights, code, 249 stars).

Adobe put generation inside the Premiere timeline on September 8, ahead of IBC. The new Generative Media tool creates video and sound effects at a spot you select, as fully editable clips, with a choice of Firefly, Google Veo, Kling, Runway or Luma. Generate Video can sample reference frames from your own project so the new footage matches the surrounding edit. Generate Soundscape, in beta, reads up to 15 seconds of video and builds layered ambience timed to the cut. Generate Music, also beta, ships with tempo and looping controls and commercially safe synchronization rights. Separately, Enhance Audio splits dialogue, music, ambience and effects inside one clip for independent control, and Separate Crosstalk pulls two overlapping speakers onto their own tracks. AI Assistant is now in public beta in After Effects, working across a whole project rather than one composition (announcement).

Open and local

The second AMD package in two days landed, and its numbers are smaller than yesterday's on purpose. EliovpAI/Wan2.2-FastWan-5B-Paiton-RDNA4 (September 9, 21:38 UTC, Apache 2.0) targets one Radeon AI PRO R9700 on gfx1201 and reports a measured Paiton latency reduction of 2.8 to 4.7 percent on FastWan, while stating that base image-to-video cases did not improve end to end at all. It also says the distillation is identical between engines and is not a compiler gain. A vendor package that publishes a case where its own product did nothing is rarer than it should be.

multimodal-art-projection/YuE: the codebase behind the day's music release, now carrying YuE2's editable-score pipeline. Why now: 6.5k stars, and the model card points here for everything past the quick start (repo).
Eyeline-Labs/ID-V2V: restyle a shot without re-filming it, with five documented recipes and a runnable sample folder for each. Why now: a same-day mirror of the July weights pushed it back into the listings, and 249 stars on eight-week-old code says most people still have not run it (repo).
Soul-AILab/SoulX-FlashHead: real-time streaming talking heads on one RTX 4090, 1.1k stars. Why now: a mirror of the weights appeared on the Hub September 10 at 02:55 UTC, but the original work is from February 2026, so this is redistribution and not a launch (repo).
comfyanonymous/ComfyUI: still the assembly point every package above eventually targets, 132k stars (repo).
m-a-p/SheetSage2: a 57.2M-parameter transcriber that turns a recording into ABC notation and MIDI, and ships an abcjs renderer plus a piano soundfont so it can draw and play the result. CC BY-NC 4.0 (weights).

Creative workflows

1. Cover an existing song by editing its score, not its audio. (YuE2 quick start)

The steps. Transcribe the source recording with SheetSage2 and save the melody ABC without chord symbols as melody.abc. Get the lyrics separately, either transcribed with Qwen3-ASR or found and checked by hand, organized into sections that match the recording, saved as cover_lyrics.txt. Then call the pipeline with cot="melody", your two files, and a target style string such as "Jazz-funk, warm lead vocal, Rhodes piano, electric bass, tight drums."

How it works. Melody-only planning tells the model to honor the pitch contour you supplied and invent everything else. The published SHS100K numbers show what that buys: with a full score, CLEWS Hit@1 is 71.3 percent; without chords, 67.3 percent; without any score at all, 0.3 percent. Song identity lives almost entirely in the symbolic conditioning.

Why it is good. You are editing notation, which is diffable, versionable and reviewable by a musician who has never touched a model.

Where it breaks. cot="melody" does not strip chord symbols for you, so a score exported with harmony will silently constrain the arrangement you were trying to replace. The transcription tools want their own Python environments. And the same table shows the cost of control: dropping the score entirely gives the best musicality (5.691) and style match (0.474) and the worst identity, so a tight cover is a deliberate quality trade, not a free win. Everything here is CC BY-NC 4.0.

2. Restyle a shot you already filmed, using a keyframe you painted. Carried forward deliberately: this is July work (HF createdAt July 15, code pushed July 29) that today's mirror surfaced again, and it is the most complete worked example of the editable-intermediate idea anyone has published. (ID-V2V restylization recipe)

The steps. Build a sample directory holding source.mp4, stylized_first_frame.png and prompt.txt, optionally with keyframes/40.png and keyframes/80.png for mid or end anchors. Run SAMPLE_DIR=... SAM_PROMPT=person bash scripts/preprocess.sh, then bash scripts/examples/restylization.sh.

How it works. Preprocessing uses SAM3 to segment the person and gray out the background, writing preprocessing/orig_pixel.mp4. That foreground-on-gray video is the single VACE control condition. The model relights whatever is inside the mask and regenerates everything outside it from your prompt and keyframes. Sources longer than 81 frames generate in overlapping clips, each conditioned on the end of the last.

Why it is good. The output includes flip_test.mp4, which plays your result and freezes at up to five evenly spaced frames to alternate generated against source. A built-in A/B is a small thing that changes how fast you can tell whether it worked.

