FervorCreative AI
Live Latest 21.11.26 · morning 422 tools tracked 1040 workflows indexed 651 topics Hot: Claude Code, Gaussian splatting, ComfyUI

The new tools stop at the second take, so creators can repair or remake one part of a finished piece (a far face, a single sentence, a sprite angle) without regenerating the whole thing.

AURAKyutaisprite-stackingH3 Face FixEchoLocalAnimated SOGMiniMax H3avatar-genvideo-genvoice-clonecomfyuilocal-creative-aicreative-workflows

Creative AI Briefing: Saturday, November 21, 2026

Draw a car four times, from the top, side, front and back, and an agent skill released yesterday carves those four drawings into a pixel-art car you can spin, drive around a rainy neon street and export as a 16-direction sprite sheet. That is sprite-stacking, and it passed 240 GitHub stars in its first day. It fits the pattern of the week's best releases: they work on the part you already have. Kyutai's new AURA model animates a portrait from a voice as the voice arrives, a ComfyUI workflow repairs only the melted faces in a finished MiniMax H3 clip and keeps the sound, and a local voice tool flags the exact words in a dubbed take you should listen to again. Less rerolling, more fixing.

New models

AURA turns one portrait and a live voice into a talking head as you speak. Kyutai published AURA in a blog post dated November 20; the two model files appeared on Hugging Face on November 18 (API createdAt). What you can make: a face for a voice character, a streamer avatar or a narrator that moves while the audio is still coming in, because each short chunk of video needs only its own slice of audio, with no look-ahead. Memory stays flat however long the stream runs. AURA-1.3B outputs 576p at 16 frames per second and, by the authors' own measurement with an optimized pipeline, runs at 20.7 fps on one H100 (the public code manages about 16 fps, still real time for its playback rate). AURA-5B reaches 720p at 25 fps but runs at 9.7 fps, so it is not real time. The catch is the license: the 1.3B weights are CC BY-NC 4.0 and the 5B weights use Snap's non-commercial research license, so neither belongs in paid client work. The 5B card says it was trained and evaluated only on English speech, and the code needs a CUDA card; the authors benchmark on a data-center H100. Free hosted demo: AURA x Unmute lets you talk to Kyutai's voice assistant with an AURA face at 480px (the page warns the face model can take about ten minutes to start).

Image

sprite-stacking makes fake-3D pixel art from four flat views. danielwh2/sprite-stacking (MIT, created November 20) is an agent skill for Claude Code, Codex, Cursor, OpenCode, Gemini CLI and Pi. Give it a sheet with top, side, front and back views; it keeps only the blocks all four views agree on, slices the result into flat layers and stacks them with a one-pixel lift, the old indie-game trick behind spinning cars and trees. A standalone HTML viewer adds turntable, day-to-night lighting, rain and snow, and exports a 16-direction PNG sheet for your engine. No image generator? The agent draws the views itself in Python. The catch: shapes that hide detail from all four sides (for example an overhang or a hollow) carve wrong, and the author credits the technique to prior art; the new part is the four-view carving.

H3 Image EasyFix cleans the grainy skin and fur in MiniMax H3 pictures. PedroMarinhoDev/ComfyUI-H3-Image-EasyFix (Apache 2.0 code, created November 19) adds a "Steering" node and a re-noise sampler that make H3 Image redraw fine texture instead of keeping speckle, plus a size picker above ComfyUI's 16-megapixel cap. Side-by-side crops on the repo show the difference at full size. The catch is H3 itself: its community license, as summarized on the model card we covered November 17, excludes users in the EU, UK, South Korea and the USA.

Video

H3 Face Fix repairs far-away faces in a finished H3 clip, audio untouched. Pixelailabs/ComfyUI-H3-FaceFix-PixelAILabs (MIT, created November 19) finds the face, redraws it large from your character photo and pastes only the face back. Its one new node fades the fix by face size: faces 110 pixels tall or smaller in 1080p get the full repair and faces 180 pixels and up stay exactly as H3 drew them (the thresholds scale with video height), so a push-in does not grow new wrinkles at the close end. The authors time it at about four minutes per ten seconds of 1080p on a 32 GB RTX 5090. The catch: you need H3 already set up (about 34 GB for the model file alone, plus a 27 GB text encoder), one face per run, and the same territory license as above.

