Creative AI Briefing October 15, 2026 morning edition 11 min read

The distance between a launch and a version that runs on a creator's own machine shrank to about a day this week, so the question that decides whether you can use a model moves from your graphics card to the licence.

Fish Audio Drama 3 FlashVSR+SMD MiniMax H3 Multics Taste Skill audio-gen voice-clone upscaling video-gen local-creative-ai open-weights

Creative AI Briefing: Thursday, October 15, 2026

You can now write a short radio scene as a script, with stage directions in square brackets and a laugh placed exactly where the line needs it, and get it back performed. That is Fish Audio's Drama 3, launched Tuesday with open weights. By Wednesday morning the community had a version running on a Mac. The same compression shows up across the week's open releases: a 4K video upscaler built from the start to fit one 24 GB consumer card, and a four-step workflow for MiniMax H3 that clips memory to about 24 GiB. When the hardware gate drops that fast, the gate that is left is the licence, and today's three releases sit under three very different ones.

New models

Fish Audio Drama 3 turns a script into a directed voice performance. Released October 13, Drama 3 is a speech model for audiobooks, audio drama, game dialogue and short video. You direct it the way you would mark up a script for an actor: a bracketed instruction before a line (mood, a whisper, a laugh), which can change mid-line; inline cues such as breaths, which land where you place them; and paired tags like <emphasis>…</emphasis> that touch only the words inside. Direction can be written in any of its 83 languages, regardless of the language being spoken. The model card adds speaker tags for up to 16 characters in one text and voice cloning from a reference clip plus its transcript. Fish Audio's own advice is the honest catch: "Takes vary, so render a few and keep the one that plays." The headline editing features, regenerating one changed word inside a finished take and asking for a take "softer," are in invite-only testing. Licence: Fish Audio Community License v1.0. Free for research and non-commercial use, and free for commercial use if you and your affiliates earn under US$500,000 a year; above that, or to run it as a voice service, you need a deal with Fish Audio. Cloning a real person's voice without consent is prohibited. Where it runs: the fish.audio web app and phone apps, the API (Fish lists $50 per million characters, cut to $10 until November 12), OpenRouter, and open weights. The full model is about 35 billion parameters, roughly 3 billion active at a time, and Fish's own speed figures come from an H200 data-center card. The pronunciation scores in the post (93.2% versus 88.1% for Fish's previous S2.1 Pro) are Fish's own report of a third-party benchmark. Try it free: the Drama 3 Space on Hugging Face (created October 13, running on shared ZeroGPU hardware).

Video

FlashVSR+ upscales a whole 4K frame on one 24 GB card. OpenImagingLab/FlashVSR_Plus (paper arXiv 2610.17552, on Hugging Face October 14; GitHub repo created October 14; weights repo ruchangyao/FlashVSR_plus created September 27 and refreshed October 14) is a one-pass video upscaler from Shanghai AI Lab, CUHK and Tsinghua. "Untiled" is the useful word: it upscales the full 4K frame at once instead of stitching tiles, which is where visible seams and flicker usually come from. The project page shows the use case most creators will want, a MiniMax H3 clip at 1344x768 taken to 3840x2160. Its own RTX 4090 chart gives 0.98 frames per second for the Turbo version and 0.50 for Base, against 0.18 for SeedVR2, so a five-second 24 fps clip is roughly two minutes of waiting on Turbo. All of those figures are the authors' own, and the page publishes no standard quality scores, only side-by-sides and an unquantified "consistent gains in user preference." Licence: Apache 2.0 on both code and weights, so client work is fine. Catches: Linux, NVIDIA and CUDA only, no ComfyUI node yet, frame sizes must divide by 128 (the README says input in one place and output in another), and one of its attention backends needs NVIDIA's compiler to build, though the README recommends the simpler Triton backend for an RTX 4090. Try it free: hugging-apps/flashvsr-plus (Space created October 14).

A four-step MiniMax H3 workflow makes a two-character clip with sound. zjwfufu/SMD-MiniMax-H3-Ref2VA-4step-ComfyUI (created October 13) packages a fast version of MiniMax H3 as a complete ComfyUI workflow: two reference pictures plus a prompt give about five seconds of 1344x768 video at 24 fps with synchronized audio. The compact profile peaked at about 24.3 GiB on a memory-limited A100; real 24 GB consumer cards are untested, and the download is 42 GB. No speed or quality comparison is published. It sits under the MiniMax H3 Community License, whose grant excludes the European Union, the United Kingdom, South Korea and the United States; creators there need a separate licence from MiniMax.

