Creative AI Briefing November 4, 2026 morning edition 11 min read

As generation drops to a handful of passes and a few seconds per edit, the slow part of making things moves to the edit itself, and today's most useful free release attacks exactly that.

Pauseless Krea 3 Seedance 2.5 Wave Forcing Idol Turbo MiniMax H3 ai-editing video-gen creative-workflows local-creative-ai open-weights 3d-gen

Creative AI Briefing: Wednesday, November 4, 2026

A podcaster can now drop an hour of talking-head footage into a browser tab, have every pause and most of the "um"s marked in seconds, and open the result as a Premiere, Resolve, Final Cut or CapCut timeline without uploading a single frame. That free tool, Pauseless, landed on Show HN overnight, and it fits the shape of the week. Image edits now finish in about eight seconds on Krea 3, a community add-on cuts a photo-to-3D model from 50 passes to 5, another squeezes a photo editor into a single pass, and a research team streams video at more than a hundred frames per second on a rack of GPUs. When making the raw material gets this fast, the hours move to choosing and cutting, and that is where the useful tools are showing up.

New models

Wave Forcing streams video faster than it plays, on serious hardware. SJTU-DENG-Lab/WaveForcing-1.3B (created October 30, 18:31 UTC; its paper topped Hugging Face's daily papers for November 3 with about 100 upvotes) is a small text-to-video model built on Wan 2.1 that makes video in a continuous stream instead of one finished clip at a time. The repo reports the authors' own numbers for the four-pass version: roughly 150 frames per second steady and about 135 end to end at 832 by 480, on eight NVIDIA H200 cards. (The table has been revised since launch, so check it before quoting.) For a creator the point is direction, not your desk: live, steerable video (a stream that reacts while you watch) is becoming a systems problem rather than a model problem. License: the weights and runtime are Apache 2.0, so commercial use is allowed; the reference training code, derived from another project, carries a separate academic-use license. The catch: only the 1.3B model is out (in four- and five-pass versions), larger checkpoints are "coming soon", the 14B row in the table measures speed with random weights and says nothing about quality, and nothing here runs live on a single consumer card. There are hosted demos at akhaliq/wave-forcing and hugging-apps/wave-forcing.

Image

Krea 3 publishes its homework. Krea's technical report (November 3) puts numbers on the editing model it launched in October (ComfyUI added a Krea 3 partner node in v0.42.0 on October 18): edits finish within about 8 seconds in production (4 to 5 at the fastest setting), it combines up to 8 reference images per edit, and it trains at native 2K, which Krea credits for better lettering and fabric, hair and skin detail. Krea's own API comparison prices Krea 3 Fast at $80 per thousand edits against $150 for Nano Banana Pro. The report also lists what it still gets wrong: very dense small text, holding a face or product identity across edits, and "not yet real time." Weights are not mentioned, so treat it as a hosted tool. Try it at krea.ai/krea-3.

A one-pass photo editor, for sketching only. akhaliq/Fibo-2.0-edit-IDM-1step (created November 2, 16:43 UTC) is a community add-on that makes Bria's Fibo 2.0 Edit Turbo work in one pass instead of four. It is best at colour, expression, object swaps and style; the card says structural edits and in-picture text "remain hardest" at one pass, and its own small test scores it slightly below simply running the teacher for one pass. License follows Bria's base: non-commercial. It runs in Draw Things and through Colab notebooks.

Video

MiniMax H3 keeps getting faster add-ons. Alibaba's TaoLive team posted TaoMate-HD-preview (created November 2, 07:50 UTC), an add-on for MiniMax H3 that generates speech and high-resolution video together in short chunks, with a paper listed the same week. Its example commands assume eight GPUs, so this is research to watch, not something to install. A separate add-on, Perflow-Shuai/H3-AR-Few-Step-LoRA (created November 3, 04:34 UTC), produces video with stereo sound at three passes per 1.4-second chunk, but the card says only a single 5-second, 1344 by 768 clip has been validated and it needs its own custom runner. Both carry the MiniMax H3 community license.

Replicate wrote the H3 buying guide. Which MiniMax H3 model should you use? (November 3) recommends Highspeed ($0.06 per second at 768p) for everyday reference-driven work, Lite ($0.025 at 480p, $0.04 at 768p) for storyboards and batches you expect to throw away, and full H3 ($0.08 at 768p, $0.13 at 2K) for finals. Two useful warnings: typing "2K" in the prompt does nothing (set the resolution field), and Lite "is also the model where you should look hardest at whether it did what you asked."

