Creative AI Briefing November 9, 2026 morning edition 11 min read

The weekend's useful releases build the director's checks into the tool, so a creator approves stills, counts props and compares fixes before paying for a render.

HyperFlow H3 Flash Video Director ComfyUI-PhotoDeblur InstantHMR Africa TTS Retro CCTV Video Adobe MAX video-gen creative-workflows comfyui local-creative-ai ai-editing voice-clone

Creative AI Briefing: Monday, November 9, 2026

You can now hand Claude Code a brief for a short ad and get back two or three labeled still frames of the same moment to choose between, a storyboard page you can pin notes to, and a motion plan, all before a single frame moves. That is Video Director, published Friday, and it sets the tone for a weekend when the big labs stayed home. The community shipped judgment instead of models: a surveillance-footage skill that counts props shot to shot, a photo deblur node that tells you to compare two models on the same picture, a body-capture node with a license table you can actually read. The one new video model of the week, a compressed MiniMax H3 that makes picture and sound on a home card, arrives with a license that shuts out most of this newsletter's readers. Adobe MAX opens tomorrow in Miami Beach, so expect the labs to speak next.

New models

HyperFlow H3 Flash makes video with a stereo soundtrack on a 24 GB card. Video Rebirth posted the weights on November 4 (Hugging Face API created 13:54 UTC; the model page showed 40,788 downloads last month this morning) and opened the code repo and ComfyUI nodes on November 6. It is a cut-down MiniMax H3: 13.7 billion parameters against the original's roughly 31 billion, run in 4, 6 or 8 steps, making 5.2 to 15 seconds of 24 fps video at 1344 x 768 with 32 kHz stereo sound baked in. What that means at a desk, by the team's own benchmarks: a 5.2-second clip with sound in 157 seconds on an RTX 4090 and 315 seconds on an RTX 3090, using the 9.1 GB compressed file in ComfyUI; a 15-second clip takes about 14 minutes on the 4090. A --fast-disk launch flag lets PCs with only 16 GB of system memory run it. Four ComfyUI templates cover text to video, first frame, first and last frame, and up to nine reference images. The catch is the license. The code is Apache 2.0, but the weights fall under the MiniMax H3 Community License, which the model card says is "licensed worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America"; people there "must obtain MiniMax's prior authorization before downloading." The card also lists an AI-content disclosure rule and a $20 million revenue threshold. Graphics cards with 16 GB are "not measured yet." No hosted demo exists; Video Rebirth's showcase page is examples only.

Image

ComfyUI-PhotoDeblur rescues shaky handheld photos for free. porfeng/ComfyUI-PhotoDeblur (created November 7, MIT, 77 stars by the GitHub API) wraps two published deblur models, HI-Diff (Apache 2.0) and MISCFilter (MIT), as native ComfyUI nodes that work at your photo's original size by processing it in overlapping tiles. Its README is unusually candid about what works: camera shake and motion trails on signs, storefronts and window frames improve; defocus, blown highlights and very small faces "are usually harder to restore," and a heavily blurred stage face and a cat's face "showed limited improvement." It tells you to develop RAW files first and save 16-bit PNG if you will keep grading. The weights download from the authors' Google Drive with checksum checks. MISCFilter needs extra NVIDIA setup.

Recraft added GPT Image 2.6. Recraft's November 5 post puts OpenAI's new image model inside Recraft Studio and runs 12 prompts through it and Recraft V4.1 Pro, and five of them through Nano Banana 2.1 too. It costs 10 to 57 Recraft credits an image; Sunburst works on the Free plan at default settings, while Flare needs a paid plan.

Video

A skill that directs short videos from brief to render. SallyRS/video-director (created November 6, MIT, 26 stars) is a Claude Code skill for product films, explainers, social shorts and live-action ads from about 6 seconds to 3 minutes. Its rule: "Nothing is animated until you have approved it as still frames, and every visual decision is a pick from 2-3 real options, never a reaction to one take." It builds the final film in HeyGen's open HyperFrames engine, plans sound alongside motion, fits music to the cut on bar lines, and runs Python render checks on cuts, text reading time and sound levels. The README says its text-size and timing rules come from measuring 59 films. Full recipe in the workflows section.

Fake surveillance footage, with continuity rules. 0xMomo-NGClubs/retro-cctv-video (created November 6, MIT, 53 stars, docs in Chinese) turns a historical or invented story into old CCTV footage that seems to have caught it by accident (our translation of the README). The value is its checklist, learned from failed takes: write each beat as trigger, action, contact, result; track how many of each prop exist and who holds them; keep the camera fixed and never punch in; draw the timestamp text on a separate low-resolution layer and soften only that layer. It ships a client only for a RunningHub MiniMax H3 workflow; any other generator means bringing your own.

