Creative AI Briefing: Monday, November 16, 2026
You can take any 3D model you own this morning, a scan, a sculpt or a character from a generator, and turn it into an orbiting, animated pixel-art diorama in your browser with one command. That look did not exist as a free tool on Friday. Over the weekend one artist's recipe (shape the world in code, cover it in tiny 3D blobs, run a retro pixel pass) became at least three independent open-source toolkits, each with its own way in: write a scene, convert a model, or ask Claude for one. The labs were mostly quiet. The community shipped a visual style, and it shipped it as tools you can keep.
New models
Robin: an open voice model you train on your own recordings, at film-standard 48 kHz. Fixed-Seed/Robin (code repo created November 15; the Hugging Face weights repo was created November 9 per the HF API and had 3 likes and 1 download when checked, card refreshed November 16) is a 3.58-billion-parameter speech model for English and Japanese, or both in one line. It does three jobs: design a voice from a text description, re-perform one of your own takes (up to 45 seconds) in another voice while keeping your words, timing and delivery, and train a dedicated add-on that always produces one exact voice. The README opens with "Zero-shot cloning is never good enough for studio work" and recommends at least 15 minutes of clean speech and says its test voices averaged 20 minutes in about 200 clips. It ships with ComfyUI nodes, two drag-in workflows and a browser app. Memory decides whether it runs on your card: the README lists peaks of 8.2 GB to speak, 9.1 GB to change a voice and 9.9 GB to train, so a 12 GB card clears speaking and voice changing on paper; the project page says training "fits on a single sixteen-gigabyte graphics card", so budget 16 GB to train. The license is the Fixed Seed Open License, Apache 2.0 plus a revenue cap. Commercial use, including of the audio you make, is allowed until your "Attributable Revenue" passes US$2 million in any 12 months, and that counts the full gross revenue of any product, service or content made with Robin, not a share. Past it you must tell Fixed Seed within 30 days, and commercial rights end 90 days later unless you sign a separate deal. Personal, research and evaluation use is uncapped. Add-ons you train count as derivative works under the same terms. The project page, audiorobin.io, offers free test inference through a hosted Lattice demo. The catch: no published listening tests, only two languages, and the voice changer needs an exact transcript of your take.
Video
Seedance 2.5 Flash arrived on Runware on November 15. The Runware changelog calls it "a lower-cost variant of ByteDance's Seedance 2.5 video model": 4 to 30 second clips with synchronized audio at 480p or 720p, in six aspect ratios from 21:9 to 9:16, under the ID bytedance:seedance@2.5-flash. The changelog gives no price, so check Runware's model pricing before you plan a batch. Thirty seconds of drafting with sound in one call is the useful part for animatics.
video-short-animation is a cartoon studio written as code, with no AI video model. DuongVuIT/video-short-animation (created November 15, 17 stars) renders flat 2D characters in a painted three.js world inside Remotion, with local text-to-speech timed to the word, synthesized music and effects, automated frame checks, and one story cut as a 60 to 90 second vertical short plus a 5 to 8 minute widescreen version. Every frame is TypeScript, so you fix one line and re-render one scene. Docs are in Vietnamese; the README says MIT, while GitHub's license detector reports "Other", so read the license file before reuse.
Audio and music
Haru-4B gives a character a voice that talks back in real time. higurashi-ai/Haru-4B (HF createdAt November 12, Apache 2.0, nine languages) is a full-duplex speech model: "A text card sets the character, a short reference clip sets the voice," and it listens and answers at the same time. That makes it a candidate for interactive installations and game characters. It needs about 12 GB of graphics memory on NVIDIA; the repo also has a Mac build. The card calls it "a pretrained base model, released for research," so expect rough edges. There is a hosted demo Space. Use only voices you have rights to.
Suno Bridge puts Suno inside Ableton. Suno's blog dated it November 13; the Bridge page describes a standalone Mac app that connects to Ableton Live 12 and generates instrument and vocal stems matched to your session's key and tempo, placed on the timeline. It is a beta, free to download, and requires a Premier subscription; generations spend Suno credits. Other DAWs are planned without a date, and it is an app beside your DAW, not a plugin.
Open and local
The weekend's local story is a style becoming infrastructure. Four repositories we found, created November 15 and 16, build the same look from a recipe credited to Silicon Jungle (@JungleSilicon, post not reachable from our tools, so the original date is unconfirmed): signed distance shapes (3D forms described by math instead of polygons), covered in Gaussian splats (millions of tiny soft 3D blobs, the format phone scans use), then drawn through a low-resolution pass with a small palette and dithering. Everything runs in a browser on WebGL2. Star counts below are GitHub API totals, checked this morning.
