FervorCreative AI
Live Latest 18.11.26 · morning 401 tools tracked 992 workflows indexed 626 topics Hot: Gaussian splatting, ComfyUI, ZONOS3-Mini

This weekend's free creative tools ship with their own proof built in, so a creator can check a tool on their own machine in minutes instead of trusting someone else's demo reel.

tiny-vocaSınırsız FotoZ-Image-TurboOpen ExplainerFormaComfyUI Smart GuidesLumalocal-creative-aivoice-cloneimage-genvideo-gen3d-gencreative-workflows

Creative AI Briefing: Sunday, November 15, 2026

You can now put a narrator voice inside a web page, a game or a phone app that weighs 18 MB, reads English about 70 times faster than real time on four processor threads, and cost its maker $9.56 of rented graphics time to build. That model, tiny-voca, landed on November 14 alongside a photo-quality image recipe for a 4 GB laptop card, a beta toolkit that cuts an explainer video on the words of your own narration, and a free iPhone and Mac 3D scanning app. The labs were mostly between announcements. What the community shipped instead has one habit in common: every tool comes with a way to check it yourself, whether that is a pixel hash, a quality report on the finished video or a review sheet before anything renders. That turns "does this free thing really work?" from a leap of faith into a ten-minute test.

New models

tiny-voca: a two-voice English text-to-speech model small enough to ship inside your project. n4ze3m/tiny-voca (weights on Hugging Face, created November 14) has 4,537,291 parameters, an 18 MB file with the sound generator included, and two American English voices, F2 and M2, at 24 kHz. It runs in PyTorch, as an ONNX file, or straight in a browser, and the author reports about 70 times real time on four CPU threads. That size is the point for creators: a game developer, an interactive-fiction writer or a web designer can bundle a voice that works offline with no API bill. Code and weights are Apache 2.0, so commercial use is allowed; the dataset is CC BY 4.0. The training voices are fully synthetic: both were designed from text descriptions with Qwen3-TTS and no real speaker was recorded, which sidesteps the consent problem that hangs over most voice tools. The author says the whole project, failed tries included, cost $9.56 on one rented RTX 4090, and the final run took 7.7 hours. The catch, in the model card's own words: English only, two fixed voices, no cloning, a slightly synthetic sound, no streaming, one sentence at a time (up to 32 seconds per call), and its accuracy scores come from 8 to 16 clips, "a sanity check, not a benchmark." Try it free in the browser demo, which downloads about 40 MB on first visit and offers voice and speed controls.

No new foundation model for image, video, audio or music came from a major lab in the last 48 hours. Midjourney V9 is still only a target, "aiming to ship within 30 days, maybe sooner" per its November 11 office hours (secondary, via the Geeky Curiosity recap).

Image

Sınırsız Foto puts Z-Image-Turbo at full quality on a 4 GB laptop card. alexturkey/z-image-turbo-pascal-4gb (created November 14, v0.2.0 the same day, Apache 2.0 scripts, MIT patches) comes from the author of yesterday's Wan 2.2 video recipe. It makes a 768 by 1344 picture with the 8-bit Z-Image-Turbo file in 193 to 228 seconds on a GTX 1050 Ti, peaking at about 3 GB of graphics memory while the model's weights sit in system memory. Same seed, same pixels, and generate.py prints a pixel hash so you can prove it. Three patches to the stable-diffusion.cpp engine stop one part of the model from running twice per step, route the final decode through NVIDIA's cuDNN library (10 to 15 seconds instead of 23 to 27, and closer to a reference decode), and try decoding the whole image in one piece first. Setup downloads about 12.2 GB. The catch is the machine it was measured on: Windows only, one 16 GB laptop, three image sizes, and free memory drops to a few hundred megabytes while it runs, so close everything else. The README says the model card is silent on who owns generated pictures.

Luma's agent now tidies your board. Luma's smart organizer (November 14) groups and labels assets on a board, frame or selection by product, campaign, concept, production state or channel, up to 10,000 assets, and "nothing moves until you say yes." Plan and price are not stated on the page.