Audio and music

Africa TTS speaks 92 African languages on a plain CPU. AfriSpeech/africa-tts (repo created November 6) builds on Kyutai's Pocket TTS, gives every language a default voice, and clones a voice from a 3 to 10 second clip. The README reports 2.4x faster than real time on a 4-vCPU cloud machine and says it beat MMS-TTS on intelligibility in 21 of 27 languages, both the authors' own tests. The first download is about 440 MB. The code is MIT but the model is CC BY-NC 4.0 by the Hugging Face API, so narration for a paid client is out. A free demo Space is running (created November 6). For a documentary editor working in Twi, Yoruba or Swahili, this is a free scratch narrator worth an afternoon.

Open and local

The local story is HyperFlow: a MiniMax H3 descendant that fits a 24 GB card with picture and sound together, at about two and a half minutes per 5-second draft on a 4090. Readers outside the excluded territories can test it today. Everyone else gets a preview of where home video is heading.

Video-Rebirth/hyperflow-h3-flash: video plus stereo sound in 4 to 8 steps on a home GPU. Repo opened November 6, 152 stars by the GitHub API (repo).
porfeng/ComfyUI-PhotoDeblur: free camera-shake repair inside ComfyUI. Created November 7, 77 stars (repo).
jtydhr88/ComfyUI-InstantHMR: 3D body capture from photos, video or a live webcam that can drive a rigged character. Created November 7, 51 stars (repo).
0xMomo-NGClubs/retro-cctv-video: fixed-camera found-footage skill with prop and causality checks. Created November 6, 53 stars (repo).
SallyRS/video-director: stills-first short film direction for Claude Code. Created November 6, 26 stars (repo).
liangdabiao/whiteboard-teacher-video-motion: self-drawing explainer videos with an offline sample; its engine, Frametell, is AGPL-3.0. Created November 8, 29 stars (repo).
AfriSpeech/africa-tts: CPU voices for 92 African languages, non-commercial weights. Created November 6, 21 stars (repo).

InstantHMR deserves its own line. The upstream InstantHMR shipped v1.0 on November 6 and v1.1 on November 8: a 20 to 32 million parameter body-capture model that runs in about 1 ms per person on a desktop GPU, per the README (0.70 to 0.98 ms on an RTX 4070). The ComfyUI node adds a Live Capture mode where a phone on the same network becomes the camera, and the pose can drive a .vrm, .fbx or .glb character or export an animated GLB. Read the license table: four sizes are "SAM License, non-commercial." Only medium-commercial is cleared for paid work, and it is the least accurate of the five (72.2 mm error against 48.8 mm for the default medium on the EMDB-1 benchmark, lower is better).

Creative workflows

1. Stills-first short film with Video Director (repo, skill in skills/video-director/)

The steps: Install in Claude Code with /plugin marketplace add SallyRS/video-director then /plugin install video-director@video-director. Answer the brief questions: purpose, audience, tools, whether to study reference films. Pick a direction from 2 to 3 labeled stills of the same moment. Open the HTML storyboard page, flip through each scene's frames, pin numbered notes, and copy them back to Claude. Approve the motion spec and sound plan. Watch the animated test, then the full HyperFrames build, then the render checks.

How it works: Every decision happens on cheap stills before expensive motion. The motion spec names each move with timing and easing, and the sound plan is written at the same moment, so music and effects land on cuts instead of being laid under them later.

Why it is good: The expensive mistake in AI video is animating the wrong idea. Here the wrong idea dies as a still. The README's example is a real 52-second ad.

Where it breaks: You need Python, ffmpeg and HyperFrames installed; generated footage needs a separate clip tool, and the README names Higgsfield. Music comes from Suno, which needs "a paid plan for anything commercial." Free sound libraries are not bundled. Without the optional tools, the skill "says what it can't check."

2. Compare-before-you-trust photo deblur (ComfyUI-PhotoDeblur.json)

The steps: Clone into ComfyUI/custom_nodes, install requirements.txt and requirements-download.txt, restart. Drag in the example workflow. Load a developed sRGB TIFF or PNG. Start at variant RealBlur_J, tile 768, overlap 96, TTA off. Run MISCFilter and HI-Diff on the same photo and inspect text, faces and edges at matching zoom. Try RealBlur_R and GoPro variants next. Save 16-bit through Save Image (Advanced).

