Creative AI Briefing: Thursday, October 29, 2026

A video maker can now take a rough 768p MiniMax H3 clip and re-render it at 2560 by 1440 in three passes with an open add-on from NVIDIA's research group, a finishing job MiniMax only sold as a closed API. A narrator can clone their own voice from ten seconds of audio and have it read a script in Tamil, Hindi or ten other languages under an Apache licence. The pattern is open twins for the paid layer: the finishing and voice steps that used to mean a per-job bill now have downloadable versions. What decides whether you can use them is no longer the price. It is where you live and what the base model's licence says about your output.

New models

Open-H3-Regenerate-2K: the paid 2K finish, opened. Efficient-Large-Model/Open-H3-Regenerate-2K (created October 26, 14:55 UTC, per the Hugging Face API; project page) is a refinement add-on for MiniMax H3, the open model that makes video and sound together. You make a cheap 768p draft first, then this pass enlarges it and redraws the detail in three steps, outputting 2560 by 1440. The model card documents 2K only; the project page adds 1080p and 4K (4K by running the enlarger twice), clips of 5 to 15 seconds, and drafts from other models such as LTX-2 and SANA-Video. MiniMax's own H3-Regenerate-2K is API-only and, per NVIDIA's page, "only accepts unmodified H3 768P output." On the authors' own 150-clip benchmark the open version scores higher on aesthetics and image quality and slightly lower on motion smoothness, and wins 174 of 200 judgments in their small user study. Speed, also the authors' own: 22.45 seconds for a 5-second clip on one NVIDIA B200 with their fast decoder, against 87.26 seconds without the speedups. The honest catches are two. First, every timing is on a B200 data-centre card and the card gives no memory figure, so nobody has shown it on a desktop GPU; the code lives in the sol-engine branch of NVlabs/Sana, not in ComfyUI. Second, the refiner and decoder sit under the MiniMax H3 Community License, which lists the United States, the European Union, the United Kingdom and South Korea as excluded territories and bars using the works "or any of their Outputs" outside the permitted territory. If you work from one of those places, this is a story about what is coming, not a tool you can ship with. The project page links no demo, but a community hugging-apps Space (created October 28) is running; we did not confirm it runs the full model.

Maya 2 Native: voice cloning in 12 Indian languages, Apache 2.0. maya-research/maya2-native (created October 26, 02:58 UTC) is a streaming text-to-speech model for Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, with 20 built-in voices. Cloning needs 3 to 10 seconds of reference audio plus an exact transcript of it, and the reference language can differ from the output, so a Hindi recording can read Tamil copy. That makes it a dubbing tool for anyone narrating across Indian markets. The weights and code are Apache 2.0, apart from the upstream Kyutai Mimi audio codec (CC BY 4.0), so commercial use is allowed with attribution; the card asks you to get the speaker's permission before cloning. It needs a recent NVIDIA card: one stream holds about 7.7 GB steady and 9.2 GB while loading. Try it free in the maya2-native-studio Space, published under Maya's own account this morning, or the hugging-apps Space; the model card links neither. The catch, from the card itself: no built-in voice has an hour of Assamese training audio and Assamese output can drift toward a Bengali timbre, one- and two-word lines wobble, numbers and URLs trip it, emotion tags like [laughs] are best effort, and there is no audio watermark. Its quality numbers come from Maya's own speech-recognition judges, with no human listening study; the card says an outside leaderboard, VoiceArena, listed it second for real-time Hindi when checked October 27, though the rank interval is 2 to 4 and the entry is not tied to a specific checkpoint.

Image

A style dial separate from your prompt. heathcliff01/Kaloscope-style-adapter (created October 23; new MOSS checkpoints followed on October 25, step10000, and October 26, step17000, which the card says separates artists better) gives image style its own input. You describe the content in the prompt and set the look three ways: by an artist's name from a bank of 44,129 names, by a weighted blend of several, or by one reference image. Each add-on is about 50 MB and works with Krea 2 Raw, Krea 2 Turbo or MOSS-Image-Base in ComfyUI through the Chenkin-x node pack (19 stars, shields.io), with ready workflow files. Licences differ by part: the MOSS adapter is Apache 2.0, the Krea adapters follow the Krea 2 Community License, and the reference encoder and artist bank follow the DINOv3 License. The card does not say where its roughly 3.6 to 4.3 million training samples came from, and says nothing about the rights of the artists in its name list. The workable use is the reference-image route with art you own. The card warns that reference images and high strength can break faces and composition.