MiniMax H3 Image Conditioning Plus (tritant, created November 14, Apache 2.0) is a drop-in ComfyUI node that feeds neutral visual patches to MiniMax H3's vision side when you have no reference image. The author claims better textures and fine detail; that is the author's own claim, with one screenshot and no comparison set.

Video

Open Explainer cuts a narrated explainer on the words of your own voice. Maria-Aleana-De-Leon/Open-Explainer (created November 14, v0.3.0 dated November 14 in its changelog, MIT) takes a voice recording and a storyboard written as phrases, not timestamps, and renders a 1080p MP4 where every cut, diagram step and mascot gesture lands on a word. It runs on a laptop processor with no graphics card, and an agent (Claude Code, Codex) can write the storyboard. Full details are in Creative workflows below. The catch: public beta, tested on macOS only, CI has not passed.

Runware retires Gemini Omni Flash 1.1 on December 3. Runware's changelog (November 13) points users to Omni Flash 1.2, which raises reference images from 7 to 10 and clip length from 10 to 15 seconds. If you have saved workflows calling 1.1 there, move them before December.

Audio and music

tiny-voca leads above. ElevenLabs' November 11 Summit NYC post is enterprise-agent news (an ElevenAI work assistant in free early access, voice-model fine-tuning "in the coming months"), with nothing new for voice, music or dubbing creators.

Open and local

The local story today is proof you can run at home. The image recipe prints a pixel hash, tiny-voca publishes its accuracy check and its full training recipe, and Open Explainer measures its own finished video for loudness, sync and timing. Star totals are GitHub API figures at run time; shields.io and Trendshift were unreachable, so there are no daily deltas.

  • n4ze3m/tiny-voca: an 18 MB, two-voice English narrator you can bundle into a game or web page. Created November 14, 110 stars, Apache 2.0 (repo).
  • alexturkey/z-image-turbo-pascal-4gb: full-quality Z-Image-Turbo stills on a 4 GB GTX 1050 Ti in about three and a half minutes each. Created November 14, 41 stars, Apache 2.0 (repo).
  • nomadoor/ComfyUI-Smart-Guides: snapping, alignment and even spacing for nodes on the ComfyUI canvas, so a messy workflow becomes readable. Created November 14, 20 stars, MIT (repo).
  • Maria-Aleana-De-Leon/Open-Explainer: narrated explainer videos timed to your voice, with your own cartoon mascot. Created November 14, 1 star, MIT (repo).
  • RobiiAragon/Forma: on-device 3D scanning for iPhone, iPad and Mac: LiDAR rooms, photographed objects to USDZ, and Gaussian splats trained on the device. Created November 14, 1 star, MIT (repo). Interface in Spanish for now; the iOS build must be sideloaded and the Mac app is not notarized.
  • tritant/ComfyUI_Minimax_H3_Image_Conditioning_Plus: a no-reference texture boost for MiniMax H3 text-to-image. Created November 14, 6 stars, Apache 2.0 (repo).

Creative workflows

1. Make a narrated explainer timed to your own voice (Open-Explainer, README Quick start, docs/GETTING-STARTED.md, showcase/ai-agents-60s/)

The steps: Install FFmpeg and uv, then uv tool install '.[transcribe]' and oex doctor. Try oex new demo --demo && oex -C demo render for a film with no recording. For your own: oex estimate script.txt --target 3:00 to check length, oex new my-film --format 16:9 --style paper, drop your recording into audio/narration.wav, write storyboard.json (or ask your agent), then oex -C my-film transcribe, render --preview, sheet, and a final render.

How it works: A speech-to-text model finds when you said each word. The storyboard names phrases ("the agent checks the price"), and the toolkit pins each scene change and diagram step to those words. Editing one scene re-renders only that scene.

Why it is good: Your voice is the edit decision list. The QA pass reads the finished MP4 for frame count, audio-video offset, cue timing, loudness, peaks, black frames and captions, and a credits file tracks the license of every sound effect.

