Creative AI Briefing September 12, 2026 morning edition 13 min read

Three releases this window published the map of where they fail alongside the map of where they work, so a creator can now decide whether a model suits the job before spending an hour and six gigabytes finding out.

AuK AuK-Flash Evoke-Turbo YuE2-3B MiniMax-Music3 ComfyUI voice-clone audio-gen video-gen local-creative-ai open-weights music-gen

Creative AI Briefing: Saturday, September 12, 2026

Feed a noisy field recording into ComfyUI, type "remove the noise and reverberation while retaining the speech," and get clean dialogue back from a model small enough to sit in 880 megabytes on disk. Then read the list the person who packaged that model wrote, which tells you plainly that the same interface will probably fail if you ask it to swap a word in a lyric. Three releases in the last forty-eight hours did that: they shipped the map of where the model breaks next to the map of where it works. Tencent's AuK reached ComfyUI and Apple Silicon with a published reliability ranking of eighteen tasks. Alaya's Evoke-Turbo claims a state-of-the-art score and then explains, on the same card, that its own row was chosen from six prompt variants at one seed against numbers somebody else measured. A community packager quantized YuE2 to nineteen profiles and printed runtime tables while writing "no fidelity table is included." None of these is modest. All of them save you an evening.

New models

Evoke-Turbo is a three-step world model you can steer mid-shot, and its own benchmark table is an argument against itself. Alaya Lab shipped the weights September 11 at 13:32 UTC under Apache-2.0: 14.3 billion parameters stored in FP32, about 57 GB of safetensors, generating 384x640 at 24 fps in three steps with no classifier-free guidance. One chunk is 36 frames, so NUM_CHUNKS=20 gives a thirty-second clip, and the repo quotes 1.5 seconds of video every 2.11 seconds on a single H200. The reason it stays coherent past a normal context window is that scene geometry does not live in the denoiser at all. A monocular depth model unprojects each emitted chunk into a persistent point cloud, the current camera pose reads that bank back as a warped image plus a visibility mask, and only what the present view needs gets retrieved. The context stays bounded however long you run.

Now the honest part, which the Evoke-Turbo card states rather than buries. Turbo posts a WBench average of 82.0003 against base Evoke's 80.8210, but it wins Navigation (83.8978 to 78.6300), Consistency and Physical while losing Video Quality (81.8914 to 82.7900) and Setting (82.0518 to 83.7600). That is a controllability-for-fidelity trade, and Alaya labelled the release "controllability-optimized" instead of pretending otherwise. Then the footnote: "Turbo uses seed-44 results selected from six prompt variants; Evoke uses published reference scores, with different prompts and poses." The two rows were not produced the same way. The GitHub README still leads with "state of the art on WBench" and does not carry that sentence.

Two more things before you clone it. The required depth backend, pkqbajng/ViGeo, ships CC-BY-NC-4.0, stricter than the Apache-2.0 on the weights, and every shipped recipe uses it, so the documented pipeline is not commercial. And the distilled models were trained on video-to-video conditioning alone, so MODE=i2v or MODE=t2v on Turbo is zero-shot by the repo's own admission. The community Space at hugging-apps/evoke is asleep and points at a different repository id, so there is no working hosted demo.

Video

The Director Web UI that landed with Evoke on August 30 is the part a filmmaker cares about and most coverage skipped it: a local single-GPU app with per-chunk prompts, independent movement and look camera sticks, streaming chunk previews and non-destructive revision branching. Run python ui/app.py and open port 7860. It binds to localhost with no authentication, which the repo says out loud.

Elsewhere in video the week is compression and porting rather than capability. Perflow-Shuai published four LongLive DMD LoRAs for Wan2.2-TI2V-5B on September 10 (rank 128, four-step, non-autoregressive, Apache-2.0, with several iteration checkpoints so you can pick where the distillation stopped helping). Yi30 posted a W4A8 SVDQuant conversion of Wan2.2-TI2V the same day. EliovpAI followed its MiniMax H3 Radeon work with a FastWan-5B build compiled for RDNA4. And drawthingsai/MiniMax-H3 appeared September 11, which puts H3 in the Draw Things Mac app for people who will never open a terminal.

