Creative AI Briefing: Friday, September 18, 2026
Ask for fourteen seconds of 768p video with synced sound and get it back in about nine. That happened yesterday, twice, from two unrelated groups. A Berkeley-led team published the architecture and the weights behind a MiniMax H3 variant that denoises a 14.4-second clip in 6.9 seconds; fal published the platform write-up behind a model that makes five seconds in under three, plus the streaming layer that lets you steer the output while it runs. The number that matters is not the speedup multiple. It is the crossover: generation is now quicker than playback, which changes a video clip from something you submit and abandon into something you sit with and revise. Everything else today is the local tier catching up to that, and one licence that says you may not.
New models
Video DeltaNet on MiniMax H3, or VDN-H3, is the release. Weights sat on Hugging Face since September 2 (createdAt September 2, 05:03 UTC; card refreshed September 17, 05:46 UTC), but the architecture write-up and the arXiv paper both landed September 17, and the paper is what makes the numbers checkable. The team is Haocheng Xi and colleagues at UC Berkeley, Impossible Inc. and UT Austin. It is not a new model so much as a new interior for one you may already use: MiniMax H3 with its attention rebuilt.
What you can make with it is the same thing H3 already makes, text-to-video or first-frame or last-frame conditioned, with synced audio, at roughly the same quality. What changes is the clock. Their published table for a 14.4-second 768p job, denoising only: a single B200 running stock H3 at 50 steps takes 13.95 minutes. VDN-H3 at 8 steps on that same single card takes 51 seconds. Eight cards takes 11.23 seconds on their own reference stack, and their SGLang serving path gets denoising to 6.9 seconds with a finished file back in about 9. On an H200 the single-card 8-step figure is 90.5 seconds. Those are steady-state numbers that exclude model loading, warm-up and the audio and video decode, which they say they would run on separate machines in a live setup. Treat them as the authors' own measurements.
The paper and the project page do not quite agree with each other, which is worth knowing before you repeat a number. The arXiv abstract states 6.70 seconds of denoising for a 14.3-second 768p clip on eight B200s, a 14.5x speedup over the 50-step dense baseline on the same eight GPUs. The project page and the model card use 14.4 seconds and 6.9 seconds, and quote the much larger multiple against a single-GPU baseline. Same work, two roundings and two reference points. Nine seconds to a finished file is the figure that describes what you would actually experience.
The honest catch is the licence, and it is a big one. VDN-H3 inherits the MiniMax H3 Community License Agreement, whose applicable territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America. The card states plainly that use outside that territory is not authorised and invites people in excluded regions to contact MiniMax. If you are reading this in any of those places, the open weights are not open to you. The download is also 82 GB, of which 72 GB is the H3 base you may already have. Free hosted demo, if you just want to see it: akhaliq/vdn-minimax-h3-demo (Space createdAt September 8, currently RUNNING on zero-a10g), with OpenVDN's own Space alongside it.
Video
fal published how H3 Max gets a five-second clip out in under three, and what it built on top of that headroom. The platform write-up went up September 17 at 23:36 UTC. H3 Max itself launched August 26; this post is the machinery. Two parts matter to anyone making things rather than serving them.
The first is a rule the team says it held to during post-training: an optimisation only shipped if the model kept its rank in the quality evaluations. Their point, in paraphrase, is that plenty of speed tricks look excellent on a latency chart while costing you the output without announcing it, and that those were rejected outright rather than shipped with a caveat attached. Worth holding a vendor to.
The second is WMA, the World Model Accelerator, which is the part you will care about in six months. It serves a model over WebRTC, the protocol video calls use, instead of over request and response. The client connects once, the runner holds a live session, media streams as a browser video track, and control messages ride back on a data channel so a new prompt lands in the running stream within seconds instead of becoming a new job in a queue. fal.live runs on it today: continuously broadcasting channels steered live by whoever is watching, where one channel costs the same GPU at ten viewers or ten thousand. WMA is marked experimental, as is the multi-node inference H3 Max serves on. But the shape is clear: once a model renders faster than it plays, the queue is the thing standing between you and a dial you can turn.
ShotStream deserves naming here, with its date corrected. KlingTeam/ShotStream surfaced near the top of Hugging Face's recently-updated video feed yesterday, which is a card refresh, not a release. The two primary sources put the release in March and disagree by ten days: the Hugging Face API gives a repo createdAt of March 17, 2026, while the GitHub changelog dates the code and checkpoint release to March 27. Either way this is months-old work, and ECCV accepted the paper on June 18. It is still the clearest statement of where the streaming idea goes: a causal multi-shot architecture that generates the next shot conditioned on what came before, at 16 FPS on a single NVIDIA GPU, driven by prompts you type while it runs. Built on Wan2.1-T2V-1.3B, Apache 2.0, 181 stars. A small model doing interactive storytelling on one card is the version of this that a working creator can actually hold.
Open and local
The local tier spent yesterday making LTX-2.5 fit on cards people own, and putting two speech models on Apple Silicon.