Audio and music

A 7 MB voice that runs faster than real time on a CPU. klein-ml/Klein-TTS-X1.5-Flash (created November 3, 16:16 UTC) is a 1.7 million parameter text-to-speech model with five preset voices in American and British English, French and Italian, Apache 2.0, with a Node package for offline use. The card is upfront: no voice cloning, no emotion control, and only English was tested. Good enough for scratch narration on a timeline; not a finished voice.

Premiere mobile gets label music, with a fence around it. Adobe announced (November 3) that the free Premiere mobile app will add clips from Universal and Warner, "coming soon", with no date. The terms matter: the music is licensed for personal, non-commercial use, and Adobe points commercial projects to Adobe Stock Audio. Fine for a personal reel, wrong for a client spot.

Writing

Live subtitles and translation for streams on a Mac. wquguru/cicada (Apache 2.0, 873 stars on shields.io, trending on Trendshift) captures call or stream audio and recognises speech on the device with Apple's speech engine. Subtitles and translation work without an account (full sentences go to Ling's cloud model, partial lines to Google Translate, with Apple's on-device translation as fallback); meeting notes need your own Ling key, which the README estimates at about 0.15 to 1 RMB per hour. It needs macOS 26 or later on Apple Silicon. Useful for anyone interviewing across languages; most translation is not offline.

Open and local

The local story today is fewer passes. Idol Turbo makes photo-to-3D a five-pass job, Fibo goes to one pass for edits, and the H3 add-ons chase streaming. The other half is a browser tool that never touches a GPU at all.

tahaygun/pauseless: removes pauses and filler sounds from talking video in the browser and exports an editor timeline. Repo created October 30, MIT, Show HN November 4 (repo).
akhaliq/idol-turbo: Idol photo-to-3D in 5 unguided passes instead of 50 guided ones, about 0.75 seconds on an H200 (author's figure). Created November 3, Apache 2.0, demo Space (card).
SJTU-DENG-Lab/WaveForcing: streaming text-to-video runtime and 1.3B preview weights. Weights created October 30, Apache 2.0 weights and runtime (repo).
klein-ml/Klein-TTS-X1.5-Flash: tiny CPU voice in four accents across three languages. Created November 3 (card).
akhaliq/Fibo-2.0-edit-IDM-1step: one-pass Fibo image edits, non-commercial. Created November 2 (card).
wquguru/cicada: on-device live captions for Mac calls and streams, mostly cloud translation. Trending on Trendshift (repo).
Appllama/appllama-skills: agent skills for designing Expo and React Native app screens from patterns in top-grossing apps; the research half needs a paid plan. 4.5k stars on shields.io, trending on Trendshift (repo).

Creative workflows

1. Cut the dead air out of a talking video and keep editing in your own app, with Pauseless (repo)

The steps. Open the page in Chrome or Edge on Windows or macOS. Drop in an MP4, MOV, M4A or WAV. Set the silence threshold (or Auto), the minimum pause length, the padding kept around speech, and the minimum clip length so coughs and clicks drop out. Switch on filler sounds and set sensitivity. Scrub the timeline: click a cut to keep it, drag to mark a filler it missed. Export .fcpxml for Final Cut or Resolve, .xml for Premiere, a CapCut project, or .edl, and fill in the optional video folder path so the editor finds the source without relinking.

How it works. It decodes only the audio track, finds stretches below your loudness threshold, and runs a tiny on-device detector for "um" and "uh". The timeline it exports points at your original file, so nothing is re-rendered unless you ask for an MP4.

Why it is good. It is free, needs no account, and the video never leaves your computer. The author benchmarks a one-hour, 2.7 GB file at about 5 to 6 seconds of analysis. Because you get a timeline rather than a baked video, every cut stays reversible in the editor you already use.

Where it breaks. The README says the filler detector finds about 70% of filler sounds with 70 to 80% precision on held-out English and Turkish speech, so expect to fix misses by hand. It needs Chrome or Edge (no Safari or Firefox). MP4 rendering re-encodes the picture and needs H.264 or HEVC. The project is days old with one star.

2. Make a dialogue-free illustrated short with Krea 3 stills and Seedance 2.5 (Krea guide, November 2)

The steps. Design each character as a still (the film's creator antrofrog credits Midjourney 8.2 for one lead and Krea 3 for the other). Paint each setting in one flat style, then place the saved characters into those plates with Krea 3, using up to 8 references per edit. Animate each beat with Seedance 2.5 on Krea from a start frame: one action, one camera move, ambient sound only, mouths closed. Keep clips at 4 to 8 seconds and draft at 480p before rendering at full quality. Add music in the edit.

How it works. The stills carry the costume, palette and line weight, so the video model only has to move what is already designed. Short clips with one action each keep it from inventing new story.