- cobanov/pixelart-gaussiansplat: converts any glTF or GLB model, rigged ones with their animations, into pixel-art splats, and builds a little island under it. Created November 15, 24 stars, MIT (repo).
- patch-gg/pixel-splats: five live dioramas in one dependency-free HTML file, plus a Claude Code skill that models, renders three angles, looks and fixes, and a tool that films a scene to MP4. Created November 15, 2 stars, Apache 2.0 (repo, demo).
- iamvillain/SyntaxSplat: scenes written as code with baked lighting, click-to-dissolve parts, a soft or pixel look, and standard
.plyexport for other splat tools. Created November 15, 25 stars, MIT (repo). - siliconjungle/shapeshift-labs-ink-studio: the recipe author's own toolkit for sculpting, animating, physics and headless rendering of pixel-art splats from Node. Created November 16, 30 stars, marked
UNLICENSED, so look, don't build on it (repo). - Fixed-Seed/Robin: train a 48 kHz voice add-on from 15 to 20 minutes of your recordings, inside ComfyUI or a browser app. Created November 15, 6 stars, Fixed Seed Open License (repo).
- benjiyaya/ComfyUI-H3Image-LayoutKit: draw boxes on a canvas inside ComfyUI and get a placement prompt MiniMax H3 Image follows, no language model required. Created November 15, 4 stars, MIT, marked in progress (repo).
- n4ze3m/tiny-voca: Saturday's 18 MB browser narrator, on its second day. Created November 14, now 471 stars, up from 110 in Sunday's briefing, Apache 2.0 (repo).
- nomadoor/ComfyUI-Smart-Guides: snapping and even spacing for ComfyUI nodes. Created November 14, 37 stars, MIT (repo).
Creative workflows
1. Turn any 3D model into an animated pixel-art diorama (cobanov/pixelart-gaussiansplat, docs/bring-your-own-model.md, tools/convert-model.mjs, tools/build-island.mjs)
The steps: Clone the repo, npm install, npm run dev (Node 20.19 or newer, or 22.12 or newer on the 22 line). Run npm run convert path/to/model.glb mymodel, then optionally npm run island mymodel ruins (themes: ruins, desert, swamp, crag, dais). Open /?m=mymodel. Tune in the address bar: px=4 pixel size, colors=24 palette size, bands=5 light steps, anim=run for a clip. For video, use the panel's record button, or &capture=1 and window.viewer.renderAt(seconds) for frame-exact output.
How it works: The converter samples your model's surface and base colors into about 110,000 flat discs. Each disc carries its surface direction so it can be lit live. The viewer draws them hard-edged at low resolution, snaps every pixel to about 15 colors and dithers between neighbors. If the model has a skeleton, its clips play; keys 1 to 9 switch them.
Why it is good: A model from any source, including an image-to-3D generator, becomes a consistent retro asset with real camera moves. The pixel grid is fixed to the screen, so zooming never changes pixel size.
Where it breaks: It ignores normal maps, metalness and glass, and lighting comes from the mesh alone. Rigs top out at 256 joints with four weights per point, bones must not scale during animation, and blend shapes (morph targets) are ignored. Raising detail past the defaults makes surfaces look sandy, per the docs. Use an even pixel size for video, since odd sizes blur after compression.
2. Ask Claude for a pixel-art diorama and film it (patch-gg/pixel-splats, skills/pixel-splats/REFERENCE.md, tools/shot.mjs, tools/film.mjs)
The steps: In Claude Code, /plugin marketplace add patch-gg/pixel-splats then /plugin install pixel-splats@pixel-splats. Ask: "Make a diorama of a witch's hut in a swamp, with fireflies and a bubbling cauldron." Check it with node tools/shot.mjs <scene> --angles 3, and film it with node tools/film.mjs <scene> --seconds 8 for a 1920x1080 MP4 (or --out out/scene.gif).
How it works: Each scene is a few dozen math shapes in plain JavaScript (the fox is 14), covered in 35,000 to 50,000 splats, lit with a shadow map drawn from the splats themselves, then rendered about 216 pixels tall with outlines on depth edges. The skill loops: model, render three angles, look, fix the worst thing, repeat.
Why it is good: One HTML file, no dependencies, and an agent that checks its own pictures before showing you.
Where it breaks: Motion is rigid poses plus four preset deformations (sway, water, swim, breathe), so no character acting. The tools need Node 22 and Chrome, filming needs ffmpeg, and publishing to Patch needs its desktop app on Windows.