Audio and music

AuK is the release of the window, and it arrived properly this week rather than last. Tencent Hunyuan's AuK is a 1.5B speech foundation model under a plain MIT license that exposes sixteen tasks through one natural-language instruction field: zero-shot and instruction TTS, content editing, lyric editing, pitch, speed and volume, emotion, timbre, de-accenting, nonverbal sounds, whisper conversion, enhancement, speaker separation, music separation and target-speaker extraction. The HF repositories were created August 18 and August 21, and the card's own News line dates the open-source announcement to September 9. What changed in the last two days is that it stopped being a Gradio demo. A ComfyUI node pack with eight nodes and three example workflows appeared September 9 at 20:58 UTC with its converted weights following at 00:37 the next morning, and a native MLX port for Apple Silicon landed September 12, Flash at 01:53 UTC and Base two hours later at 03:50.

The quantized ladder is why this matters on a small card. The ComfyUI repack publishes the diffusion transformer at 6.122 GB in FP32 down to 0.880 GB at W4A8, with the Qwen2.5-Omni encoder at 8.070 / 4.669 / 3.179 GB and the 0.637 GB VAE deliberately left unquantized and computing in FP32. One reported peak with a quantized model and encoder pair was around 8 GB, described as an observation rather than a floor. Base runs 32 Euler steps with guidance 2, Flash four fixed steps with guidance disabled, output mono 24 kHz. On a Mac the MLX port reports the four-step pass on ten seconds of audio dropping from 3.820 s under PyTorch MPS to 0.992 s, with the 8-bit build cutting backbone memory from 5.70 GB to 0.56 GB. Those are the author's own numbers on unstated hardware in a repository with no stars, so treat them as a claim to verify rather than a measurement to plan around.

The unusual thing is the prompt adherence ranking the integrator published before anyone asked. Eight tasks are called stable: description TTS, voice cloning, volume, enhancement, singing extraction, speaker separation, denoise, quality repair. Ten are called less stable, and the bottom line bundles insert, replace, lyric editing and whisper conversion together, which is to say the marquee capabilities in the announcement. Two workarounds come with it, both transferable: setting the target duration 0.2 to 0.5 seconds shorter than the source improves difficult edits (5.5 seconds for a 6-second input), and some unstable tasks succeed more often with the Chinese cookbook instruction than the English one. Duration conditioning and instruction language both move the result. Nobody else is writing that down.

Music got two useful repacks. ngquocvinh/YuE2-3B-GGUF landed September 11 at 02:59 UTC with nineteen main profiles for the audio.cpp yue2 loader, SHA256 sums, a reproducibility manifest, and a measured runtime table on one A10M: BF16 plus F32 VAE took 144.9 s of wall time for 224.8 s of audio at 11,095 MiB peak, Q8_0 plus F16 took 112.2 s at 8,033 MiB, and Q4_0 plus F16 took 96.3 s at 6,441 MiB. The Q2_K run hit the semantic-token limit and the table says so in a column. The card also states that no imatrix is claimed because audio.cpp's converter will not take one, and that the smoke tests are not a fidelity evaluation. Weights stay CC BY-NC 4.0, so this is a practice tool. Separately, TeamAudiyo posted MiniMax-Music3 GGUFs September 11 down to 1.49 GB at Q4_K_M for 8-10 GB cards, though the card's own download commands fetch from Abiray/MiniMax-Music3-GGUF rather than itself, which tells you where it came from.

On the business side, Universal Music Group and ElevenLabs announced a multi-year strategic agreement on September 10. No creator-facing product shipped with it, and licensed catalogue is what eventually decides which voice and music tools you can publish with.

Image

Quiet, and worth saying so rather than padding. The one release with teeth is TransNormal-2 from September 9, a FLUX.2-klein-base-9B LoRA that estimates surface normals for transparent objects, which is the case every photogrammetry and relighting pipeline fails on. It is research-licensed, not Apache. Everything else in the last two days on the image side is mirrors, personal LoRAs and repackaged Qwen-Image-Edit adapters.

Open and local

The pattern in the repos is porting rather than inventing. Speech and music models that shipped in the last two weeks are being cut into quantized ladders and wired into the two surfaces creators actually use, ComfyUI and Apple Silicon, usually by a single person over a weekend.

