Creative AI Briefing: Saturday, November 14, 2026
You can now make a five-second, 832 by 480 video clip on a 2017 gaming laptop with a 4 GB GTX 1050 Ti, using the full Wan 2.2 5B model at its official 50 steps. It takes about two and a half hours, and the author proves with file hashes that none of the speed-ups changed a single pixel. That recipe landed on November 13 next to a local voice-cloning and video-dubbing app, a ZONOS3-Mini desktop app for Windows, Mac and Linux, and a free Gaussian splat trainer that turns a phone video of a room into a Blender set. The big labs were mostly between announcements. The pattern from everyone else: the hardware you already own is back in the game, and the price is paid in waiting time instead of a new card or a monthly plan.
New models
We found no new foundation model for image, video, audio or music from a major lab in the last 48 hours. The closest thing to lab news is a promise: at its November 11 office hours, Midjourney said V9 is aiming to ship "within 30 days, maybe sooner," with a new base model, a new upscaler, better text and a Thinking mode, and that V9 images "could cost double" (secondary, via Geeky Curiosity; nothing is final and nothing is available). Today's real releases are tools built on models that already exist, and they are below.
Image
Seedream layer splitting reached Runway's API, then got cheaper a day later. Runway's API changelog added Seedream 5 Pro Layers on November 10 at 18 credits per decomposition, and on November 11 switched it to per-image billing: 3 credits per image at 1K or 1.5K and 5 credits at 2K or auto (secondary, via the Releasebot mirror of Runway's changelog; Runway's own site was unreachable from this run). If you liked yesterday's Seedream 5.5 layer story, this is the same idea on the older Pro model inside a tool you may already pay for. The same catch applies: a split that returns many layers bills many images.
Video
wan22-pascal-4gb puts Wan 2.2 on a 4 GB card from 2017. alexturkey/wan22-pascal-4gb (created November 13, MIT scripts and patches) is a measured recipe plus four patches for stable-diffusion.cpp, a small C++ engine that runs image and video models without Python. On one GTX 1050 Ti laptop card it makes a 2-second clip (49 frames) in about 26 minutes and a 5-second clip (121 frames) in about 158 minutes, peaking at 2.4 to 2.7 GB of graphics memory and about 7.1 GB of system memory. The author's rule is the interesting part: a patch only goes in if the finished frames hash exactly like the unpatched engine's, so every speed-up is free in quality. Faster tricks that blur the picture (step caching, the distilled Turbo model, a lighter decoder) were tried and rejected, with numbers. Windows binaries are on the v0.1.0 release; models are about 10.9 GB. Wan 2.2 TI2V-5B weights are Apache 2.0, so commercial use is allowed. The catch, stated by the author: one prompt, one seed, one card, Windows only, and only the 5B model was measured. Search results point to a hosted try of the same model in a Wan-AI/Wan-2.2-5B Space; Hugging Face was unreachable from this run, so we could not open it.
ClipForge turns a long video into edited Shorts through Claude. lowdevco/ClipForge (created November 13, MIT) is an agent skill: give it a YouTube link or a file, get a scored shortlist of moments, pick the ones you want, and it renders vertical clips with a color grade, animated captions and, in its punchier styles, zooms on the big moments. The YouTube-link route needs Claude Code; in the chat app you hand it a file. Useful for podcasters and streamers who hand-cut clips today. Rights in the source video are yours to sort out, not the tool's.
Audio and music
Mimvox clones your own voice, reads books aloud and redubs video, all offline. tarpetos/mimvox (created November 12) is a desktop app for Windows and Linux, with an Apple Silicon build the README calls "best effort" and not validated. You read a guided text into your microphone (about 1 minute for an instant, untrained voice; about 6 to 7 minutes for the validated trained setup; 15 to 20 for the best match), and it speaks typed text in about 60 languages in your voice. The README describes the chain: OmniVoice speaks the words, then an RVC timbre model trained on your recording makes it sound like you. Book mode turns EPUB, DOCX, PDF and other files into one audio file per chapter. Redub translates a video, separates the music and effects from the old voice, keeps them, and lays your voice on top. It recommends 8 GB of graphics memory, and the runtime takes about 30 to 33 GB of disk (setup asks for 34 to 40 GB free). The catch is the license: the app code is PolyForm Noncommercial and the OmniVoice and NLLB translation weights are CC BY-NC 4.0, so this is for personal projects, not client work. The README also says plainly: "Early version: expect bugs."
