Creative AI Briefing November 5, 2026 morning edition 11 min read

This week's tools take finished media apart into pieces you can edit, so a GIF becomes frames, a film mix becomes stems, a reference video becomes a shot list and a timeline becomes a script, and the craft moves to reworking parts instead of rerolling wholes.

ChatGPT Images 2.6 GPT-Image-2.6 Flare GPT-Image-2.6 Sunburst iZotope Photon Kino AI rea-video agent-screenplay Luma MCP image-gen ai-editing audio-gen creative-workflows local-creative-ai

Creative AI Briefing: Thursday, November 5, 2026

Type a prompt into ChatGPT today and you can get back an animated GIF, then change its speed and fix one bad frame without regenerating the loop. That is the headline feature of OpenAI's Images 2.6, released late on November 4, and it rhymes with nearly everything else that shipped this week. iZotope's new Photon splits a finished film mix into dialogue, music and effects on your own machine. A free open-source tool breaks any reference video into shots, captions, stems and a remake plan. A young editor stores the whole timeline as code an AI agent can change. The finished file is turning back into parts, and the people who know which part to touch will save the most time.

New models

ChatGPT Images 2.6 makes GIFs you can edit frame by frame. OpenAI's community announcement (posted November 4, 23:49 UTC) introduces Images 2.6 in ChatGPT and says two models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, are "coming to the API with 50% faster image generation" (OpenAI's own figure, not broken down by job). Both already have pages in OpenAI's developer docs. The creative news is the GIF: "turn a prompt or image into an animated GIF. Adjust the speed, edit individual frames, and share it on repeat." Flare is the speed tier ("as little as six seconds"); Sunburst "brings higher-quality generation and greater consistency across edits." OpenAI also mentions image markup (draw on the picture to point at what to change) and SVG exports, a vector file you can scale for print. Secondary reporting (The Mobile Indian) says it reaches every ChatGPT tier on web, desktop and mobile, plus Codex, and that you can save and tag reference images of people, pets and characters for reuse. Price: OpenAI's developer model page bill by tokens (image output listed at $30 per million tokens), and no per-image price is published, so budget by test run. The honest catch: the API guide we could read still describes the 2.5 models, and frame limits for API GIFs circulate only in third-party write-ups and forum posts. Treat any "64 frames" figure as unconfirmed until OpenAI documents it. (openai.com blocked our fetcher; the announcement above is OpenAI's own forum post.)

Image

Recraft published a logo-maker walkthrough (Nov 2) showing how to go from a few business details to logo concepts you can compare and export. It is a tutorial, not a new model. If you are judging AI logo tools this month, pair it with a vector check: open the exported file and count the paths before you send it to a printer.

Video

Luma now works from inside Claude and ChatGPT. Luma's MCP connector (November 3) lets you turn one product photo into product shots, new camera angles, background swaps, relighting and short promo clips without leaving the chat. It "uses your Luma credits," and setup lives at lumalabs.ai/mcp. The catch: the post names no plans, prices per job, or which Luma models run underneath, so a chat request can spend credits you cannot predict. Run one small job and check your balance.

Kino AI is a free video editor where the timeline is code. The YC S23 company's Launch HN (posted October 31, still drawing comments this week) says "all editing, all motion graphics... and all 3D scenes are expressed as code, built on Remotion," so your own AI agent can edit the cut through its MCP access. The pricing page calls the full editor "Free forever," with 100 MCP tool calls a week for your agents; the founders' post puts Pro at $25 per editor per month (the page itself did not render the price). It ships as a desktop app for now; the founders say there is no browser version yet.

Research, no tool yet: MoviEdit, a long-form multi-shot video editing paper, sits near the top of Hugging Face's November 4 daily papers (65 upvotes at our check). Worth a bookmark, nothing to run.

Audio and music

iZotope Photon splits a finished mix on your own GPU. Per MusicTech (published around November 4; iZotope's own page blocked our fetcher, so this is secondary), Photon runs 12 custom-trained models locally, with "none of your files" uploaded. It separates music into four stems, drums into five (kick, snare, hi-hat, toms, cymbals), up to six different voices, and its Cinema model splits "a video or audio file into dialogue, music, sound effects, and everything else." Three cleanup models handle speech restoration, room noise and outdoor wind and traffic. For an editor, the Cinema split is the story: replace temp music under dialogue you cannot reshoot. It costs $349, or $259 in a sale through December 7, as a standalone app and an ARA2 plugin (one that opens inside your DAW's timeline) in Pro Tools, REAPER and Cubase, plus Sequoia and Samplitude on Windows. Needs an Apple Silicon Mac or Windows 10+. Catch: no trial or GPU minimum was stated in anything we could read.