3. Train a voice add-on from your own recordings (Fixed-Seed/Robin, comfyui-workflow-examples/Robin - LoRA Voice Training Official.json)
The steps: Write a script, then record it line by line: one or two sentences per clip, 3 to 15 seconds, one quiet dead room, same mic and distance, peaks around minus 6 dBFS, WAV or FLAC at 44.1 or 48 kHz. Save each clip with a .txt of exactly what it says, numbers written as spoken. Run python -m robin prep my_voice/, then python -m robin train my_voice/ --model . --out my_voice_lora, and speak with --lora my_voice_lora/lora.safetensors. In ComfyUI, use Robin Check Dataset then Robin Train LoRA.
How it works: Training writes a small add-on file for one voice. A checkpoint is saved every 200 steps and reads a sample line in the new voice, so you pick the best step by ear.
Why it is good: The prep step lists clips it will drop and flags clipping, digits, low sample rates and mismatched transcripts before you spend hours.
Where it breaks: You need about 10 GB of graphics memory (the project page names a 16 GB card for training) and 15 to 20 minutes of clean speech. A calm read trains a calm voice, so record the delivery you want back. Train only on voices you have rights to.
Worth testing
- pixel-splats demo at patch.gg. Tradeoff: our fetch saw only "Loading the demo…", so rendering is unconfirmed outside a real browser.
- Robin on Lattice (link). Tradeoff: sign-in terms are unclear, and a hosted demo won't show you training, the part that matters.
- Haru-4B demo Space (link). Tradeoff: a research base model; latency and character consistency will vary.
- cobanov's converter on a model you already own. Tradeoff: example creatures in the repo come from a game and are not MIT licensed, so don't reuse those.
- Suno Bridge if you are on Ableton Live 12 and Premier. Tradeoff: Mac only, beta, and every stem spends credits.
What actually matters from today's signal
The pixel-splat weekend is the clearest example yet of how a look spreads now. Someone posts a recipe. Within a day, three strangers have written it from scratch, each picking a different entrance: code, model conversion, an agent skill. For a creator, that means an aesthetic no longer has to be bought as a filter or a subscription. You can own the whole machine, change its palette and film it at any length. The cost is time in a terminal, not money.
Robin points the same way for voice. Its pitch is an afternoon of recording in exchange for an exact voice file you keep, which is slower than ten-second cloning and more honest about what studio work needs. Read the license closely, though. The $2 million cap counts the whole gross revenue of anything made with it, so a successful game or channel can cross it faster than the number suggests, and a voice add-on you train stays under Fixed Seed's terms.
The counter-signal: almost everything today is days old, single-maintainer and barely starred. Robin's weights had one download when we checked. The splat tools have no rights story for converted assets beyond whatever you bring. Treat this week's toys as sketchbooks, not production pipelines, until someone ships a real project with them.
Source access notes: Adversarial fact-check pass ran (Sonnet subagent). It caught: the Robin license summary understated the revenue test (the cap counts full gross revenue of anything made with Robin, with a 30-day notice and 90-day wind-down), now fixed; a framing that called the README's opening line the team's motive, reworded; the video-short-animation license note, softened; the Node version requirement, completed; ink-studio stars updated to 30. Its tiny-voca star doubt was re-checked against the GitHub API (471). SyntaxSplat's README does not credit Silicon Jungle; cobanov and patch-gg do. Hugging Face (site and API), Hacker News Algolia, shields.io and most vendor sites were blocked from both of our shells this run; Hugging Face API data (Robin createdAt 2026-11-09T20:57:27Z, lastModified 2026-11-16T00:43:08Z; Haru-4B createdAt 2026-11-12T11:05Z) and vendor pages were read through a fetch agent. GitHub repo metadata and star totals come from the GitHub API (shields.io blocked); READMEs and the Robin LICENSE were read from raw.githubusercontent.com with cache busting. No HF model in the text-to-video, image-to-video, text-to-image, text-to-speech, text-to-audio or image-to-3d lists was created on or after November 13 with meaningful traction. x.com was blocked, so the Silicon Jungle post is unverified. OpenAI news failed to load; Runway's news page loaded without dates; blog.comfy.org and PetaPixel loaded without usable dated posts. Adobe, Krea, ElevenLabs, Replicate, fal, BFL, Stability and DeepMind had nothing in the window. Luma's November 13 to 14 posts were listicles except the November 14 board organizer, covered yesterday. Two GitHub repos posing as "desktop" versions of commercial AI apps were excluded as likely lures. MIRAGE (OQ-J, created November 15) was excluded because its author says the posted weights produce unsatisfactory results.