Creative AI Briefing September 13, 2026 morning edition 14 min read

No lab shipped a creative model this window, and yet the ceiling on what a 24 GB card can make went up roughly threefold, because the work that landed was conversion work on models you already downloaded.

LTX-2.5 Wan 2.2 TI2V 5B Z-Image Turbo YuE2-3B DreamX-Creator LingBot-World 2.0 Anima ComfyUI video-gen image-gen music-gen local-creative-ai open-weights licensing-provenance

Creative AI Briefing: Sunday, September 13, 2026

A ten-second LTX 2.5 video, 249 frames of it, now renders on a single 24 GB card with the model resident the whole time and nothing swapping to system RAM. That file went up at 03:16 UTC this morning, and it is not a new model. Every significant thing in the last twenty-four hours is a conversion of a checkpoint you could already have downloaded: LTX 2.5 at 3.36 times lighter, Wan 2.2 TI2V at 3.39, Z-Image Turbo at 3.75, the autoregressive half of YuE2 at 1.87. The labs published nothing. The converter layer published four times, and it moved the hardware floor further in one weekend than most model launches move it in a month.

Video

LTX 2.5's 22B distilled transformer now fits in 11.66 GiB, and the person who shrank it published the number that argues against his own file. JoaoZaokk/LTX-2.5-22B-distilled-W4A8-ConvRot went up September 13 at 03:16 UTC: 39.13 GiB of BF16 re-encoded into ComfyUI's own asym_w4a8_int8 format, 4-bit weights with an 8-bit activation path across 1440 Linear layers. On one RTX 3090 a complete 249-frame video at 512×512 and 25 fps took 400.9 seconds against the original's 780.7, so 1.95x faster. It also scores MAE 7.81 and PSNR 25.39 dB against BF16 where Lightricks' own INT8 build scores 4.10 and 29.71 dB, and the card says so plainly: 1.90x less faithful than the official file. The contact sheet backs it up, a brick-coloured tower with dark smoke where the reference has a thin white lighthouse and a light beam. Both are good ten-second videos. Only one is the video you asked for. The speed comes from residency, not arithmetic: at 11.66 GiB the model stays on the card and skips the weight traffic both other arms pay every step. Distribution is under the LTX-2.x Community License with the statement of changes Section 3.3 requires. One prompt, one seed, 512 px, and the audio branch was never decoded.

Wan 2.2 TI2V 5B drops to 2.75 GiB, and brings a VAE trap that will hand you a plausible wrong picture. JoaoZaokk/Wan2.2-TI2V-5B-W4A8-ConvRot, September 13 at 08:54 UTC, Apache 2.0. From 9.31 GiB to 2.75, 1.59x faster per step over six prompts, output the author calls indistinguishable at a glance. The pure W4A4 build of the same source ships alongside it, labelled as a measured negative: same 4-bit weights, 4-bit activations instead of 8, and every face and brick smears. Read the running note before you queue anything. This model needs the 48-channel wan2.2_vae.safetensors. The 16-channel wan_2.1_vae sitting next to it in many installs will decode the latents into something that looks fine and is wrong, caught by counting channels rather than by looking.

DreamX-Creator got a ComfyUI package, which is the difference between a paper and a Sunday afternoon. t8star/DreamX-Creator-Comfy, September 12 at 14:29 UTC, Apache 2.0: a 7B joint video and audio DiT with cross-modal attention, a separate SR-DiT 5B refiner, a CreatorDAC audio decoder, plus the Wan 2.2 VAE and UMT5-XXL encoder, arranged into the layout the T8mars nodes expect. Upstream weights are ten days older (HF createdAt September 2). Entry cost is honest and large: 20 files, about 54.25 GB, and Creator and Refiner run as separate phases so both large models are never resident together. Try the hosted Space before committing the disk. Paper: arXiv 2608.31106.

An interactive world model now runs on a MacBook Pro with a memory ceiling that does not grow with frame count. PandacatAI/lingbot-world-v2-mps, September 12 at 14:26 UTC (its own news line says September 13), adds single-device Apple Silicon support to LingBot-World 2.0's 1.3B causal-fast checkpoint. On an M4 Pro with 48 GB: 109 written frames at 480×832 in 14.4 minutes at 16.1 GB peak footprint. A 361-frame request came in at 33.5 minutes and the same 16.1 GB, because the text encoder is released after encoding, the DiT before decode, and the VAE runs spatially tiled. The 8-bit MLX text encoder is 5.6 GB instead of 10.6, with embeddings about 3% off bf16, which the author notes is the same gap as torch-on-MPS against torch-on-CPU. 4-bit was measured at 17 to 28% error and rejected. License is CC BY-NC-SA 4.0, so this is your own experiments, not client work. Free real-time hosted version at Reactor.

