Creative AI Briefing: Friday, November 20, 2026
Paste a link to an 85-minute YouTube stream into a free app on your own PC, and about a minute and a half later you have three vertical, captioned clips with a title and a hook on each. That is NeonClipping, and it picked up roughly 790 GitHub stars in its first day. It is the loudest example of a wider shift this week: the jobs creators have been renting by the month (clipping, sticker-style app animations, pixel-art cleanup, demo videos) are turning into free tools you install once and own. The big labs were quiet. The community shipped the plumbing.
New models
Kling Image 4.0 makes sets of up to ten matching images, with real text in them. Kling announced it on November 17 (press release) and in a blog post dated November 18 on the page. Our earlier briefings missed it, so it leads today. What you can make: a storyboard, a product page or a slide set where up to ten images keep the same subject and style from one prompt, plus posters with up to 2,000 characters of text, more than 70 preset fonts and the option to upload a picture of your own font as a reference. Text works in seven languages, including English, Japanese, Korean and Spanish. Kling says native 8K output is live and that the press-release version lets you edit text and images on separate layers (the blog does not mention layers). It is "now live on Kling Creative Studio and available via API to enterprises and developers," and the blog adds that it "is now available to all users." The catch: neither Kling page states a price or commercial-use terms. An API-pricing review on Aireiter's blog puts third-party gateway rates at about $0.063 per image at 2K, $0.094 at 4K and $0.157 at 8K, but those are third-party rates, not Kling's own credit price. The leaderboard claim appears only in Kling's paid press release.
A four-step add-on makes MOSS-Image roughly fourteen times cheaper per picture. lvladikov/MOSS-SFT-Distill-4step-LoRA (Hugging Face createdAt 2026-11-19 08:45 UTC, Apache 2.0) is a small add-on file that teaches the open MOSS-Image model to finish a picture in 4 passes instead of 50. The author measured 22 images at 1024x1024 in 26.6 seconds of billed GPU time versus 387.3 seconds without it, on Comfy Cloud. Those are the author's own figures. It ships eight ComfyUI workflows (text-to-image, layout, image edit and the 4K PiD decode), each in local and Comfy Cloud versions. The catch: it is a community project, not endorsed by OpenMOSS, and the card gives no memory figure for running it, only a training peak on a data-centre card.
Image
Pixagon puts an AI pixel-art assistant inside Aseprite. blkdmr/pixagon (Apache 2.0, created November 18) generates a sprite from a prompt, edits one by instruction ("make the sword glow blue"), redraws only a selected region, or turns a character to a new angle for turnarounds. Results land as a new layer, so your own pixels are never touched. Locally it runs FLUX.2 Klein 4B on an NVIDIA card with 12 GB or more of memory (8 GB works in a slower mode) or an Apple Silicon Mac, with a 15 GB download. Without a GPU you can route through OpenRouter on your own bill, which sends your sprite to a cloud provider. The author says local runs on Windows and macOS still need more real-hardware testing.
Pixel Studio snaps any image model's "pixel art" to a real pixel grid. alanfrigo/pixel-studio (MIT, created November 19) is a self-hosted web app that finds the fake grid inside a 1024-pixel AI image, recovers the art at its true size, strips the fake checkerboard background and exports sprites, icons, tilesets and UI as a package a game can read. A built-in simulated provider lets you try every step without an API key.
Bria's RMBG-3.0 cuts out hair, mesh and fine holes, but the free weights are non-commercial. briaai/RMBG-3.0 (createdAt 2026-11-10) treats background removal as an image edit and returns a soft alpha matte. Its own card lists glass, nets, veils and smoke as weak spots that tend to come out opaque. It needs 32 GB or more of GPU memory, and the open weights are CC BY-NC 4.0; commercial use goes through Bria's API or fal. Try it free in the official Space.
Video
NeonClipping turns long streams into vertical clips, locally and free. iamajmeena/neonclipping (MIT, created November 19) is built by a Jaipur-based colorist and Resolve educator. Fast mode downloads only the audio, borrows YouTube's own word-timed captions, scores loudness jumps and raised voices, has a small local language model pick and title the best moments, then downloads only those sections. The author's timings on an RTX 2060: 97 seconds for three clips from an 85-minute stream, 136 seconds from a three-hour Twitch VOD. It has 11 layouts for facecam and gameplay, five caption styles, Hinglish captions, live clipping on chat spikes, and a five-star rating that re-weights its picks to your taste after eight ratings. Honest limits from the README: it finds loud and quotable, not guaranteed viral, and its language model reads text only, never the screen. Our caution, not the README's: check you have the rights before clipping someone else's stream.