Video

Genie 4 makes a scene you steer, with sound. Google DeepMind's Genie 4 (October 27) is a real-time world model: you move through a generated scene with the keyboard and change it with text (weather, camera, added people or objects), and it now generates music, ambient sound and dialogue alongside the picture. Google says testers already use it to run multiple takes of a scene for previsualization. Access is a research preview for trusted testers, with Project Genie coming to Google AI Ultra subscribers and early-access Gemini API partners; there is a waitlist. Google's own caveats: visual details and object permanence "can still shift or drift over time," and complex physics, audio accuracy, legible text, spatial accuracy and long scenes remain open problems. All video and audio it generates carries a SynthID watermark.

Motif-Video-2B in an FP8 file. axteam/Motif-Video-2B-FP8 (created October 29) is a smaller-file repack of Motif Technologies' Apache 2.0 text- and image-to-video model from April, tagged for ComfyUI. It is a 2.6 GB file, about half the original size; the card gives no speed or memory figures.

Audio and music

"Logic Pro 13" is coming with an AI assistant (secondary sources). German trade site gearnews (October 27) and MusicRadar (October 28) report a Logic Pro update arriving in November (the version number itself is community consensus, not Apple's), free for existing owners: an assistant built on Apple and OpenAI models that can start projects and generate a vocal track from a text prompt, a Browser that analyses and tags your sounds, and a Stem Splitter that can now isolate individual drum pieces. Apple's Logic Pro page did not mention any of this this morning and we found no Apple press release, so treat the details as reported, not confirmed.

YuE2 on a Mac, and a licence claim worth checking. mlx-community/YuE2-3B-bf16 (created October 28) runs the full-song model through a Swift package on Apple Silicon: a 58-second song in 41 seconds on an M5 Max at full precision, with peak memory of 11.9 to 14.6 GB for a worst-case six-minute song depending on the compression tier, per the publisher. Its card says M-A-P lets individual creators publish and sell songs they make while companies need a licence. The YuE2 LICENSE file itself is plain CC BY-NC 4.0 and grants no such permission in its text, so that claim still needs M-A-P's own wording before you sell anything.

One-step sound generation, non-commercial. xingyuchen-8/DriftAudio (created October 28) makes general audio from text in a single step; demo page. CC BY-NC-SA, needs separate upstream parts, and gives no timing figures.

Open and local

Today's local story is Macs catching up on video and song. An independent MLX port runs Lightricks' LTX-2.5 on a 24 GB M5 at a 9.3 GB peak, though a 5-second 1024 by 576 clip takes about 7.3 minutes and comes out silent, per its author. The YuE2 port above renders songs faster than real time on a high-end M5 Max. On the PC side, the interesting files are add-ons that replace paid steps, and the interesting fine print is in the base licences.

Efficient-Large-Model/Open-H3-Regenerate-2K: draft at 768p, finish at 1440p in three passes; created October 26, MiniMax H3 licence with territory exclusions (card).
NVlabs/Sana: the research code that runs it, sol-engine branch; 9.7k stars, shields.io (repo).
maya-research/maya2-native: 12-language voice cloning from 3 to 10 seconds of audio, Apache 2.0; created October 26 (card).
MayaResearch/maya2-native: the streaming server and inference code; 19 stars, shields.io (repo).
Chenkin-x/Kaloscope-Style-adapter-Comfyui: set style by reference image or artist name, separate from the prompt; 19 stars, shields.io (repo).
zero0303/ltx-2.5-mlx-dynamic-8bit: LTX-2.5 video on Apple Silicon without PyTorch; created October 28, LTX-2.x Community License, which requires a paid licence above US$10M revenue (card).
mlx-community/YuE2-3B-bf16: full songs with vocals on a Mac; created October 28, CC BY-NC 4.0 (card).

Creative workflows

1. Style from your own reference image, not a famous name. Source: Kaloscope card and its workflows/Kaloscope_MOSS_reference.json (newer MOSS examples, Kaloscope_MOSS_B8_reference.json, in workflows/moss-b8-step10000/), plus the node pack.

The steps. Install the node pack. Download the repo with hf download heathcliff01/Kaloscope-style-adapter --local-dir Kaloscope-style-adapter, then run python Kaloscope-style-adapter/install_models.py --comfyui "/path/to/ComfyUI" --variant moss --with-encoder (the encoder is only needed for the reference route). Drag the non-_api reference workflow into ComfyUI, point it at one image of your own work, and test at 512 pixels before scaling. Card settings for MOSS: CFG 3, style guidance 6, shift 5, strength 1.0. For Krea 2 Raw or Turbo the default strength is also 1.0; compare 0.8 on a fixed seed. The installer puts in step10000 by default; select the step17000 file by hand in the adapter loader if you want the newer one.