H3-Tiny is a slimmed-down MiniMax H3 video model, untested. H3-Tiny/H3-Tiny-17B-Pruned-FL2VA (created November 20) is a pruned copy of H3's first-and-last-frame and text-to-video model at about 17 billion parameters. Second-day work, unaffiliated with MiniMax, no quality numbers yet, and it still needs the official H3 text encoder and decoders. Watch, do not build on it.

Audio and music

EchoLocal clones your voice into 23 languages offline and checks every line. Ares9323/EchoLocal (created November 19) wraps Resemble AI's Chatterbox Multilingual for cloning, Meta's NLLB-200 for draft translation and Whisper for subtitles and a read-back check. Write a script once, tick languages, and get one file per language plus a report listing words Whisper heard differently. About 8 GB of graphics memory is enough on an NVIDIA card; Apple Silicon works; CPU takes about a minute per sentence. The catch: EchoLocal's own code is PolyForm Noncommercial and NLLB-200 is CC BY-NC 4.0, so this is a personal and testing tool as shipped. Chatterbox embeds an inaudible watermark in every clip.

Open and local

The local story today is repair kits for models that already exist. MiniMax H3 launched November 3; this week the community shipped a texture fix, a face fix and a smaller copy for it, all inside ComfyUI. Two releases point somewhere new: AURA brings real-time avatars to open (if non-commercial) weights, and Animated SOG packs moving 3D scans with sound into a file a browser can stream.

Trending and new repos (stars from the GitHub search API this morning):

  • danielwh2/sprite-stacking: four pixel-art views in, a spinnable fake-3D sprite and 16-direction sheet out. 241 stars, created November 20 (repo).
  • CryptoYuyyy02/stickman-video-director: a Chinese-language agent skill that plans, voices, scores and edits stickman story shorts scene by scene, adapted from kaomei's MIT original. 80 stars, created November 20 (repo).
  • PedroMarinhoDev/ComfyUI-H3-Image-EasyFix: cleaner fine texture from H3 Image. 46 stars, created November 19 (repo).
  • arrival-space/animated-sog: a format for moving Gaussian-splat clips (volumetric video) with sound that streams in PlayCanvas and three.js; the authors turn a 3.6 GB folder of 180 frames into a 12.4 MB file. 34 stars, created November 20 (repo).
  • Pixelailabs/ComfyUI-H3-FaceFix-PixelAILabs: fix melted far faces in H3 video, original audio kept. 27 stars, created November 19 (repo).
  • kyutai-labs/aura: code for real-time portrait-plus-voice avatars. 21 stars, created November 19 (repo).
  • Ares9323/EchoLocal: offline voice cloning in 23 languages with a Whisper check. 4 stars, created November 19 (repo).

Creative workflows

1. Turn four drawings into a spinning game sprite. Repo: danielwh2/sprite-stacking, script skills/sprite-stacking/scripts/views_to_slices.py, viewer viewer.html.

The steps: Install with npx skills add danielwh2/sprite-stacking (or copy the folder into ~/.claude/skills/). Draw or generate one row of four orthographic views: top, side, front, back, same scale. Run python3 views_to_slices.py views.png slices.png --slices 36 --preview check.png --html viewer.html. Use 15 to 20 slices for chunky, 32 to 48 for detailed; add --round for pots and towers and --emissive for headlights and windows. Open the viewer, spin it, then export the 16-direction sheet.

How it works: Think of four cookie cutters pressed through a block of clay from four sides. Only clay all four cutters leave behind survives. The tool then slices that block into thin layers and draws them stacked, each rotated and nudged up a pixel, so the eye reads depth.

Why it is good: You stay in 2D. Four drawings replace a 3D model, and the output is plain PNG strips any engine can draw.

Where it breaks: Hollows and overhangs that none of the four views reveal get filled in. Views drawn at different scales will not line up. Day-old repo.

2. Repair the small, melted faces in an AI video without regenerating it. Repo: Pixelailabs/ComfyUI-H3-FaceFix-PixelAILabs, workflow example_workflows/PixelAI_H3_FaceFix.json.