- comatto/LTX-2.5-Distilled-GGUF: the distilled LTX-2.5 video model in seven sizes, so you can pick the one that fits your card instead of the one the lab shipped. Why now: createdAt September 16, 22:51 UTC, with a published size table from 12.6 GB (Q3_K_S) to 23.6 GB (Q8_0), Q4_K_M at 15.7 GB named as the recommended baseline, and two ready ComfyUI workflow files in the repo. One wrinkle: the README's own links point at a different repo id (
Abiray/LTX-2.5-Distilled-GGUF), so this looks like a re-upload with the links left alone. Files are present here either way. - majentik/Audio8-TTS-Preview-0.6b-MLX-4bit: voice cloning and eleven languages of speech on a Mac, no Python GPU stack. Why now: createdAt September 15, 16:33 UTC, Apache 2.0, 4-bit MLX build of Edge0/Audio8-TTS-Preview-0.6b, with an 8-bit sibling published a minute later.
- cloud0day3/antalia-1: a single-speaker Turkish speech model, useful if you have been making do with a multilingual model that gets Turkish prosody wrong. Why now: createdAt September 14, 09:57 UTC, roughly 0.3B parameters, OpenRAIL-M, and it ships prosody presets and a timbre profile as separate files rather than burying them. Free hosted demo at cloud0day3/antalia.
- OpenVDN/vdn-minimax-h3: the fast H3 above, listed here because the release includes the optimised inference stack and the training code, not just weights. Why now: 311 likes and 603 downloads as of this morning, and sixteen community Spaces already point at it.
- vanch007/FastVideo-FastH3-8-Step-V2-MLX-INT8: yesterday's Mac build of the other fast H3, carried forward because it is the only way to run this family on Apple Silicon and the author published his measured 31 minutes per clip on an M3 Max.
Audio and music
Nothing shipped from ElevenLabs, Suno or Adobe inside the window; their most recent posts are September 10, and Adobe's newest creative-tool news is the September 8 Premiere and After Effects timeline generation announcement. The movement was all in the open speech tier, which is now producing a steady drip of Apple Silicon and ONNX conversions within a day or two of each upstream model. The Audio8 MLX pair above is the one to grab if you narrate your own work on a Mac.
Creative workflows
1. Try the faster H3 on someone else's GPU before you commit 82 GB and a licence problem.
The steps. Open akhaliq/vdn-minimax-h3-demo, which runs the 8-step VDN-H3 checkpoint and returns video with a soundtrack. Write your prompt the way MiniMax asks for it: the VDN card explicitly recommends rewriting the prompt through H3-Context-IR or the official prompt-writing skills before you generate, and says it greatly improves output. If you want it locally afterwards, hf download OpenVDN/vdn-minimax-h3 --local-dir ckpts pulls all 82 GB, and bash scripts/8nfe_tuned_fp8.sh is the single-GPU first render. The first run compiles kernels and takes several minutes; later runs reuse the cache.
How it works. The team's own measurement is that plain attention eats more than 85% of H3's runtime. Their fix splits the video's attention in two. Nearby frames get exact, full-strength attention in a sliding window, which is what keeps local detail and short-term stability. Distant frames get handled by a cheap running summary carried forward and backward through the clip. On top of that, every frame also attends to the first and last frames as anchors, which they measure as adding only 3.57% more attention work while noticeably improving long-range stability. Two learned volume knobs balance the exact path against the summary path so one does not drown the other.
Why it is good. It is a bolt-on. The checkpoint adds a separate branch plus two small adapters that merge into the backbone at render time without touching the base weights, so this is not a fork of H3 you have to maintain. And the training code shipped with the weights, which is rarer than it should be.
Where it breaks. The headline nine seconds needs eight B200s. On one card you are at 51 seconds (B200) or 90.5 seconds (H200) for the 8-step version, which is still a different working day from 14 minutes but is not interactive. The published figures exclude loading, warm-up and the decode step. And the licence excludes the US, UK, EU and South Korea, so for many readers the hosted Space is the only lawful route.
2. Pick your LTX-2.5 file by your card, not by the download button.
The steps. Go to comatto/LTX-2.5-Distilled-GGUF and read the size table before you click anything. Grab one video file: LTX-2.5-Distilled-Q4_K_M.gguf at 15.7 GB is the maintainer's recommended baseline, Q3_K_S at 12.6 GB is the floor, Q8_0 at 23.6 GB is described as nearly indistinguishable from the original. Then get the three supporting pieces from the upstream Lightricks/LTX-2.5 repo, because they are not in this one: the text encoder (take gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors at 15.4 GB, not the 26.3 GB bf16 one, unless you have room), ltx-2.5-video-vae-bf16.safetensors at 1.47 GB, and ltx-2.5-audio-vae-bf16.safetensors at 365 MB if you want the synced sound. Then load video_ltx2_5_t2v_GGUF.json or video_ltx2_5_i2v_GGUF.json from the same repo in ComfyUI.