Multics is a research model for shared game-like worlds. TencentARC/Multics (created October 12; code Apache 2.0, while the README says weights keep their upstream terms; built on Wan 2.2 TI2V 5B) generates video for four players at once, and the same model file runs 6, 8 or 10 players without retraining. The cost is memory: the README reports about 60 GiB peak for four players and 106 GiB for ten at 640x352 per player. No hosted Space; a research exhibit for now.

Audio and music

The Drama 3 story did not stop at launch. Within about a day, three smaller builds appeared. An int4 version that Hugging Face's AK posted on October 13 (akhaliq/drama-3-int4) was re-hosted under Fish Audio's own account on October 14 (fishaudio/drama-3-int4). It halves the download to 22.1 GB and the weights to 20.8 GB of card memory, though the only tested hardware is an A100 80 GB and an H200, where the full serving stack peaked near 58 GB. Its card reports speech accuracy on par with the original and a request about 8% slower, and warns that tag-heavy takes are occasionally cut short. For Apple Silicon, mlx-community/fish-audio-drama-3-4bit (created October 14) is about 22 GB on disk and peaks around 26 GB while speaking, so it wants a Mac with 32 GB of memory at the very least. Its card reports about twice real time on an M3 Max. It is an unofficial conversion ("Fish Audio did not release or endorse it") and needs mlx-audio installed from an unmerged pull request branch. Both carry the same Community License as the original.

Open and local

The local story today is speed of arrival. Drama 3 went from launch to a Mac build in under a day, FlashVSR+ was designed around one consumer card, and SMD's H3 workflow trades sampling steps for a memory budget near 24 GiB. What none of them changes is what you may sell: Apache 2.0 on FlashVSR+, a revenue ceiling on Drama 3, a licence on H3 that excludes the US, EU, UK and South Korea.

OpenImagingLab/FlashVSR_Plus: one-pass 4K video upscaling on a single consumer card, Apache 2.0. Why now: paper and repo landed October 13 and 14; 68 stars on shields.io (repo).
fishaudio/fish-speech: the code home for Drama 3's open release. Why now: Drama 3 shipped October 13; 33k stars on shields.io (repo).
mlx-community/fish-audio-drama-3-4bit: Drama 3 on Apple Silicon, unofficial. Why now: created October 14 (repo).
zjwfufu/SMD-MiniMax-H3-Ref2VA-4step-ComfyUI: two-reference H3 clips with sound in four steps, as a drop-in ComfyUI workflow. Why now: created October 13 (repo).
TencentARC/Multics: multi-player generated worlds, research stage. Why now: on Hugging Face's daily papers this week; 65 stars on shields.io (repo).
Leonxlnx/taste-skill: an agent skill that builds a landing page from one of 20 researched design directions, then checks it in a real browser at three widths. Why now: on Trendshift's daily list; 97k stars on shields.io (repo).
iterative-computer/powermove: an AGPL-licensed Mac motion and video editor that Claude Code or Codex can operate through undoable edits. Why now: Show HN on October 13 (repo dates to August); 307 stars on shields.io (repo).

Creative workflows

1. Direct a multi-voice scene in Drama 3: Fish Audio Drama 3 post and model card

The steps. Write the scene as a script. Put a bracketed direction before each line where the delivery matters, place breaths and laughs inline where they happen, and wrap single words in paired tags such as <emphasis> (in the web app you select the words and pick from the Tags menu). For several characters, use the model card's numbered speaker tags, up to 16 per text. Try it in the free Space first, then the fish.audio app or the API with the model header set to drama-3. Render three or four takes of each line and keep the best.

How it works. Fish tuned the model on a single reward, a model of which take people preferred, so it aims for the most playable reading rather than the most literal one.

Why it is good. Direction lives in the script, so a rewrite or a new language keeps the performance notes attached to the line instead of in your head.

Where it breaks. Takes vary by design, and the fix-one-word editing is invite-only for now. Commercial use stops being free above US$500,000 in annual revenue. The open weights are heavy: even the int4 build was only tested on 80 GB-class cards.

2. Take an AI clip to 4K without tiling seams: FlashVSR+ (inference/run_stream, ckpts/turbo/)

The steps. On a Linux machine with an NVIDIA card, create a Python 3.10 environment, pip install -r requirements.txt and pip install -e . --no-deps, then put the Turbo files from Hugging Face in ckpts/turbo/ with TCDecoder.ckpt in ckpts/. Run python -m inference.run_stream --input clip.mp4 --mode turbo --output clip_4k.mp4 with your target --height and --width (the README's closest 4K example is --height 2176 --width 3840 --decode-every 28). Switch to --mode base for a hero shot. For fast motion, lower --temporal-compression from 4 to 2 or 1; for long clips, add --decode-every.