Two Claude Code skills record and narrate app demos and edit talking-head Reels. OrrrGE/autoRecording (MIT, created November 19) drives a real browser through a web app from one Python "story file," renders 1080p60 with camera moves and captions, and adds optional local narration. Liamrjohnston/ai-video-editor (MIT, created November 19) splits a Reel edit into five skills (cut, style for layouts, captions, sound-effects, motion-graphics) cuts first, then stops to show you a table of which layout each line gets before it builds graphics, captions or sound.
motlab is a code-drawn video engine with a review viewer. staryone/motlab (MIT, created November 16, 237 stars) lets you click a frame, drag a box or mark a time range and type a note, which the agent then works through by number. The author reports it rendering 2 to 4 times as fast as Remotion and HyperFrames on a GPU, on one benchmark scene on one laptop; in software mode it was slower.
Audio and music
Inworld open-sourced a 48 kHz voice isolator you can use commercially. inworld/voice-isolator (createdAt 2026-11-14, Apache 2.0) pulls a voice out of a noisy recording; subtract it from the original and you also get the backing track or room noise on its own. A realtime version adds 63 ms of algorithmic delay at its 20 ms look-ahead setting, small enough for live use. This is second-day news, six days old, and we had not covered it. It is not a speaker selector: with two voices it keeps the louder one. Try it in the community fffiloni Space.
Grey is a free ten-language voice model with a Windows app, for non-commercial work. Randy-H0/Grey-tts (created November 18) speaks ten languages, clones a voice from a short recording, does voice-to-voice, and trains your own voice on cards from 6 GB in its offline Grey Studio app. You can try it free online. The license file opens with "Free for non-commercial use," so read it before a paid job.
Pratidhvani builds consent-first voice clones for Indian archives. svayatta/pratidhvani (Apache 2.0 code, created November 19) aims at families, schools and archives with old, noisy tapes, keeps accent and dialect, and watermarks every generated clip so its own detect command can find it.
Open and local
The quieter local story is about who owns the setup. On November 18 Comfy launched Custom Builds for Comfy Cloud: your own custom nodes, models and add-ons packaged once and run in the browser on a rented GPU at $1.75 to $11.23 an hour, included in every paid plan from Standard up. GPU time keeps billing during a keep-alive window after the last job, so stop workers you are done with. For a team whose ComfyUI only worked on one person's tower, that is the difference between "send me the file" and "open the link."
Trending and new repos (stars from the GitHub search API this morning):
- iamajmeena/neonclipping: long video or live stream in, captioned 9:16 clips out, all on your PC. About 790 stars in its first day (repo).
- staryone/motlab: write video scenes as code with an agent, review and annotate frames in a viewer. 237 stars, created November 16 (repo).
- Nyxera0x/simulator: a Claude Code skill that rebuilds a factory, airport or worksite as a toy-scale animated simulation in the format of a reference video. 28 stars, created November 18 (repo).
- OrrrGE/autoRecording: narrated web-app demo videos from one story file. 14 stars, created November 19 (repo).
- blkdmr/pixagon: AI sprite generation and region edits inside Aseprite. 2 stars, created November 18 (repo).
- nathankim0/image-to-lottie: AI stills to Lottie app animations, checked frame by frame. 1 star, created November 19 (repo).
- danchar12138/scan3d-windows: photos or phone video to a Gaussian splat, a textured mesh and a collision mesh on one Windows PC with an NVIDIA card. 1 star, created November 19 (repo).
- remorses/splat-paint: repaints any 3D splat scan as an oil painting of 3D brush strokes, in the browser. 1 star, created November 19 (repo).
Creative workflows
1. Turn three AI stills into an app-ready "gift opens" animation. Repo: nathankim0/image-to-lottie, example folder examples/gift-open/ with config.json, and the rules in skills/image-to-lottie/references/pitfalls.md.
The steps: Generate the base, lid and reward on one transparent sheet in a single request. Run image-to-lottie split sheet.png -o parts --names base,lid,item. Run fit on the base and lid, open fit.png, and choose the smallest overlap where the opening is fully hidden (the example settled on 182). Put that number in config.json with the gift-open preset and run build with --lottie for a single .lottie file. Run check, then render --frames ... --sheet and look at the contact sheet.
How it works: Product teams often animate 3D-looking rewards by moving a few flat pictures with simple transforms. The tool does that part for you and writes Lottie, the animation format apps and websites already play. Presets cover opening boxes, hops, floats, pop-ins, confetti poppers and flipbook frames.
Why it is good: The checks are real. check warns when anything leaves the canvas or an image is far bigger than it displays, and render pins exact frames, so you see the 0.3-second burst a screen recording would miss.
Where it breaks: Parts generated one at a time do not match, so a bad sheet means regenerating, not fixing. Motion a transform cannot fake (turning, morphing) needs frames, which grow the file. It is one day old, with one star.
2. Clip a long stream into shorts without uploading it anywhere. Repo: iamajmeena/neonclipping, Windows installer NeonClipping-Setup.bat from Releases, or run.ps1 / run.sh.