Where it breaks: macOS only so far; Linux and Windows are untested. Lip sync is cartoon-grade, driven by the sound of the voice, not by phonemes. The free draft voice (macOS say) is licensed for personal use only, not for published films. Full-body lip sync from generated mouth sheets is unverified.

2. Turn a photo of yourself into a lip-syncing mascot (Open-Explainer docs/mascot/SHEETS.md, prompts/sheets/)

The steps: In ChatGPT or another image tool, upload your photo with 01-master-mascot.md and refine until you like it. Then generate a 3x3 pose sheet, a 3x3 expression sheet, the same expressions with mouths removed, and a seven-mouth sheet. Run oex assets split pose-sheet.png --kind poses --out ~/OpenExplainer/my-mascot, then oex assets place-mouths and oex mascot prepare. Check the review sheet before rendering.

How it works: The splitter finds the real gaps between drawings instead of assuming equal cells, cleans faint transparency, aligns poses on one canvas and flags cuts through artwork.

Why it is good: Your photo never leaves your image tool, and the review sheet catches drift before you spend a render on it.

Where it breaks: The project says a generated sheet is "not animation-ready by itself." The mouth sheet has no labels, the mouthless sheet is redrawn rather than edited, and lip sync from sheets is tested only on a head-and-shoulders mascot.

Worth testing

  • tiny-voca's browser demo with a paragraph of your own script. Tradeoff: two fixed voices, English only, and names it has not seen often come out wrong.
  • Open Explainer's demo film (oex new demo --demo). Tradeoff: macOS-tested only, and the free macOS draft voice is for personal use, not published work.
  • ComfyUI Smart Guides via ComfyUI Manager on your messiest workflow. Tradeoff: two days old, and it is a layout aid, not a speed-up.
  • Sınırsız Foto on any old 4 GB NVIDIA card. Tradeoff: only a GTX 1050 Ti on Windows was measured, and other cards need their own engine build.
  • Forma Desktop on an Apple Silicon Mac. Tradeoff: Spanish interface, unsigned builds, and its own README says the results are estimates, not measurements.

What actually matters from today's signal

The most useful thing a free tool can ship now is its own test. A creator deciding whether to spend an evening on an open project used to have a demo reel and a star count to go on. Today's releases hand you a pixel hash, a word-error check, a review sheet and a finished-video report. Ten minutes now tells you whether it works on your machine.

Small is getting good enough to ship. An 18 MB voice that runs in a browser will not win a listening test against a paid voice service, and its author does not pretend otherwise. But a game jam team, a museum kiosk or a children's story app does not need the best voice in the world. It needs one that is free, offline and safe to use commercially, made without anyone's voice taken without permission.

The counter-signal is the sample size. Every tool in today's list was measured by one person on one machine, and two of them have a single star. Treat each "it works" as "it worked for them," run the built-in check, and keep your deadline work on tools you have already proven.


Source access notes: Hugging Face was reached through a fetch agent; tiny-voca's date is from the HF API createdAt (2026-11-14T15:46:27Z). HF listings for text-to-video and image-to-video returned nothing newer than late August, which may be an API ordering quirk; HF likes and trending top 15s showed no image, video, speech or audio model created on or after November 12. Blocked or unreachable from this run: img.shields.io, Trendshift, hn.algolia.com (no HN discovery pass), openai.com/news (fetch failed). Checked with nothing in the 48-hour window: blog.fal.ai, replicate.com/blog, bfl.ai/blog, stability.ai/news, elevenlabs.io/blog (one Nov 11 enterprise post), Releasebot's Runway mirror (newest Nov 11), the Kombitz Nov 14 roundup (items dated Nov 6 to 11). Midjourney V9 is secondary-sourced. Star counts are GitHub API figures at run time. Repo dates are GitHub API created_at. Adversarial fact-check (Sonnet) ran on this draft: it found no major errors and caught seven minor ones, all fixed: workflow steps cited to the wrong Open Explainer doc, an overbroad "Pascal only" limit, a compressed description of the three image patches, "36 slices per step" (the log says 36 segments), a demo-voice claim, Midjourney "promise" softened to "target", and Luma's missing "channel" grouping.

Articles from this briefing