Tencent-Hunyuan/AuK: one instruction box that does TTS, voice cloning, denoising, stem separation and speech editing, MIT licensed. 515 stars, and the free hosted Space is running on zero-a10g right now (repo).
Saganaki22/ComfyUI-AuK: eight nodes, three drag-and-drop workflows, and the reliability ranking that should ship with every model. 25 stars, created September 9 at 20:58 UTC (repo).
vanch007/mlx-AuK: native Apple Silicon port with an A/B web demo, Flash and Base weights posted September 12. Zero stars so far, numbers unverified (repo).
AlayaLab/Evoke: three-step world model plus the Director UI for per-chunk prompting and dual-stick camera. 466 stars (repo).
ngquocvinh/YuE2-3B-GGUF: full song generation on a 6.4 GB peak instead of an 11 GB one, with the measurement table to prove it (card).
TeamAudiyo/MM3-GGUF: MiniMax-Music3 quantized to 1.49 GB for 8-10 GB cards, Apache-2.0 inherited (card).
onnx-community/VibeVoice-Realtime-0.5B-Onnx: streaming long-form TTS in ONNX Runtime, MIT, posted September 11 (card).

Creative workflows

1. Clean up a bad recording locally with AuK in ComfyUI, and know in advance which fixes will hold. Install ComfyUI-AuK into custom_nodes/, install its requirements against your ComfyUI Python, then pull three files from drbaph/AuK-comfyui: one diffusion model (auk_base_bf16.safetensors at 3.062 GB, or auk_base_w4a8.safetensors at 0.880 GB if you are tight), one encoder (qwen_omni_int8.safetensors at 4.669 GB), and auk_vae.safetensors. Drop 02_audio_edit_or_clone.json from example_workflows/ onto the canvas.

The steps. Load your audio, type the instruction, set seconds to 0 to match the source, render. How it works. The Qwen2.5-Omni encoder listens to the clip directly, so you never supply a transcript; the instruction and the audio become one conditioning tensor that the diffusion transformer denoises against. Why it is good. Denoise, dereverberate, extract a singer from a mix and separate one speaker from a crowd all sit at the top of the tested-stable list, and none of them needs a subscription or an upload. Where it breaks. Anything that changes words or accents sits near the bottom. If you attempt those, set the output 0.2 to 0.5 seconds shorter than the source and try the Chinese phrasing from the cookbook before you conclude the model cannot do it. Also note the pack's warning, in its own words, that "Native VAE Decode Audio in the tested ComfyUI normalizes loud outputs, which can alter volume edits," so use the generation node's direct decode when amplitude is the point.

2. Change a scene mid-shot instead of cutting to a new one. Evoke's segment mode conditions each chunk separately, so the prompt can change while the rollout is running. From a cloned Evoke repo with weights and ViGeo in place, run MODE=segment NUM_CHUNKS=6 MAX_CASES=0 bash scripts/inference/infer_post_distill.sh.

The steps. Six chunks at 36 frames each gives 213 frames, roughly 8.9 seconds, with the prompt switching at chunk 3. The bundled cases include an aurora igniting over a frozen tundra and a meteor shower opening over a still lake. How it works. The world state bank keeps the geometry you already established, so when the prompt changes the model repaints the same place instead of inventing a new one. Why it is good. Weather, light and event changes normally cost a cut. This is the same take. Where it breaks. You need the environment the repo actually uses (Python 3.10, CUDA 12.4, torch 2.4, and a diffusers development fork that is not on PyPI, which pip will not install for you), plus ffmpeg on PATH. The warp and attention recipe must match between training and inference; a mismatch degrades quality silently rather than erroring.

3. Generate a full song locally on a mid-range card. Build the dev branch of audio.cpp, download the YuE2-3B GGUF directory including sidecars/, and point --model at the directory, not at a single file. The steps. audiocpp_cli --task gen --family yue2 --model <dir> --backend cuda --threads 8 --lyrics $'[Verse]\n...' --request-option 'style=English, indie pop, bright acoustic guitar' --out yue2.wav --log --metrics. Q8_0 plus the F16 VAE load by default; override with --session-option yue2.model_gguf=yue2-3b-q4_0.gguf. How it works. Lyrics and style are separate conditioning inputs, and cot=melody or cot=full turns on the symbolic planning stage that YuE2 uses to write a score before it writes audio. Why it is good. Q4_0 produced 220 seconds of 48 kHz stereo in 96.3 seconds at 6,441 MiB peak on the packager's A10M. Where it breaks. The weights are CC BY-NC 4.0, so nothing you make here is commercially usable, the Q2_K profile truncated on the longform check, and the package publishes no fidelity comparison between quants at all.