The steps: Install the node plus the listed packs (Carasibana's H3-FaceRefine, MaskVidExperiments, VOSR2, KJNodes, VideoHelperSuite, Impact-Subpack) on ComfyUI 0.45.0 or newer. Load your finished H3 clip and a clear photo or character sheet of the face. Choose largest_face or smallest_face and a face number. Run. The result lands in ComfyUI/output/ as H3_FaceFix_*.mp4 with the original sound.

How it works: It tracks the face, cuts a big square around it, masks only the face, has H3 redraw that square at 768 pixels from your photo with no prompt, polishes it and stitches it back. The size fade decides how much of the fix each frame gets.

Why it is good: The take you liked survives. Only the broken part changes, and close-ups stay untouched.

Where it breaks: One person per run, heavy downloads, an H3-only pipeline, and the H3 license territory exclusions.

3. Dub one script into several languages and know which lines to re-listen to. Repo: Ares9323/EchoLocal.

The steps: Run setup.ps1 (Windows) or setup.sh (Mac). Put 5 to 15 seconds of clean speech in voices/. Open the web interface, write the script one sentence per line, tick languages, press Translate, read and fix each language box, then Generate. Open check_report.txt and listen to the flagged words first.

How it works: Each line is generated on its own because the voice model degrades on long text, then the lines are joined. Whisper transcribes every clip back and compares it to the script.

Why it is good: The check turns "listen to 23 files" into "listen to these six words."

Where it breaks: Non-commercial as shipped; translation is a draft; Whisper mishears rare words on its own; the room and microphone of your sample end up in the clone.

Worth testing

  • AURA x Unmute live demo: talk to a voice assistant with a real-time AURA face. Tradeoff: 480px, slow cold start, non-commercial weights.
  • Animated SOG PlayCanvas viewer: orbit a moving volumetric clip in your browser. Tradeoff: making your own clips needs a CUDA GPU and an external trainer under the AGPL license.
  • sprite-stacking: four views, one spinning sprite, in minutes. Tradeoff: needs Python, Pillow and NumPy, and an agent if you want it to draw the views.
  • H3 Image EasyFix: compare the 1:1 crops before you install. Tradeoff: only matters if H3's license covers where you live.

What actually matters from today's signal

For two years the default fix for a bad AI result was another roll of the dice. This week's tools reject that. Face Fix keeps your take and repairs a 47-pixel face. EchoLocal keeps the 22 good sentences and points at the one to redo. sprite-stacking keeps your drawings and builds the volume from them. That is how editors and illustrators already work, and it saves the thing rerolling burns: a take you were happy with.

AURA matters for a different reason. Real-time talking heads have mostly lived inside paid avatar services. Kyutai published weights, code and a paper that make one run as fast as speech on a single data-center card. You cannot sell work made with it, but every hosted avatar product now competes with a public recipe.

The counter-signal is licensing. Four of today's seven repos sit on a model or license that blocks paid use or whole countries: H3's territory clauses, AURA's non-commercial weights, EchoLocal's PolyForm terms. The free tools are good. Read the license before you show a client.


Source access notes: An adversarial fact-check pass (Sonnet subagent) ran on this draft and caught four issues, all fixed: AURA chunk length stated as half a second (now unstated), unverifiable H3-Tiny block and parameter figures (removed), an English-only claim extended to both AURA models (now scoped to the 5B card), and Face Fix threshold boundaries. Hugging Face dates come from the API createdAt field, read through a fetch agent; huggingface.co, img.shields.io, openai.com, blog.google and hn.algolia.com were not reachable from either shell this run, so star counts come from the GitHub search API, not shields.io. GitHub's per-repo API was not enabled for this scheduled run (only search worked); repository details come from raw.githubusercontent.com README and LICENSE files with cache-busting. The nocode.mba release tracker refused automated fetches. Lab checks (OpenAI, Google, Adobe, Runway, Luma, ElevenLabs, Suno, Midjourney) via WebSearch found nothing new dated November 19 to 21; Adobe's latest news posts are the November 10 MAX announcements already covered. The AURA blog body did not render for our fetcher; its date comes from the blog URL and page title. Trendshift, Civitai and blog.comfy.org were not checked.

Articles from this briefing