The steps: Install Python 3.10+ and FFmpeg (the Windows setup file does this), plus Ollama with ollama pull llama3.2:3b for titles and scores. Paste a link, set the clip length range (say 30 to 80 seconds), pick a facecam preset or draw the boxes, choose a caption style, press Make clips. Rate each clip with stars.
How it works: It listens before it looks. Audio jumps, shouting-range energy and key phrases shortlist moments; the local model reads each transcript and writes a title and hook; only the chosen sections get downloaded and rendered.
Why it is good: By the author's timing it handles a long video in under two and a half minutes on a 2019-era gaming card, and edits re-render without re-transcribing.
Where it breaks: Quiet, visual moments get missed because the model never sees the picture. Without YouTube captions, local transcription is slow on CPU. Deep mode was verified on a short video only.
3. Draft MOSS-Image pictures in four steps in ComfyUI. Model card and workflows: lvladikov/MOSS-SFT-Distill-4step-LoRA.
The steps: Put moss_sft_4step_rank_64_lora_comfyui.safetensors in ComfyUI/models/loras at strength 1.0. Load the MOSS-Image model, text encoder and small FLUX.2 decoder from Comfy-Org's MOSS-Image repo. Use the provided workflow, which sets four manual noise levels and guidance at 1. Skip the original negative prompt.
How it works: The add-on was trained so the model lands near its 50-pass result in 4 passes, so drafts arrive while you still remember what you asked for.
Why it is good: Iteration gets cheap enough to try ten compositions where you used to try one, and the same workflows run on Comfy Cloud if your card is small.
Where it breaks: Large decodes can fail on Macs (the card suggests tiled decode or --cpu-vae). The speed figure is the author's own, on cloud hardware. Fast versions usually trade some fine detail, so finish important frames with the full model.
Worth testing
- splat-paint live site: drop your own
.plyscan and watch it repainted as oil strokes, file stays in your browser. Tradeoff: no license file in the repo yet, so treat output as personal. - RMBG-3.0 Space: test the hardest hair or fine-mesh cutout you have. Tradeoff: glass and veils are a stated weak spot, and the free weights are non-commercial and need 32 GB to run locally.
- Grey online: ten languages and voice-to-voice in a browser. Tradeoff: non-commercial license.
- Kling Image 4.0 in Creative Studio: try a ten-image storyboard set or a text-heavy poster with your own font reference. Tradeoff: no published price or usage terms on Kling's own pages.
- voice-isolator Space: clean a noisy interview take. Tradeoff: a community Space, and with two people talking it keeps the louder one.
What actually matters from today's signal
Clipping services built a business on one honest insight: long videos hide short ones, and nobody wants to scrub three hours to find them. NeonClipping shows that insight no longer needs a subscription. Audio-first scoring, borrowed captions and a three-billion-parameter local model do most of the job on a 2060. If you clip a weekly podcast or stream, the paid tool now has to win on polish and taste, not on access.
The same shape shows up in image-to-lottie, Pixel Studio and Pixagon. None of them make better art than the models underneath. They do the boring, checkable part: cut the parts apart, find the true pixel grid, keep your existing layer untouched, pin the frame where the burst happens. That is the step where hours disappear, and it is now free and repeatable.
The counter-signal: almost every one of these tools is days old with a handful of stars, and the most capable hosted launch of the week, Kling Image 4.0, arrived without a public price or plain usage terms. Local tools give you ownership and leave you as the support desk. Hosted ones give you polish and keep the fine print. Pick per job, and read the license before the client does.
Source access notes: Hugging Face dates come from the API createdAt field, read through a fetch agent (Hugging Face and img.shields.io were not reachable from the shell this run, so star counts come from the GitHub search API instead of shields.io). Repository details come from raw.githubusercontent.com README and LICENSE files with cache-busting. openai.com/news and nocode.mba returned errors; Runway, Ideogram and Google's AI blog showed no day-level dates; Creative Bloq items were undated; civitai and Midjourney were not attempted. BFL, Stability, Luma, fal, Replicate, ElevenLabs and Krea had nothing dated November 18 to 20. Rumble-2-Flash got a Hugging Face blog post on November 18, but the model was created October 14 and only transcribes speech, so it is omitted. Writing section omitted (nothing creator-facing). Adversarial fact-check pass ran (Sonnet subagent): it caught an unsourced three-minute idle-billing figure (removed), RMBG-3.0's glass claim (its card lists glass as a weak spot), the Aireiter prices described as a reseller's own (reframed as third-party rates), the ai-video-editor order of steps and skill names, motlab's GPU-only speed scope, Pixagon's untested local platforms, voice-isolator latency scope and age, and a rights caution wrongly implied to come from the NeonClipping README. All fixed.