Phonon TTS now speaks in the browser. blixt/phonon (on Show HN this week) ports Gradium's small Phonon voice models to WebGPU: about a 44 MB first download for the English 31M model (around 54 MB for the 40M models), cached after the first visit, with start and end times for every word. The README warns those times follow the model's attention rather than a forced alignment, so check them before you burn captions. The README's own benchmark for the 31M model: 115 times faster than real time on an M3 Max (against 31 times for the older WebGPU port) and 26 times on an iPhone 17 Pro. License matters here: the code and the English-only 31M model are Apache 2.0, while the 40M models (English, French, German, Spanish, Portuguese) are CC BY-NC 4.0, which the README says rules out commercial use.

Spotify plans paid AI remixes. Creative Bloq (November 4), citing the Financial Times, reports a plan for listener-made AI remixes and covers around February 2027 at a $4 to $5 monthly surcharge. A plan, not a product. Musicians should watch the licensing terms when they land.

Open and local

The local story today is less about a big model and more about small tools that dismantle media on your own machine. Photon does it for sound. rea-video does it for a whole edit, running offline with ffmpeg and Python and calling an image-reading AI model only if you give it one. Phonon puts a voice in a browser tab. Star totals below come from GitHub's API (shields.io was unreachable this run).

JuneYaooo/rea-video: reverse-engineers any video shot by shot into a report, stems, captions and a film/generate/code remake plan. Created October 29, 161 stars, trending on Trendshift today (repo).
gmpetrov/agent-screenplay: an AI agent writes a short JSON script and the tool films a polished screen recording with zooms, callouts and voice-over. MIT, created November 2, 61 stars (repo).
blixt/phonon: browser text-to-speech with word timings for captions. On Show HN this week; star count not verified this run (repo).
3351666087/scan-to-vector-skill: an agent skill that rebuilds scanned pages as searchable PDFs with vector diagrams. MIT, 7 stars, updated today (repo).
alex0ptr/aigude: turns the news feeds you follow into a daily AI-narrated podcast with chapters. AGPL-3.0, created November 3, 6 stars (repo).
harry0703/MoneyPrinterTurbo: topic-to-short-video automation (script, voice, footage, subtitles, music). An old project back on Trendshift's daily list; total not verified (repo).
empero-ai/PiVo-1.4M: a tiny Apache 2.0 text-to-speech model, created November 1 per the Hugging Face API. Unproven; listen before you build on it (model).

Creative workflows

1. Break down a reference video before you copy its feel, with rea-video

The steps. Install ffmpeg (brew install ffmpeg) and Python 3.10+. Clone the repo and run pip install -e ".[ocr]" (per docs/cli.md). Run rea-video decompose ref.mp4 -o out/ref, then open out/ref/report.html. For better reads on AI or 3D footage, set REA_VLM_BASE_URL, REA_VLM_API_KEY and REA_VLM_MODEL to any OpenAI-compatible vision model and add --vlm. Use rea-video stems ref.mp4 -o out/ref for vocals and music. The remake step, rea-video remake-plan out/ref/breakdown.json --capsule-root $CC, turns AI shots into first/last-frame jobs for the author's separate capsule-cinema repo, so most people will stop at the report. Add --lang en for English reports. Claude Code users can install it as a plugin (/plugin marketplace add JuneYaooo/rea-video).

How it works. It cuts the video into shots, saves first and last frames, reads on-screen text, tracks camera moves and beats, separates audio (a roughly 67 MB voice-separation model downloads on first run), and writes breakdown.json, a storyboard image and SRT/ASS caption files. The README says the remake plan splits each shot into "what to film, what to generate and what to code."

Why it is good. It turns "I like the pacing of this ad" into numbers you can act on: shot count, cut points, caption style, where the music hits.