How it works. Most of the model looks only at nearby pixels across neighbouring frames, with a few layers that see the whole frame, which keeps memory low enough to skip tiling.

Why it is good. One pass, no tile stitching, and an Apache licence you can deliver client work under.

Where it breaks. Sizes must divide by 128, so true 2160-line UHD means rendering 2176 lines and cropping. No Mac or CPU path. No published quality scores. Large motion costs extra time.

3. Two characters, one clip, with sound, in four steps: SMD MiniMax H3 Ref2VA (workflow.json, SETUP.md, download_models.py)

The steps. Run download_models.py, follow SETUP.md, and load workflow.json (the compact profile). Connect two reference images and refer to them in the prompt as <Picture 1> and <Picture 2>, keeping each image matched to its tag. Leave the sampler as shipped: Euler, BasicGuider, the supplied step schedule, add-on strength 1. Do not stack H3 Lightning or Turbo add-ons on top.

How it works. SMD is a distillation method, a way of training a student model to reach in four steps roughly what the original reaches in many more.

Why it is good. A complete, tested graph instead of swapping weights into ComfyUI's default template, which the card says fails with errors like a missing audio_scale.

Where it breaks. 42 GB to download, 24 GB consumer cards untested, no timing or quality figures, Linux and NVIDIA only, and the H3 licence's territory terms.

Worth testing

Drama 3 Space: paste a six-line scene with bracketed directions and compare three takes. Tradeoff: shared free hardware means queues, and the best features are invite-only.
FlashVSR+ Space: upscale one of your own generated clips with fine text in frame. Tradeoff: no published quality metrics, so your eyes are the benchmark.
Drama 3 on a Mac via the MLX build. Tradeoff: unofficial, a pull-request install, and about 26 GB of memory while it speaks.
SMD H3 four-step workflow for two-character dialogue tests. Tradeoff: a 42 GB download before you see a frame.

What actually matters from today's signal

The lag between "a lab released it" and "it runs on my machine" used to be measured in weeks. Drama 3 shipped on Tuesday and had a Mac build on Wednesday. FlashVSR+ skipped the wait entirely by designing for one consumer card from the start. For creators this is a planning change: assume a strong open release will reach your hardware within days, and spend the waiting time writing the scripts and shot lists it will need.

What did not shrink is the fine print. Three releases, three regimes: Apache 2.0 on the upscaler (sell anything), a US$500,000 revenue ceiling and a voice-consent clause on Drama 3, and a territory-bound community licence on H3. The voice model that is easiest to run this week is also the one that asks the most of you ethically, because cloning is one reference clip away. Read the licence before the install guide.

The counter-signal: every quality number in today's batch is the maker's own. Fish's pronunciation score, FlashVSR+'s speed chart and the int4 card's accuracy comparison all come from the people who built them. Test on your own material before you promise a client anything.

Source access notes: Checked with no new creative launches in the window: OpenAI news (newest Oct 8), Adobe blog (newest creative post Oct 2; Adobe MAX is November 10 to 12 per Creative Bloq), ElevenLabs blog (newest Sep 30), Black Forest Labs blog (newest Sep 23), Runway research (newest dated Sep 3), Stability AI news (undated), Luma blog (Oct 14 posts are listicles), Suno blog (Oct 13 post is a behind-the-scenes piece), fal blog (newest Sep 17), Replicate blog (stale listing), CDM, MusicRadar, PetaPixel (AI items are commentary, not tools). DeepMind blog served an undated, stale list. blog.comfy.org needs JavaScript and was skipped. No Midjourney, Kling, Seedance, Pika, Ideogram, Recraft or Krea news dated this week surfaced in WebSearch. Image and Writing sections omitted: nothing new from a primary source. All Hugging Face dates are createdAt from the models and spaces APIs; GitHub creation dates and star totals from shields.io with cache busting; Trendshift daily list used for presence only, no deltas cited. A Hacker News item claiming an AI agent erased a full season of a Swedish TV show was left out because the link resolved to a translation proxy and the primary report could not be read. Adversarial fact-check ran (Sonnet subagent) against primary sources. It caught and we fixed: Powermove's licence (AGPL, not GPL) and star count (307, not 6); Multics' base model (Wan 2.2 TI2V 5B), weights licence caveat and its published memory figures; the H3 licence's excluded territories, now stated outright; FlashVSR+'s 128-divisibility ambiguity and a better 4K example (2176 by 3840); and an overstated build requirement. Noted: the Drama 3 weights repo carries an October 1 createdAt, ahead of the October 13 launch. arXiv rate-limited the checker, so FlashVSR+ affiliations rest on the Hugging Face papers listing.