ZONOS3-Mini got its second-day apps. Yesterday's free, word-steerable voice model now has a desktop front end. stepupgaming/booth (created November 12, Apache 2.0) offers preset voices with mood, pace and energy controls, several takes per line, whole scripts rendered line by line with a voice per line, cloning from a short clip, and a Claude Desktop extension. It runs on the processor on any modern Windows machine, with a faster engine for RTX 20 series and newer; Mac and Linux builds are listed as "new and not yet tested on real hardware," and the app is not code-signed yet. For Pinokio users, cocktailpeanut/zonos3-mini (created November 12, MIT) is a one-click launcher. Same rule as yesterday: clone only voices you have permission to use.
Open and local
The local story is patience. None of today's open releases asks you to buy anything. They ask you to wait: 26 minutes for two seconds of video on an old laptop, a few hours of voice training on a mid-range card, a few minutes of splat training per room. Star totals are GitHub API figures at run time (shields.io and Trendshift were unreachable, so no daily deltas).
- alexturkey/wan22-pascal-4gb: Wan 2.2 video on a 4 GB GTX 1050 Ti, every speed-up hash-checked to change no pixels. Created November 13, 4 stars, MIT (repo).
- voyager-so/vauss: phone video of a room in, an explorable 3D scene, a Blender pack and a walkable voxel game out. Created November 12, 7 stars, source-available Vauss Preview License (repo).
- tarpetos/mimvox: offline voice cloning, audiobooks and video redubbing in your own voice. Created November 12, 54 stars, noncommercial (repo).
- stepupgaming/booth: ZONOS3-Mini as a desktop app with takes, scripts and Claude support. Created November 12, 3 stars, Apache 2.0 (repo).
- lowdevco/Clip-Forge: long video to edited vertical Shorts through a Claude skill. Created November 13, 29 stars, MIT (repo).
- mileson/chaojifeng-video-content-creator: a Chinese-language agent skill covering script, storyboard, spoken-audio cleanup and 148 HTML animation effects. Created November 12, 74 stars, MIT (repo).
- cocktailpeanut/zonos3-mini: one-click Pinokio install for ZONOS3-Mini with a Voice Studio. Created November 12, 3 stars, MIT (repo).
Creative workflows
1. Make AI video on an old 4 GB graphics card (alexturkey/wan22-pascal-4gb, scripts/generate.py, scripts/verify_frames.py)
The steps: Download wan22-pascal-4gb-v0.1.0-win-x64-cuda12.8-sm61.zip and cudart-cuda12.8-win-x64.zip from the release and unpack both into one folder. Install the Microsoft Visual C++ runtime if you lack it. Download the roughly 10.9 GB of models listed in models.lock.json onto an NVMe drive. Run python scripts\generate.py with --frames 17 first (about a ten-minute check), then --frames 49 for two seconds. Run the five-second job detached, one job at a time, with the browser closed.
How it works: The model is shrunk to an 8-bit file, the text encoder runs on the processor (the README says processor and disk), and the final decode happens in small 128 by 128 tiles so the card never holds the whole frame at once. Four patches speed up the parts that were slow on this older chip, and each is kept only if the output frames match the unpatched engine byte for byte.
Why it is good: It is the full model at official settings, not a degraded "turbo" version, and the repo publishes what was tried and rejected in docs/TRIED.md. You can check your own run against the reference hashes.
Where it breaks: Five seconds takes about 2.6 hours. Models on a spinning disk made the first step take 329 seconds. At 121 frames, 20 steps produced a dot-grid pattern, so 50 steps are required. Only one card, one prompt and Windows were measured, and the patches track one engine version.
2. Turn a phone video of a room into a Blender set (voyager-so/vauss, vauss run, vauss export-blender)
The steps: Film a slow, steady loop around the room with overlapping views. Run vauss check room.mp4 --room for a quick check of light, sharpness and coverage. Then vauss run room.mp4 --goal "walk-through of my living room" --driver fixed --run-dir runs/room (the fixed driver needs no AI account). Open runs/room/exports/viewer/index.html to look around, and use vauss export-blender runs/room for the Blender pack. vauss shot renders new orbits and push-ins.