How it works. A small image encoder reads the look of your reference and feeds it to the generator through its own channel, so the prompt only has to carry the subject.

Why it is good. An illustrator can generate layout roughs and background plates in their own palette and brushwork without training anything.

Where it breaks. The card describes one reference image per style, faces and bodies destabilise at high strength, the card does not promise MOSS matches Krea 2's artist separation (step17000 narrows the gap), and FP8, GGUF and ControlNet combinations are "not yet fully verified." The training data is undisclosed.

2. Draft cheap, finish at 2K. Source: Open-H3-Regenerate-2K card and project page.

The steps. Generate and pick your shot as a 768p H3 draft (1344 by 768, 121 frames for 5 seconds). Fetch the bundle with the download.py script in the code folder. Run the refiner: it enlarges the draft, adds back a measured amount of noise, and redraws it in three steps. For 4K, the project page says to apply the enlarger twice; the card itself documents 2K only.

How it works. The draft already fixes layout and motion; enlarging alone adds no detail, so the three passes draw the surfaces again at the higher size.

Why it is good. You spend generation time on choosing takes at low resolution and only pay the high-resolution cost on the one you keep.

Where it breaks. Data-centre timings only, no memory figure, research code rather than a ComfyUI node, and a licence that rules out the US, EU, UK and South Korea, outputs included.

Worth testing

maya2-native-studio Space: clone your voice and hear it in another language. Tradeoff: needs an exact transcript of your sample, and numbers and URLs still trip it.
Kaloscope reference workflow: your style as a separate dial. Tradeoff: local ComfyUI only, and faces break at high strength.
Genie 4 waitlist: previz you can walk through. Tradeoff: trusted testers first, then the top Google AI plan, and scenes drift.
DriftAudio demo page: hear what one-step sound generation sounds like. Tradeoff: non-commercial and fiddly to run.

What actually matters from today's signal

The paid step is the next thing to go open. Last week's open video models gave you the draft; the closed API sold you the finish. NVIDIA just published the finish. Maya did the same for multilingual voice cloning, a job most Indian-language creators have rented from closed services. The money argument for drafting locally and finishing in the cloud gets weaker each time this happens.

Licences now do the gatekeeping prices used to. Open-H3-Regenerate-2K is free and better on the authors' own scorecard, and a creator in Los Angeles, London or Berlin still cannot use it, because the H3 licence follows the output. Maya 2 is Apache apart from an attribution-licensed codec, and usable anywhere. Kaloscope's licence is fine, but it ships a list of 44,129 artist names and no word on where its training images came from. Read three documents before you build: the base licence, the add-on licence, and whatever the card does not say.

The counter-signal is hardware. The most impressive release today has only been timed on a B200. Open does not mean runnable on your desk, and the Mac ports show the gap honestly: LTX-2.5 fits in 9.3 GB, and you wait seven minutes for five silent seconds.

Source access notes: Quiet primary-lab window apart from Genie 4. Nothing new for creators October 27 to 29 from Adobe (latest October 19), BFL (October 15), Stability (August 25), ElevenLabs (September 30), Suno (October 13), fal (October 20) or Replicate. OpenAI's news page failed to load. Runway's news page showed no dates. Luma's October 27 posts were guides and roundups, not launches. Krea's October 28 Seedance entry was covered yesterday. Midjourney via WebSearch only: nothing current. HN Algolia showed no creative stories. Apple's newsroom page returned no listings, so Logic Pro 13 rests on gearnews and MusicRadar. Adversarial fact-check ran (Sonnet subagent): it confirmed all Hugging Face dates, star counts, the Open-H3 timings and benchmark figures, the H3 licence quote, Maya's specs and the gearnews details, and caught these, all fixed: a running hugging-apps Space for Open-H3 we had called absent, a stale Kaloscope checkpoint (step17000 followed on October 26), "images" that are training samples, Maya's CC BY codec and an overstated VoiceArena rank, the unofficial Logic Pro version name, run-together YuE2 speed and memory figures, an inaccurate "second-day" label on the Motif repack, the LTX licence revenue cap, Genie's SynthID and caveat wording, and the 4K claim being project-page only. It could not open MusicRadar's article or the H3 licence's Applicable Territory definition. Cloud and local shells had no network access to Hugging Face or GitHub, so all Hugging Face dates come from the API via WebFetch (createdAt) and star counts are shields.io totals. GitHub trending and Trendshift not used; no star deltas cited. Civitai and blog.comfy.org not attempted. Omitted as covered recently: ComfyUI v0.43.0, Hy-Nomas, Qwen-Image 2.1 consistency add-ons (created September and early October).