Image

Z-Image Turbo runs at 0.341 seconds per step in 3.06 GiB, and the mixed build that beats it stops beating it on the sibling checkpoint. JoaoZaokk/Z-Image-Turbo-W4A4-ConvRot and its de-distilled sibling, September 12 at 21:18 and 21:22 UTC, Apache 2.0. 11.46 GiB to 3.06, and per-step time on a 3090 at 1024×1024 falls from 0.903 to 0.341 seconds. Thirty-six renders across six prompts and two seeds, none broken. Each repo carries a pure 4-bit build and a mixed one with 57 layers promoted to 8 bits. On Turbo the mixed build tracks the BF16 trajectory better in 11 of 12 paired runs for 0.12 GiB and 15% more time; run the same converter at the same threshold on De-Turbo and it wins only 7 of 12. Whether promoting layers pays depends on the checkpoint, not the recipe. Budget for one catch: the matching quantized Qwen3-4B encoder saves 5 GiB of VRAM and no time, because stock ComfyUI runs every text encoder through dequantized math.

Anima gets an edge ControlNet, and the recipe is the release. khanghy1000/Anima-ControlNet-VACE-Canny, September 13 at 02:18 UTC, for circlestone-labs/Anima, the Cosmos-Predict2-derived anime model with over a million downloads. Anima has had structure control since June, when TaihoC published both the VACE training implementation and a depth ControlNet built with it, alongside community LLLite adapters for tile repair, regional prompting and inpainting. What is new is a Canny model, so you can trace line art instead of supplying a depth pass, and the fact that a second person trained one and published every hyperparameter: about 15,000 Danbooru images, cv2.Canny at fixed thresholds 100 and 200, four control blocks wired to base blocks 0, 7, 14 and 21, 3000 steps, roughly 32 GB peak VRAM, about ten hours on one L40S. Two things bind you. Anima's circlestone-labs-non-commercial-license travels to the adapter. And inference needs the fix/anima-vace-hardening branch of a fork of a fork, because the upstream training fork has a duplicate control hook bug.

Audio and music

A full song out of YuE2 dropped from 329 seconds to 213 on a DGX Spark, and the more faithful sibling is in the same table. DKmode22/YuE2-3B-NVFP4 and DKmode22/YuE2-3B-FP8, September 13 at 04:20 and 04:21 UTC, CC BY-NC 4.0 inherited from m-a-p/YuE2-3B. Only the autoregressive half is quantized, the part that plans the score and emits semantic tokens; the flow-matching stage and the VAE still run bf16 and fp32 and you need those files too. AR weights go 4.33 GB to 2.31, and AR decode on a GB10 goes from 44.8 to about 100 tokens per second, because on 273 GB/s unified memory that stage is bandwidth-bound and nothing else. 142 seconds of audio in 213 seconds of wall clock. Then the evaluation puts FP8 above it: total-variation distance on the semantic phase is 6.3% for NVFP4 against 3.8% for FP8, with the bf16-against-bf16 floor at 1.2%. Calibration used sequences the model generated itself rather than a text corpus, because text never exercises codec-token activations. W4A4 was built, measured as both less accurate and slower at batch 1, and not published.

MiniMax Music 3 in 4.43 GB, as long as you read what the number measures. jayPark777/MiniMax-Music-3-3Trit-MLQT, September 13 at 01:25 UTC, Apache 2.0: ternary weights with per-block scales, routed so audio, vocal and cross-attention tensors get finer blocks than the FFN matrices. 13.35 GB down to 4.43. That saving is on disk and download. The loader node restores dense fp16 at load to stay compatible with the existing inference path, so plan VRAM against the original figure. Generation cost is steep and stated: a 60-second song on an RTX 3060 with 12 GB and low-VRAM offload took about 27 minutes 48 seconds for AR tokens plus 2 minutes 30 for 25 DiT steps, measured August 26. The card retracts its own earlier public claim of ten-minute generation as an INT8-path estimate. Worth having for the correction alone.

Late but uncovered here: Suno v6 shipped September 9, the first Suno generation built in partnership with the music industry, following the Believe and TuneCore agreement a day earlier and BMG in August. Pair it with Universal Music Group's agreement with ElevenLabs on September 10. The terms you get will follow from deals you are not party to.

Open and local

Three of these have almost no stars, which says more about how fresh the work is than how good.