Why it is good. The guide's example runs about 59 seconds across roughly ten settings, made by one person. Locking the look in stills is cheaper than fixing drift in video.

Where it breaks. The guide reconstructs a method; antrofrog has not published prompts or settings. Seedance 2.5 on Krea needs the Pro plan or higher. Krea blocks uploads showing realistic human faces, which is why the guide uses illustrated frames. The original film uses a copyrighted song, and the guide tells you to use music you own or license.

3. Compare MiniMax H3 tiers fairly before you commit a budget, with Replicate's test plan

The steps. Pick one prompt and one set of references. Run it several times each on Lite, Highspeed and H3, all at 768p, same duration and aspect ratio. Score subject consistency, camera behaviour and text. Time the runs separately from clip length. Test 2K on its own.

How it works. H3 defaults to 2K and Lite to 768p, so a casual comparison mixes resolution with model quality. Fixing every setting isolates the model.

Why it is good. At 768p, a hundred 10-second clips cost $40 on Lite, $60 on Highspeed and $80 on H3, so knowing which tier passes your test is real money on a series.

Where it breaks. Replicate itself ran one sample per tier for timing and says it "proves very little"; your own runs are the evidence.

Worth testing

Pauseless: run last week's interview through it and open the timeline in your editor. Tradeoff: roughly three in ten fillers slip through.
Idol Turbo demo: one product photo to a 3D object in about a second of compute. Tradeoff: trained on 93 images, and some results come out softer than the full model.
Wave Forcing demo: see what streaming video looks like at 1.3B. Tradeoff: preview quality, and the speed numbers need eight H200s.
Krea 3: multi-reference edits in seconds. Tradeoff: identity still drifts across edits, per Krea's own report.
Recto (Show HN, November 3): a free desktop canvas for Mac, Windows and Linux where your coding agent writes HTML, CSS, SVG or 3D while you adjust it with the mouse. Tradeoff: version 0.1.6, and it needs an agent such as Claude Code or Codex already installed.

What actually matters from today's signal

Count the passes. Idol went from 50 to 5, Fibo editing from 4 to 1, the H3 add-ons work in three passes per chunk, and Krea sells edits in eight seconds. When generation costs seconds, a creator's bottleneck stops being the render and becomes judgement: which take, which cut, which version goes to the client. Spend your attention there. Generate wide, then be ruthless.

That is why the most practical release of the day is not a model. Pauseless does the dullest job in video editing, removing dead air, and hands the result back as a timeline rather than a finished file. Tools that return editable decisions instead of baked output will beat tools that return pretty files, because the creator stays in charge of the last 10%. Expect more of this: small, free, local tools that do one edit chore and export to Premiere or Resolve.

The counter-signal is the fine print on the fast versions. The one-pass Fibo add-on is non-commercial, the H3 add-ons inherit MiniMax's community license, Adobe's new label music is personal-use only, and Wave Forcing's headline speed lives on hardware few creators will ever rent for an afternoon. Fast is not the same as finished, and free is not the same as cleared.

Source access notes: Thin day for the major labs: no new creative releases in the window on blog.adobe.com (beyond the Premiere mobile music note), elevenlabs.io/blog, blog.fal.ai, stability.ai, lumalabs.ai, bfl.ai, suno.com/blog or Krea beyond its report and guide. openai.com/news (client error), midjourney.com/updates (no content rendered), runway.com/news (no dated list), blog.google AI landing page (no dated posts), github.com/search (robots), api.github.com (client error) and the Hugging Face API from the shell (egress blocked; WebFetch used instead) were unavailable. The Hugging Face trending-models API returned a stale snapshot (newest items from September) and was ignored; Spaces trending and createdAt-sorted listings used instead. All Hugging Face dates from API createdAt. Pauseless repo creation date from shields.io ("last friday", read as October 30); Show HN dates from the HN Algolia API; cicada star count from shields.io (873) disagrees with a cached repo page (83), so shields.io used. Adversarial fact-check pass ran (Sonnet subagent): it corrected Wave Forcing's speed figures to the current README table, fixed the claim that its paper and training code were unreleased (the paper is out, training code is academic-use only, only larger checkpoints are pending), lowered its upvote count to about 100, widened Pauseless's analysis time to 5 to 6 seconds, and corrected cicada's translation routing and macOS 26 requirement; all other dates, prices, licenses and quotes checked out. Bucket D: Creative Bloq and No Film School had nothing in the window; PetaPixel served no dates. Google, Sora, Kling, Ideogram, Recraft, Pika, HeyGen and Udio checked by search only; no new release found.