How it works: Each model was trained on a different set of blurred-and-sharp photo pairs, which is what the J, R and GoPro variants mean. Tiling keeps full resolution without running out of memory. An optional tone-match node pulls color back toward the original.

Why it is good: It treats deblur as a choice between results, not a magic button, and the README tells you which photos are worth trying.

Where it breaks: Defocus and missing detail stay missing. Intentional background blur gets "fixed" too, so mask the subject back in your photo editor. TTA averages eight passes and costs roughly 8x the compute.

3. Found-footage continuity checklist (retro-cctv-video README)

The steps: Break down the reference. Write every beat as trigger, attempt, action, contact, result. List props with counts and who holds them. Build character and prop references. Convert to model prompts. Generate with a fixed wide camera and no text in frame. Re-roll only failed segments. Add the timestamp on its own low-resolution layer, outline it, soften it, scale it up and lay it over.

How it works: Surveillance footage reads as real because nothing is staged for the camera. The checklist forces the story into actions visible from far away, and keeps the camera from doing anything a bolted-down camera cannot.

Why it is good: The rules came from real failed takes the author lists: figures too big, rescues that skip to the answer, stone counts that change, breakage that looks like foam.

Where it breaks: Docs are Chinese only. The bundled client covers RunningHub and MiniMax H3 only. Fake CCTV of real historical events can mislead; label it.

Worth testing

Africa TTS demo Space (try it): free and running. Tradeoff: non-commercial weights, and the authors ask for native-speaker feedback because it is still improving.
ComfyUI-PhotoDeblur on your worst shaky night photo. Tradeoff: no hosted demo, needs a working ComfyUI and some NVIDIA setup for MISCFilter.
InstantHMR Live Capture with your phone as the camera. Tradeoff: only the least accurate size is licensed for commercial work.
Video Director on a 15-second social short you already planned. Tradeoff: the full chain needs HyperFrames plus paid clip and music tools for commercial output.
HyperFlow H3 Flash if you live outside the EU, UK, Korea and US. Tradeoff: 24 GB card minimum as measured, and about 14 minutes per 15-second clip on a 4090.

What actually matters from today's signal

The best releases this weekend are checklists wearing code. Video Director makes you approve stills before motion. Retro CCTV counts stones. PhotoDeblur tells you to run two models and look. None of them makes a better pixel; all of them stop you from paying for the wrong one. For a working creator that matters more than a new model, because the real cost of AI video is not the per-second price, it is the rerolls you burn on an idea that was wrong at the storyboard.

The counter-signal is licensing, and it keeps getting sharper. HyperFlow is a real step for home video with sound, and it is closed to the US, the UK, the EU and Korea without MiniMax's permission. Africa TTS is free code over non-commercial weights. InstantHMR's commercial model trades accuracy for clean training data. Read the license table before the README's speed table, every time.

Adobe MAX opens tomorrow. Expect Firefly and partner-model news to dominate the week, and expect it to be paid and cleared for commercial use, which is exactly the gap this weekend's free tools leave open.

Source access notes: Hugging Face pages and API, and every vendor blog, were unreachable by direct fetch from both shells (egress blocked); HF dates and licenses came through a separate web reader using the API (createdAt), and GitHub data from the GitHub API and raw.githubusercontent.com. shields.io was blocked, so star counts are GitHub API counts. HN Algolia blocked; Trendshift and github.com/trending not checked. OpenAI news returned an error; Midjourney updates and ByteDance Seed returned empty pages; DeepMind, blog.google and Ideogram listings carry no dates, so nothing could be confirmed in range. Runway (newest October 27), BFL (October 15), Stability (August 25), ElevenLabs (September 28), Suno (November 4), fal blog (October 20) and Replicate (November 3) had nothing in the window; fal model listings refuse automated readers. Civitai and blog.comfy.org skipped. Africa TTS model createdAt could not be read (oversized API response); dated by its GitHub repo and Space. Adobe MAX 2026 dates (November 10 to 12, Miami Beach, preconference November 8 to 9) from max.adobe.com and Adobe's November 5 sessions post. Fact-check pass: an adversarial pass checked about 40 claims and caught seven problems, all fixed: Recraft's Nano Banana prompt count (five, not twelve) and its credit pricing (the post does state it); a HyperFlow download figure that was the page's monthly count, not the API's; two unsupported "first" superlatives (HyperFlow on 24 GB, Africa TTS); InstantHMR's 1 ms figure sourced to the README, not the paper; a truncated Video Director quote; and a translated retro-cctv line wrongly set in quote marks.