Worth testing

tencent/AuK on Hugging Face Spaces, running now on zero-a10g with sample inputs for accents, content edits, whisper conversion and vocal extraction. Free, no install. Tradeoff: the queue is shared and the Space carries the same reliability gaps as the local build, so a failed edit here is the model, not the hardware.
AuK-Flash over AuK base if you are iterating. Four fixed steps, no guidance, same task list. Tradeoff: the pack says Base at 32 steps is the quality choice, and nobody has published a side-by-side on edit accuracy between the two.
The W4A8 AuK pair at 0.880 GB plus 3.179 GB. Tradeoff: it needs a ComfyUI and Comfy Kitchen build exposing asym_w4a8_int8, which the pack explicitly says you should not assume every release has.
Evoke's Director UI rather than the shell scripts, if you want to feel the camera control. Tradeoff: single GPU, no auth, and you still need the full CUDA 12.4 environment before it opens.
MiniMax-Music3 at Q4_K_M (1.49 GB) with the Comfy-Org text encoder and VAE. Tradeoff: the card calls Q3_K_M "may introduce audible softening" and offers no measurement to say where Q4 sits relative to it.

What actually matters from today's signal

The shift this window is not a capability. It is that the people shipping ports have started publishing the shape of the failure alongside the shape of the win, in the same document. A ComfyUI integrator ranked eighteen tasks by whether they actually work and put the warning above the download links. A quantizer printed peak memory, wall time and a truncation column, then wrote that these are smoke tests and not a fidelity evaluation. A lab published a benchmark row and immediately explained it was selected from six prompt variants at one seed against numbers measured under different conditions. Every one of those disclosures makes the release look worse and makes it far more useful.

That changes the economics of trying things. The expensive part of adopting an open model has never been the download; it is the two hours you spend discovering that the one task you needed is the one that does not hold. AuK's ranking collapses that to a thirty-second read. The YuE2 table tells you before you build audio.cpp whether your card will finish a song. Evoke's footnote tells you the navigation gain is real and the quality gain is not.

The counter-signal is that this is voluntary and inconsistent. Evoke's model card carries the caveat and the GitHub README, which is what people actually read, does not. The MLX numbers for AuK come from an author with zero stars and no stated hardware. TeamAudiyo's music GGUF card still instructs you to download somebody else's files. Honest documentation is a habit of individual maintainers right now, not a norm, and the moment a release gets big enough for a marketing team the footnote is the first thing to go. Read the card, then read the README, and when they disagree, believe the one that makes the model look worse.

Source access notes: ComfyUI's blog (blog.comfy.org) is a Substack JavaScript wall and returned no content; ComfyUI GitHub releases via the API returned empty. Hugging Face's daily_papers endpoint exceeded the fetch size limit and was skipped. Midjourney, Civitai and Suno were not reachable this run. The text-to-image createdAt feed returned results ending July 27 without cache busting and was re-fetched with a cb parameter; the stale response is why the image section was verified separately. al3al3/supertonic-3 surfaced as new on the September 12 feed and was dropped after the API showed the real Supertone/supertonic-3 was created 2026-05-06. Every Hugging Face date here comes from the API createdAt field, cache-busted. GitHub star counts come from cache-busted shields.io JSON. An adversarial fact-check subagent ran against this draft before publication and returned three corrections, all applied: the ComfyUI-AuK code repository was created September 9 at 20:58 UTC, not September 10 (that is the weights repo); the pack's VAE warning says a normalizing decode "can alter volume edits," not that it silently undoes them, and is now quoted verbatim; and the MLX port is two repositories posted two hours apart, not one. Every other load-bearing figure, date, license and quote in this briefing was checked against primary sources and matched, including the full WBench table to four decimals and the Evoke-Turbo seed-44 footnote's absence from the GitHub README.