Creative AI Briefing November 10, 2026 morning edition 12 min read

This week's useful tools let a creator lock the part they already approved and redo only the part that is wrong, a failure Adobe's own creator survey names.

Open-H3-Regenerate-2K-IC Logic Pro 13 Cinematic Skill Viggle Remake Adobe MAX video-gen upscaling music-gen creative-workflows comfyui ai-editing

Creative AI Briefing: Tuesday, November 10, 2026

You can now take a cheap 768p draft of a talking scene and get it back at 2560 by 1440 with the same mouth shapes, the same eyes and the same lettering on the sign behind the actor. That is NVIDIA's Open-H3-Regenerate-2K-IC, posted November 9, with a community ComfyUI version the same morning. It belongs to a pattern that runs through the whole weekend: tools that lock what you already approved and redo only the rest. Logic Pro 13 splits a finished drum loop into kick, snare and hats so you can turn down one piece. Two new ComfyUI packs stretch Viggle Remake swaps into long takes, one by locking the last frames of each window into the next. A code-drawn documentary skill re-times its whole film from a new voice without touching the picture. Adobe MAX opens today in Miami Beach, and Adobe's own survey from yesterday names the problem these tools answer: about a quarter of creators who quit an AI project mid-way said they could not change one part without redoing the rest.

New models

Open-H3-Regenerate-2K-IC finishes a draft at 2K without moving the faces. NVIDIA's SANA team posted the weights on November 9 (Hugging Face API created 01:50 UTC) with code in the sol-engine branch of NVlabs/Sana. It is an add-on for MiniMax H3, the open model that makes picture and sound together. You make a 768p draft (1344 by 768, about five seconds), and this pass redraws it at 1080p, 2K (2560 by 1440) or 4K (3840 by 2160) while keeping "the draft's faces, lip sync, lettering and layout," per the model card. The draft sits beside the new video as a reference the model looks at while it draws, the way a colorist keeps the approved grade on a second monitor. The draft's audio is copied across unchanged. NVIDIA's project page now says that version "draws more texture but moves eyes, wrinkles and mouth shapes away from the draft; the IC version keeps them," and shows MiniMax's hosted 2K API turning a "PRIVATE CAR PARK" sign into unreadable marks where the new one keeps the words. Those are the authors' own picked examples; there is no benchmark table or timing for the IC version yet. The official script is Linux and NVIDIA only. The community ComfyUI port (created November 9, 08:17 UTC, 38 stars, node code Apache 2.0) ships a draft-to-2K workflow and a converted 5.87 GB add-on file, and says plainly that "2K on 32 GB VRAM requires model offload / low-VRAM attention handling" and that 4K is "usually impractical on a 32 GB GPU." The catch is the license. The weights sit under the MiniMax H3 Community License, and the card says "the European Union, the United Kingdom, the Republic of Korea and the United States of America are excluded." No hosted demo exists.

Video

Cinematic Skill makes documentaries with no generated video at all. Moh4696/cinematic-skill (created November 9, MIT, 63 stars by the GitHub API) is a skill for Claude Code, Codex, Cursor and Gemini CLI that writes a script, records a scratch voice, pulls freely licensed photos from Wikimedia Commons (public domain, CC0, CC BY, CC BY-SA) with their credits, and flies one continuous camera over an evidence-wall board where "stamps slam, highlighters swipe and numbers roll up on the exact word the narrator says." Every frame is drawn in Python, so nothing drifts. The README says the 92-second Wright brothers example renders at 1080p in about six minutes on a 16 GB laptop. Full recipe below.

Viggle Remake now runs past one window. Viggle Remake, the outfit-and-actor swap add-on for MiniMax H3, renders about 12 seconds at a time per its model card. Two ComfyUI packs posted November 9 stitch longer takes. SGUN-father/ComfyUI-Viggle-Remake-LongVideo (Apache 2.0) targets what its author describes as a flaw: ComfyUI's built-in context windows slice only the output, so later windows repeat the opening moves. This pack gives each window its own slice of the source clip. hojenaia/viggle-remake-comfyui-long-take hands each new window the previous take's last 17 frames as locked starting material, and says to check identity "at every window seam." Both are three-star experiments. Viggle's card puts one 345-frame window at about 28 minutes on a 32 GB card with about 60 GB of system memory, and the weights sit under the same MiniMax H3 Community License.