How it works. LTX-2.5 is a split-component model: the generator, the part that reads your prompt, and the parts that turn numbers back into pixels and audio all live in separate files. The GGUF files here only compress the generator. That is why the quantized file size is not your memory requirement: the text encoder is the same size again or larger, and it is the piece most people forget to account for.
Why it is good. Seven tiers with published sizes means you can size the download to the card you actually have. The included workflow files skip the usual hour of wiring loaders by hand. And the LTX-2.x community licence permits commercial and production use at no cost for entities under $10M annual revenue, which for most people reading this is a real yes rather than a hedged one.
Where it breaks. Budget for the encoder and the decoders or you will run out of memory with a correctly sized generator file loaded. The repo's own links point at a different account, so double-check you are pulling from the repo you think you are. And the model card's own limitations list is blunt about prompt following depending heavily on prompting style, and about artifacts in complex scenes.
Worth testing
- akhaliq/vdn-minimax-h3-demo, free, for the fast H3 with sound. Tradeoff: it is a community Space on shared hardware, so the nine-second figure from the paper is not what you will experience at the queue.
- H3 Max on fal if you want the fastest hosted path and will pay per generation. Tradeoff: closed, metered, and the latency you get depends on whether a warm runner exists when you press go.
- fal.live to see what steering a running stream feels like before you build anything on it. Tradeoff: WMA is explicitly experimental, and running a model continuously is expensive today.
- comatto/LTX-2.5-Distilled-GGUF if you have 16 GB or better and want video with audio on your own machine. Tradeoff: no hosted demo, and the real memory cost is the generator plus the encoder plus two decoders.
- Audio8 on MLX for narration on a Mac. Tradeoff: it is a preview model in a 4-bit conversion by a third party, with no published quality comparison against the original.
What actually matters from today's signal
The interesting thing about yesterday is not that video got faster. Video has been getting faster every month for two years. It is that two groups independently crossed the same line and then said out loud what the line is for. fal's phrasing is the clearest: the crossover produces video you iterate on rather than wait for. That is a different creative act. When a clip takes fourteen minutes you write the most complete prompt you can and hope. When it takes nine seconds you make the bad version on purpose, look at it, and change one thing. Every working creator knows which of those two loops produces better work, and it is not the careful one.
Watch what the fast labs are building next, because it is not faster clips. It is the connection that stays open. WMA is WebRTC, which is to say it treats your generation as a phone call rather than a delivery. ShotStream, six months old and running on a 1.3-billion-parameter base, already generates the next shot conditioned on the shots before it while you type. Put those two together and the unit of AI video stops being the clip. That is a genuinely new medium, and nobody has figured out what to make in it yet, which is the fun part.
The counter-signal is money and jurisdiction, and it is sharp. Every headline number today requires eight B200s. On the single card most people can rent, the same model takes 51 seconds, which is excellent and is not a stream. And the open weights that make this checkable are governed by a licence whose applicable territory excludes the United States, the United Kingdom, the European Union and South Korea, which is to say most of the people who will read the paper. The architecture is public, the code is public, the training recipe is public, and for a large share of the audience the lawful way to use any of it is to pay someone else to run it. That gap between what is published and what you are permitted to run is the real constraint on this week's speedup, and it is not getting narrower.
Source access notes: An adversarial fact-check pass ran on this draft and found three problems, all now fixed. First, the draft asserted a single March 27 ship date for ShotStream from the GitHub changelog without reconciling it against the Hugging Face API createdAt of March 17; both dates are now named and the ten-day gap is stated rather than hidden. Second, the VDN-H3 headline numbers differ between sources, 14.3 seconds and 6.70 seconds and a 14.5x speedup in the arXiv abstract against 14.4 seconds and 6.9 seconds and a larger multiple on the project page and model card; that disagreement is now spelled out instead of silently averaged. Third, a sentence from the fal post was presented as a direct quote with a word dropped and the punctuation altered; it is now marked as paraphrase. Everything else the pass checked, every Hugging Face createdAt, the full B200 and H200 latency tables, all file sizes, the licence territory and revenue terms, the star count and the Space statuses, verified against primary sources unchanged. Hugging Face dates come from /api/models/... createdAt, never listing timestamps. GitHub star count via shields.io with cache busting. curl has no egress from this workspace, so all API reads went through the fetch tool. The ComfyUI GitHub releases feed returned a response too large to read and was skipped this run. Midjourney updates and Civitai were not reachable in previous runs and were not retried. OpenAI, Adobe, Replicate, Stability, Black Forest Labs, Luma, ElevenLabs and Suno published nothing inside the 24 to 48 hour window; their most recent items are dated September 4 to September 10. No music-generation item cleared the bar today, so that section is folded into Audio. Image section omitted: the only fresh items were the Ideogram v4 quantized builds and the Krea 2 line-art adapter, both covered in yesterday's briefing.