Where it breaks. The README is in English, but the detailed docs are in Chinese. The repo carries an Apache 2.0 license file, though GitHub's metadata does not detect it and the README never mentions it. The "code" layer gives text tracks and layout but does not build a Remotion project yet (per the Chinese CLI doc), and automatic remakes of AI shots need a separate repo and cost money. The README itself says not to reuse other people's footage, music or likeness.

2. Script a product demo instead of recording it, with agent-screenplay

The steps. Run npx agent-screenplay init (downloads Chromium once and adds the MCP server to Claude Code, Codex or Cursor). Ask your agent for a scenario, or write one: a JSON file with a url, a style, and steps such as click, type, callout, highlight and say. Check it with npx agent-screenplay check demo.json, then npx agent-screenplay record demo.json.

How it works. It drives a real browser through the steps, then renders demo.mp4 (1920x1080 at 30 fps by default, up to 4K) and demo.storyboard.png, with optional GIF and voice-over through a Cartesia or ElevenLabs key or your own audio files.

Why it is good. A fix to the UI means editing one line and re-rendering, not re-recording a take.

Where it breaks. check and record perform the steps for real ("a 'Send' click sends"), so point it at a demo account. Linux desktop recording needs X11, not Wayland, and desktop recordings cannot run headless. Some sites re-ask bot checks on every visit and cannot be filmed automatically.

Worth testing

ChatGPT Images 2.6 GIFs: on every tier, per The Mobile Indian's launch report. Tradeoff: no published per-image API price and undocumented API frame limits.
Kino AI: free desktop editor with agent access. Tradeoff: young, desktop-only, and agent calls cap at 100 a week on the free plan.
Phonon in the browser: narration plus word timings with no server. Tradeoff: the commercial-friendly model is English only, the multilingual 40M models are non-commercial, and word times are approximate.
iZotope Photon: local dialogue/music/effects split for editors. Tradeoff: $259 on sale with no trial confirmed.

What actually matters from today's signal

The expensive habit in AI media has been the reroll: one bad hand, one wrong word, one stray music cue, and you regenerate the whole thing. This week's tools attack that habit from three directions. Images 2.6 lets you edit a frame instead of the loop. Photon lets you swap the music under a voice instead of re-cutting the scene. rea-video and Kino turn an edit into text a machine can read and change. Each one moves your time from waiting to deciding, and decisions are where taste lives.

The money question splits cleanly. Local tools (Photon, rea-video, Phonon) cost once or nothing and keep your files on your disk. Hosted tools (Images 2.6, Luma inside chat) bill in tokens or credits that nobody prices per job. The more an agent does on your behalf inside a chat window, the less you see the meter. Test with a small job and read your balance before you hand an agent a whole campaign.

The risk is the same one that shadows every "take it apart" tool: taking apart someone else's work. rea-video's own README warns against reusing footage, music or likeness, and Photon will happily pull a clean acapella out of a commercial record. The tools make learning from a reference trivial. They do not change who owns it.

Source access notes: Blocked or unreadable this run: openai.com (both announcement and news pages; used OpenAI's community forum post and developers.openai.com instead), izotope.com Photon page (Photon facts are secondary via MusicTech), Ideogram blog (both URLs), Midjourney updates (JavaScript-only, no content), The Verge (blocked), Hacker News Algolia API and Hugging Face API from the shell (reached via a fetch agent instead), shields.io (star counts from GitHub's API), news-radar Nov 3 digest. Civitai and blog.comfy.org skipped. Checked with nothing new in window: DeepMind, Google blog, BFL, Stability, ElevenLabs, Suno, fal blog, Pika, ComfyUI releases (latest v0.43.0, Oct 27). Krea's technical report, Replicate's H3 guide and Premiere mobile's label music were covered yesterday and are omitted. Fal's $16B funding round omitted (funding). Hugging Face dates are API createdAt. Adversarial fact-check pass ran (Sonnet subagent). It caught: Kino's Launch HN was October 31, not today (redated); OpenAI says the 2.6 API models are "coming," not launched (reworded); Phonon's speed and size figures belong to the 31M model and its word times are not forced alignment (corrected, multilingual 40M noted); MoviEdit upvotes 63 to 65; the rea-video remake command needs --capsule-root and feeds a separate repo (fixed); an unverified Kino platform list and an unfound auto-scoring claim (removed); tier availability now attributed to The Mobile Indian. The later article fact-check found rea-video does carry an Apache 2.0 LICENSE file that GitHub's API misses; corrected here.