JoaoZaokk/comfy-quant-bench (3 stars): converters and a harness for 4-bit ComfyUI checkpoints that verify the kernel actually ran instead of falling back to dequantized math without saying so. Why now: three of today's four compressions came out of this repo in twelve hours.
T8mars/Comfyui-DreamX-Creator-T8 (5 stars): ComfyUI nodes for DreamX-Creator's joint audio-video generation and 2x refiner. Why now: it turns a September 2 weights drop into something you can queue.
AMAP-ML/DreamX-Creator (257 stars): the original project, native audio-video at 2K. Why now: the reference implementation the package wraps.
robbyant/lingbot-world-v2 (1.7k stars): an interactive world model with an unbounded interaction horizon and a real-time distilled variant. Why now: its 1.3B causal-fast checkpoint, released September 10, is what made the Mac fork possible.
TaihoC/sd-scripts-animaCN (0 stars): VACE ControlNet training for Anima on a kohya-ss fork, published alongside a depth adapter in June. Why now: somebody other than its author used it to ship a Canny model and printed the settings.
PineCookie/ComfyUI-Advanced-ControlNet (0 stars): the fix/anima-vace-hardening branch that adapter requires. Why now: without it the ControlNet fires a duplicate hook.
vllm-project/llm-compressor (3.8k stars): the GPTQ and compressed-tensors toolchain behind YuE2 NVFP4. Why now: the path from a music model's bf16 weights to something vLLM serves on Blackwell.

Creative workflows

1. Get a real ten-second LTX 2.5 video out of a 24 GB card. Download ltx-2.5-22b-distilled-transformer-w4a8.safetensors from JoaoZaokk/LTX-2.5-22B-distilled-W4A8-ConvRot into ComfyUI/models/diffusion_models/, paired with the gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot encoder the measurement used.

The steps. Launch with --disable-dynamic-vram. Set 249 frames at 25 fps, 512×512, 3 steps, cfg 1.0, euler, and enter the sigmas manually: 0.909375, 0.725, 0.421875, 0.0. Frame count must satisfy 8n+1, so 250 is rejected and 249 is the legal neighbour. Expect roughly 400 seconds on a 3090. How it works. The transformer's Linear layers are stored as 4-bit weights on an 8-bit activation path, small enough that the whole model stays on the card. The PCIe traffic the other arms pay every step never happens, and that is most of the speedup. Why it is good. The BF16 original does not render this clip on this hardware at all: three of four attempts died with an out-of-memory error or a process-killing access violation, and the one that worked spread 40 GB onto the CPU. Where it breaks. Composition drifts. The published sheet shows a different tower in the same scene. If you need the reference picture, use Lightricks' INT8 build and pay 8.4 GiB and 20% more time. Nobody has decoded this build's audio branch, so treat audio as untested.

2. Train a control type your model does not have yet, for about ten GPU-hours. The Anima Canny adapter card publishes the whole recipe, which makes it a template rather than an artifact.

The steps. Clone the feat/anima-vace-controlnet branch of TaihoC/sd-scripts-animaCN. Assemble roughly 15,000 captioned images at 1024×1024. Generate edge maps with cv2.Canny at fixed thresholds 100 and 200, not randomized. Attach four control blocks to base blocks 0, 7, 14 and 21. Train 3000 steps at batch 16 (2 per GPU, gradient accumulation 8), learning rate 5e-5 cosine with 150 warmup steps, AdamW_adv with stochastic rounding and bf16 state, shift timestep sampling at discrete_flow_shift=3.0, caption dropout 10%, full bf16, no gradient checkpointing. How it works. A ControlNet is a small parallel stack that reads your edge map and injects hints into a handful of the base model's blocks. Four taps across a 22-block model steer composition without retraining the part that knows how to draw. Why it is good. Ten hours on one L40S is rentable for a two-figure sum, and the fixed Canny thresholds mean your inference-time edge maps match training exactly, which is the commonest reason a home-trained ControlNet underperforms. Where it breaks. Peak VRAM is about 32 GB without gradient checkpointing, so a 24 GB card needs the recipe changed first. Anima's non-commercial license travels to whatever you train. And inference needs the hardened fork branch or a duplicate hook fires.

3. Swap a faster autoregressive checkpoint into YuE2 without forking the pipeline. From the YuE2-3B-NVFP4 card.

The steps. pip install "yue2_infer[fast]", run the repo's patch_fast.py against your installed yue2/fast.py to add a YUE2_AR_CHECKPOINT environment override, download the quantized checkpoint, point that variable at it, and call YuE2Pipeline.from_pretrained("m-a-p/YuE2-3B", backend="vllm", memory_budget_gib=16) as normal. How it works. The official pipeline derives its own Qwen3-shaped AR checkpoint internally for vLLM. The patch lets you hand it a pre-derived one. Nothing else changes; flow-matching and VAE run untouched. Why it is good. One idempotent file rather than a fork, so upstream updates do not orphan you, and the gain lands entirely on the stage that dominates wall-clock time. Where it breaks. vLLM is only used when cfg_scale is 1, the default for cot=full and cot=melody. Set cot=off and guidance becomes 1.01, the pipeline silently falls back to torch, and your checkpoint does nothing. On DGX Spark the pinned vllm==0.19.0 wheel has no working NVFP4 path and returns wrong logits without erroring, so you need a build with SM121 support and a widened version check. Take the FP8 sibling if fidelity matters more than speed; by the two cards' own token rates it runs roughly a fifth to a third slower.