Audio and music

Logic Pro 13 splits a drum loop into its kit pieces. Apple released Logic Pro 13 on November 5 (Production Expert, same day; Apple's release notes page is dated November 6), and a long hands-on from CDM, landed November 9. The useful new trick is in Stem Splitter: a "Drums (Individual Kit Pieces)" preset pulls kick, snare, hi-hat, toms, cymbals and other percussion onto separate tracks in a summing Track Stack, per Apple's notes. That means rebalancing a sampled or bounced drum part without the original multitrack. The new Browser tags your own sample folders on the Mac and lets you search them by description or with Find Similar (Apple silicon required for Find Similar). Both are free for owners of the $199.99 app. The intelligent assistant (beta) is not: it needs an Apple Creator Studio subscription ($12.99 a month or $129 a year), macOS 27, Apple silicon and U.S. English, and its monthly usage limit depends on "the complexity of conversations," per Apple's intelligence features page, with paid top-up packs once you hit it. CDM's Peter Kirn "never did find a single task that worked effectively" with it and found the plain-language browser search "hit-or-miss." Apple's support page settles who runs it: the assistant is "powered by Google Gemini."

Open and local

The local story is a finishing pass that refuses to change what you approved. If you are outside the excluded territories and have a 32 GB card, Open-H3-Regenerate-2K-IC turns MiniMax H3's cheap 768p drafts into 2K deliverables in ComfyUI today, with offloading. Everyone else gets code-drawn video skills, like Cinematic Skill and the launch-video skill, that use no generated footage at all.

slmonker/ComfyUI-OpenH3Regenerate2KIC: draft-to-2K finishing for MiniMax H3 clips in ComfyUI, faces and lettering held. Created November 9, 38 stars (repo).
kkget/ComfyUI-Open-H3-Regenerate-IC: a second port that cuts memory use with chunked attention and tiled upscaling. Created November 9, 3 stars (repo).
Moh4696/cinematic-skill: narrated archive documentaries and explainers drawn entirely in code. Created November 9, 63 stars (repo).
SGUN-father/ComfyUI-Viggle-Remake-LongVideo: Viggle Remake swaps past 15 seconds without repeated opening moves. Created November 9, 3 stars (repo).
hojenaia/viggle-remake-comfyui-long-take: continuous long takes by locking 17 overlap frames between windows. Created November 9, 3 stars (repo).
ZSeven-W/h3-gpt-image-mv: music video skill from keyframes to frame-accurate cut. Created November 9, 4 stars (repo).
theboringhumane/launch-video-skill: product launch films in HyperFrames with voice and loudness checks. Created November 9, 2 stars (repo).

Creative workflows

1. A 90-second archive documentary from freely licensed photos (cinematic-skill, skill in skills/cinematic-film/)

The steps: Install with npx skills add Moh4696/cinematic-skill (or /plugin marketplace add Moh4696/cinematic-skill in Claude Code), then pip install -r requirements.txt and put ffmpeg on your path. Ask for "a 90 second documentary about" your subject. Approve beats.md, the beat table with the facts. Let fetch_commons.py pull photos with licences into photos.json; it refuses non-free files. Time the film with the free scratch voice (tts_say.py, then split_takes.py, then build_narration.py). Approve stills of the opening and the look, then a 30-second draft at 1280 by 720. Record the final voice one section at a time, rerun the split, then add sound effects, mix and captions. Ship credits.md with the film.

How it works: Every stamp and highlight is tied to a spoken word (doc.W(line, "word")), not a timecode, so a new narrator re-times every hit by itself.

Why it is good: Nothing is generated video, so nothing drifts, and the rights trail is written as you go. The rules ban stand-in faces, AI images of real people and "recreated" documents that add words nobody wrote.

Where it breaks: The camera choreography is code, so revisions go through your agent, not a timeline. The references/voice.md file refuses tempo fits beyond 0.90 to 1.15x. The references/performance.md file puts 4K at about three hours for five minutes on a 16 GB laptop. Commons pages can be wrong, and the skill says to check the stated source for anything that matters.

2. Draft at 768p, finish at 2K with the faces held (open_h3_regenerate_2k_ic.json)

The steps: Install the node pack from ComfyUI Manager (Registry ID comfyui-open-h3-regenerate-2k-ic). Place the MiniMax H3 model files the README lists, plus three add-ons loaded in order: HyperFlow 8-step, LMS Sharp, then the converted IC file. Put an existing 768p H3 draft in ComfyUI/input/, open open_h3_regenerate_2k_ic.json, select it in Load Video, paste the original prompt into the Official Prompt node, and run. For 1080p set both resolution nodes to 1920 by 1088.

How it works: The draft rides along as a clean reference while the 2K version is drawn next to it in three steps, so the new detail has to agree with the draft's faces and text.

Why it is good: It turns the cheapest H3 setting into the drafting stage. You judge timing, performance and lip sync at 768p and pay for 2K only on takes you keep.

Where it breaks: The territory license excludes the US, EU, UK and Korea. A 32 GB card needs offloading; 4K is impractical there. Clips are about five seconds (124 frames at 24 fps, at least 113 real frames). It only knows H3 drafts at 768p. No published timing exists for the IC version.

3. Rebalance a bounced drum loop in Logic Pro 13 (Apple release notes)

The steps: Update to Logic Pro 13. Select the drum audio region, open Stem Splitter, choose the Drums (Individual Kit Pieces) preset. Logic creates Kick, Snare, Hi-Hat, Toms, Cymbals and Other stems inside a summing Track Stack. Mix each like a multitrack: tame the hats, compress the snare, swap the kick with a sample.