How it works: A Gaussian splat is a scene made from millions of tiny colored 3D blobs. Vauss picks frames, works out where the camera was for each, then learns where to place the blobs so every view matches your footage.
Why it is good: The authors report 3 to 6 minutes of training per scene and say their trainer ran 5.4 times faster than gsplat-MCMC at about half the memory in their own October tests on server cards. One run writes a browser viewer, a web file, a Blender pack and a playable voxel game. The license lets you use it commercially during the preview, and you own what you make.
Where it breaks: Those speed numbers come from L40S and RTX PRO 6000 server cards; the README suggests 8 GB of graphics memory and 16 GB of system memory as a starting point, with no timing published for that. It is source-available, not open source, and later versions may be paid. The optional vauss ad product-film step uses paid fal.ai generation.
3. Direct a voice line in a desktop app with Booth (stepupgaming/booth)
The steps: Install Booth, pick a preset voice, set mood, pace and energy, generate several takes and keep the best. For a full script, render it line by line with a voice per line.
How it works: Booth wraps Zyphra's own ZONOS3-Mini engine and downloads the model weights on first launch.
Why it is good: Takes and scripts are how voice directors already work, and it runs on the processor.
Where it breaks: Mac and Linux builds are untested on real hardware, the app is unsigned, and the faster engine needs an RTX 20 series or newer card.
Worth testing
- Wan 2.2 5B in a hosted Space before you commit a night to the local recipe. Tradeoff: we could not open the Space this run, and a hosted run is not the 4 GB setup.
- wan22-pascal-4gb at 17 frames on any old NVIDIA card you have. Tradeoff: only a GTX 1050 Ti was measured; your hashes may differ.
- Vauss on one room video with
--fastfor a preview. Tradeoff: no published timing for consumer 8 GB cards. - Mimvox Quick preset with one minute of your voice. Tradeoff: noncommercial only, 30-plus GB of disk, early and buggy by its own account.
- Booth on Windows with a ZONOS3 preset voice. Tradeoff: unsigned installer, and cloning still needs consent.
What actually matters from today's signal
The upgrade treadmill is losing its grip on who gets to make things. The wan22-pascal-4gb author put it plainly: "the card is modest, and that is the point." A laptop that was retired for gaming can make honest-quality video if you will let it run overnight. For a student, a hobbyist or anyone in a market where a new card costs a month's pay, that is a bigger change than another hosted model.
Time becomes the budget. Two and a half hours for five seconds means you cannot iterate on the old laptop; you draft somewhere fast and render finals slowly. The useful habit is the one the recipe models: test short, keep the settings that worked, and only then commit a long run.
The counter-signal is licensing and proof. Mimvox is noncommercial, Vauss can change its terms in later versions, and almost every number today is one author on one machine. Read the license before you build a client job on a tool, and treat single-machine benchmarks as a starting point.
Source access notes: Hugging Face (site and API) was blocked from both shells this run, so no Hugging Face dates were checked and no Hugging Face item is dated here; the Wan 2.2 Space link comes from search results. shields.io, Trendshift, HN Algolia, openai.com, blog.google, deepmind, Adobe's blog, Runway, BFL, Stability, Luma, ElevenLabs, Suno, Replicate, fal, Recraft and all Bucket D trade sites were unreachable by direct fetch; discovery leaned on GitHub (API and raw files), web search and fetched digests. The Neuron's November 13 digest failed to load. Adobe MAX (November 10: Firefly Creative Agent, partner models, Lightroom Open Eyes, Premiere Color Mode) predates this window and was covered earlier in the week. ElevenMusic's section editor is from August. Retell Tell-1 (contact-center voice) and NVIDIA BirdDreams (driving research) are out of scope. ComfyUI-Banzhang-All was covered November 12. Writing section omitted. Adversarial pass: ran (Sonnet subagent). It caught: Mimvox disk size understated and Apple Silicon support overstated, ClipForge punch-ins and YouTube input overstated, the Wan Space unverifiable (now hedged), the text-encoder wording, and a flat "no new model" claim (now "none we found"). It reported lower star counts for Mimvox (35) and ClipForge (15) than our run-time GitHub API reads (54 and 29); we kept our reads and flag the gap. It also flagged Vauss license terms as unverified from the README; we confirmed them in the LICENSE file itself.