Where it breaks. The published hardware is 8× A100-80GB. Checkpoints run about 96 GB and SAM3 needs a gated-repo login. Pick the wrong one of the two .pth files and nothing errors, you just get worse video. And the authors call it a research exploration released "for demonstration and inspiration purposes only," which is a question your client will ask before you do, Apache 2.0 weights or not.

Worth testing

YuE2 in a browser, no install. A community Space, mrfakename/yue2-3b, went up September 9 at 22:39 UTC. Tradeoff: community-run rather than official, so uptime and configuration are nobody's promise, and the CC BY-NC 4.0 terms follow the audio out of the box (Space).
Premiere's Generate Soundscape beta on a cut you already finished. Feed it 15 seconds and see whether the ambience it invents is a usable sketch layer. Tradeoff: it is a beta positioned by Adobe as a starting point, and everything it makes still needs mixing.
The Ideogram 4 GGUF split. Run Q6_K conditional against Q4_K unconditional and see whether you can tell. Tradeoff: non-commercial base license, and the repo's extra contents are not curated.
ID-V2V in a browser first. The official repo lists a hosted Space, hugging-apps/id-v2v, which is the way to see whether identity preservation holds on your footage before committing to a 96 GB download. Tradeoff: eight-week-old research code, and the local path the Space is standing in for wants 8× A100 (Space).

What actually matters from today's signal

The interesting number in the YuE2 release is not 6.9632. It is 0.3 percent, the CLEWS Hit@1 for cover generation with no score supplied, against 71.3 percent with one. That gap is a measurement of how much creative control lives in a symbolic representation the model can be handed rather than a prompt it has to guess from. For two years the answer to "make it more like this" has been a longer prompt and another roll. A score file, a stylized keyframe, a clip on a timeline: each replaces the reroll with an edit, and the edit is cheap because you can see it before you spend the render.

The counter-signal is licensing, and it is severe this time. YuE2's entire stack, model, transcriber, both encoders and both VAEs, is CC BY-NC 4.0. ID-V2V calls itself a research exploration for demonstration and inspiration. Ideogram 4's open weights are non-commercial. The most editable tools in this briefing are the ones you cannot bill for. The only entry here you can put in front of a paying client without a lawyer is the one that costs a Creative Cloud subscription, and Adobe made a point of saying its generated music carries synchronization rights across any media type. That is not a coincidence, and it is worth watching whether it hardens into a rule.

The thing to be skeptical about is best-of-8. YuE2's headline win over Suno v5 comes from generating eight candidates and picking one by musicality, then prompt adherence, then phoneme error rate. Its standard two-candidate row loses to Suno v5. Eight generations is about nine and a half minutes of 4090 time, which is nothing, but the card's own protocol adds four ASR passes per candidate to score PER, so the real cost of that headline is generation plus a scoring pipeline the speed table does not price. The comparison is between a model wrapped in a selection procedure and a product that ships you what it made. Read the two YuE2 rows as two different tools, and budget for the second one.

Source access notes: Blocked or unusable this run: blog.comfy.org (JavaScript wall, no content served); the Hugging Face text-to-video listing endpoint returned a payload too large to read and was replaced by the image-to-video and text-to-audio feeds. Bucket A checks that produced nothing new in window: openai.com/news (latest creative item is ChatGPT Images 2.5, September 8, covered yesterday), elevenlabs.io/blog (latest post September 8 is an executive hire), blog.fal.ai (latest August 27), replicate.com/blog (latest August 4), runwayml.com/research (GWM Worlds 2, September 3, outside window). Every Hugging Face date above comes from the API createdAt field with cache busting, not from listing "Updated" timestamps; CurioSeaLab/SoulX-FlashHead-1_3B has a September 10 createdAt but is a mirror of February 2026 work and is labeled as such rather than dated as new. Star counts from cache-busted shields.io. WildSongBench and SHS100K figures are the YuE2 authors' own evaluations and are labeled as such throughout. Sections omitted: Writing, which had nothing in scope.

An adversarial fact-check pass ran against this draft and caught two errors, both now fixed. First and most serious: the draft led the Video section with gubernac/ID-V2V as a September 10 release. It is a personal reupload. The official Eyeline-Labs/ID-V2V has HF createdAt 2026-07-15 and a GitHub repo created 2026-07-14 and last pushed 2026-07-29, so the work is roughly eight weeks old. The item has been re-dated, moved out of the launch slot, relabeled as carried-forward in the workflows section, and repointed at the official repo. This is the same mirror trap the briefing caught for SoulX-FlashHead and initially missed on its own headline. Second: the draft called YuE2's 6.7316 row "single-candidate." The card states standard YuE2 selects the lower-PER candidate from two, so neither published row is a single shot; the text now says so. The check also noted that "nine minutes of 4090 time" priced generation only, while the card's best-of-8 protocol adds four ASR passes per candidate, and that the ID-V2V disclaimer is "demonstration and inspiration," not "demonstration only." Both corrected. All other claims in sections 3 through 10 of the check passed without amendment.