Why it is good: Loops, old bounces and client stems stop being take-it-or-leave-it, and it is free for existing owners.

Where it breaks: Separation leaves artifacts where pieces overlap, as all splitters do; no stem will match a real close mic. Apple's notes do not list which Macs are fast enough, and pro alternatives (Pro Tools 2026.11, Cubase 16) added six-stem separation but do not split the kit itself, per Production Expert.

Worth testing

Logic Pro 13 kit-piece split on a sampled break you love. Tradeoff: artifacts on cymbal-heavy material, and the assistant that gets the headlines costs a subscription and needs macOS 27.
Cinematic Skill's Wright brothers example (new_project.py ~/wright --example). Tradeoff: needs Python, skia-python and ffmpeg, and the look is a style; not every film wants an evidence wall.
Viggle Remake in the Viggle app, which the model card calls "no setup, no GPU." Tradeoff: the card itself warns it will put a real person's likeness into footage they never agreed to; label output as AI.
Adobe MAX keynote today, in person in Miami Beach and online (max.adobe.com). Tradeoff: expect paid, credit-metered features; read the plan fine print before rebuilding a workflow around them.

What actually matters from today's signal

The expensive part of AI work is not the render, it is the redo. Adobe's Creators' Toolkit survey (Harris Poll, 2,000-plus US creators, October 2026, commissioned by Adobe) says 47 percent of AI users gave up on a project mid-way in the past three months, and 27 percent of those could not change one part without redoing the rest. The weekend's best tools answer that complaint directly. The IC finishing pass keeps the approved mouth shapes. One long-take pack locks the approved frames and carries them forward. Logic separates the snare you hate from the kick you like. The documentary skill swaps a voice and leaves the picture alone.

Budget the day accordingly: spend the cheap passes deciding, and make the expensive pass a finishing step that is not allowed to change your decisions. Expect Adobe to pitch its own version of that on stage today.

The counter-signal is the same one as last week, sharper. Every MiniMax H3 weight file in this briefing, the IC finisher and the Viggle Remake weights behind both long-take packs included, sits under a license that excludes the US, EU, UK and Korea (Viggle's hosted app is a separate question we could not settle). The open video stack is getting good at finishing, and most of this newsletter's readers cannot legally ship with it. The free option with the simplest rights story today is the one that skips generated video entirely.

Source access notes: Hugging Face, vendor blogs, Apple, Adobe, CDM, MusicRadar, Production Expert and HN were read through a separate web reader; Hugging Face dates come from the API createdAt (Open-H3-Regenerate-2K-IC 2026-11-09T01:50:47Z; the converted ComfyUI file 2026-11-09T08:18:03Z). GitHub data from the GitHub search API and raw.githubusercontent.com; shields.io was unreachable, so star counts are GitHub API counts. Logic Pro 13's ship date (November 5) is from Production Expert and Consomac; Apple's release-notes page carries a November 6 page date but no date on the 13.0 entry. The assistant's model provider comes from Apple's support page 127888 (Google Gemini); MusicRadar's earlier report of OpenAI models was not used. Chimera Vocal Synth's page is JavaScript-only; only its description was readable. OpenAI news failed to load; Midjourney, Runway, Apple Newsroom and blog.comfy.org returned no dated items; Ideogram's blog carries no dates. BFL (October 15), Stability (August 25), ElevenLabs (October 20), Suno (November 4), fal (October 20), Replicate (November 3), Krea (November 3), Pika (October 29), Recraft (November 7, a guide) and Luma (November 3) had nothing new in the window. Hugging Face's newest-model lists for text-to-video and text-to-image appeared stale (newest entries October 20 and 27). The ComfyUI releases page served a stale list; not cited. Adobe MAX news had not posted at run time. Image section omitted: nothing new and verifiable in the window. Fact-check pass: an adversarial pass checked the draft against primary sources and caught eleven problems, all fixed: the Adobe survey's 27 percent overstated in the opening and thesis (it is a subset of the 47 percent who quit); the two Viggle packs described as working the same way; a ComfyUI flaw stated as fact that is the pack author's claim; MiniMax's 2K API called "paid" without a source; the territory exclusion attributed to the Viggle card, which names the license but lists no territories; Cinematic Skill's sources narrowed to public domain (it also takes CC0, CC BY and CC BY-SA with credit); the kkget port's offload description, not in its README; the Apple silicon footnote scoped to the whole Browser instead of Find Similar; Pro Tools and Cubase stems called "instrument" stems; an unsupported "first" for CDM's review; and an unsourced "Mac App Store" and "data-center" detail. The assistant's model provider was then settled from Apple's own support page. Chimera Vocal Synth (73 points on Hacker News, November 8) was cut because its author, license and method could not be confirmed.