Worth testing

LingBot-World 2.0 on Reactor: steer an interactive world in real time in a browser, free, before deciding whether 14 minutes per clip on your Mac is worth it. Tradeoff: the hosted build is the distilled real-time variant, so what you judge is not what you would run.
DreamX-Creator hosted Space: native joint audio and video, the thing that is hard to fake by generating separately and syncing. Tradeoff: a queue on shared hardware against a 54 GB download at home.
Anima ControlNet demo Space: TaihoC's June depth adapter, so you can see how Anima responds to structure control before spending ten L40S-hours training your own. Tradeoff: depth and Canny behave differently, and this is not the model published today.
Z-Image Turbo W4A4: a 3 GiB image model at a third of a second per step, with the fast build and the more faithful mixed build in one repo so you can A/B them. Tradeoff: nothing was tested with LoRAs, ControlNet or img2img, which is most of what people do.
Wan 2.2 TI2V 5B W4A8: 2.75 GiB, Apache 2.0, the cheapest serious video model here to try. Tradeoff: 33 frames is 1.3 seconds, and nothing published says how it behaves over a long generation.

What actually matters from today's signal

The release that changes your week is no longer the model. It is the file somebody made from the model. Four people working independently took public checkpoints and made them run on hardware their authors did not target, and the total effect on what a 24 GB card can produce beats any single launch this month. That is a different supply chain from the one most creators watch. If you spent the weekend refreshing lab blogs, you missed all of it.

So the skill that pays now is reading a quantization card, and the people writing them are getting unusually honest. Three of today's four published the arm that beats their own file: the LTX build states it is 1.90x less faithful than Lightricks' official INT8, the YuE2 card puts its FP8 sibling's better score in the same table, and the Wan 2.2 repo ships a broken W4A4 build deliberately, labelled, as evidence. That is not modesty. It is the only way a stranger decides whether your file suits their job.

The counter-signal is that these numbers do not all mean what they look like. Latent divergence, the metric quantizers reach for most, turns out to be biased in one direction: blur is a small latent change, so a build that softens everything scores better than a build that stays sharp and lands somewhere slightly different. Wan 2.2's W4A4 file sits at 0.3847, comfortably inside the range that renders fine on other models, and it is the mushy one. Meanwhile a headline like 13.35 GB down to 4.43 GB can describe a file that expands back to full size the instant you load it. On disk, in VRAM, at rest, and during generation are four different questions, and only one decides whether you can run the thing tonight.

Source access notes: Hugging Face API queries returned results roughly two weeks stale until ?cb= cache-busting was added; the uncached first call showed nothing newer than August 28. Every ship date above comes from HF API createdAt, never a listing timestamp. Star counts are cache-busted img.shields.io JSON read September 13. HF /papers?date=2026-09-12 returned an empty body. Hacker News Algolia results exceeded the fetch size limit and were skipped. blog.comfy.org was not reachable in usable form. No lab source (OpenAI, Adobe, Black Forest Labs, fal, ElevenLabs, Suno) published a creative-model release inside the 24-hour window; their most recent items are dated where cited. Midjourney, Civitai, Runway, Stability, Luma and Kling were checked via search only and produced nothing inside the window, so no claim here rests on them.

The adversarial fact-check pass ran on this draft against cache-busted primary sources and could not falsify any load-bearing claim. It produced two corrections, both folded in: a rough "last 30% of speed" figure for the YuE2 FP8 sibling was the briefing's own midpoint rather than a published number, and is now stated as the range the two cards support; and reactor.inc/lingbot-world-v2, the URL the LingBot-World card itself prints, redirects to reactor.inc/lingbot-world-2, so the canonical slug is used. The checker independently confirmed the quiet-lab thesis against Runway (latest research item September 3) and Black Forest Labs (latest blog item August 20), and could not re-verify the search-only sources named above. A third error the checker did not catch was found afterwards and corrected here: an earlier draft said Anima had no structure control before today. It has had a VACE depth ControlNet since June 20 and community LLLite adapters since June 1, confirmed by a cache-busted HF model search; the Canny adapter is a new control type, not a first.