Creative AI Briefing November 7, 2026 morning edition 11 min read

The useful releases this week put an AI assistant in the operator's chair of real tools, so a creator's job shifts from prompting pixels to giving notes on an editable scene, a shot list or a timeline.

vork Krea Agent Seedance 2.5 GPT Image 2.6 Pika H3 Flex Stable Audio 3 Chatterbox Multilingual 3d-gen creative-workflows video-gen local-creative-ai audio-gen image-gen

Creative AI Briefing: Saturday, November 7, 2026

You can now ask an AI assistant to build a 3D scene in code, drop it into a tracked shot of real footage, render it on your own graphics card and revise it from your notes, with no generative video anywhere in the chain. That is vork, a free research-preview renderer that went public on November 5, and it fits a pattern running through the week: the twelve stills behind Krea's one-minute monkey film came out of five short messages to an agent, Luma and Kapwing now take orders from inside Claude and ChatGPT, and the open community spent Wednesday through Saturday shrinking voice and music models onto Macs. No major video model launched. The tools that shipped give you something a prompt box never did: a result you can open, change and render again.

New models

vork turns an AI assistant into a 3D and VFX operator. The voyager-so/vork repository was created November 5 (22:37 UTC) and tagged its first release, vork 0.1.0, on November 6. The README calls it "a headless 3D and VFX engine built for agents": no editor window, just a renderer your assistant drives. "An agent writes a scene, renders it, looks at the result and revises the code." Scenes are plain JSON files (with Python and TypeScript kits), and output is PNG stills or MP4 movies. The examples include a "VFX on real footage" shot that adds a ceramic-and-bronze object to a Pexels clip using a tracked camera, occlusion mattes and stand-in geometry for relighting: "Real footage + rendered 3D; no generative video." The license is source-available, not open source: free "including commercial work" during the preview, but "this version only," so later versions may cost money, and you may not resell it or run it as a hosted service. It needs a recent NVIDIA card with hardware ray tracing on Linux or Windows (data-center cards like the A100 and H100 lack it), or Apple Silicon on macOS 26. The honest catch is in the README's own words: fog and volumes "are ignored with warnings," "Large scenes and dense transparency can fail," the first render of a new program can take "about 3–4 minutes" to compile on NVIDIA, and the Blender helper "currently fails during camera preparation on Blender 5.2.1." Its speed tables beat Eevee by 1.38x to 1.66x on stills and 1.14x to 2.32x on movies, but the README says these are "not equal-quality speedups," and its path tracer ran 5.81x slower than Cycles on one test. No hosted demo exists. Stars: 573 on shields.io, 288 on GitHub's API at our checks; treat both as rough.

Image

GPT Image 2.6 Sunburst is free in Recraft, with a catch for paid work. Recraft's November 5 post puts OpenAI's new image model (covered here on November 5) into Recraft Studio as two versions: Flare for speed, Sunburst for consistency across edits. Sunburst runs on the free plan at default settings; Flare and the quality and 2K controls need a paid plan, at 4 to 57 credits an image. Read Recraft's terms (updated October 21) before you use that free tier for a client: "no commercial use of Free Tier Assets is permitted," and those images "are owned by Recraft." Studio also outputs raster only, so the SVG export OpenAI shipped lives in ChatGPT and the API, not here.

A two-step version of P-H3-Image, for people with very large cards. Pruna's P-H3-Image (created October 27) is an image model built on MiniMax H3. On November 7 akhaliq posted a 2-step add-on (Apache 2.0 file) with a ComfyUI workflow, p-h3-image-pdmd-2step-t2i.json. The author reports "~2.9 s per image" at 2912 x 2912 pixels on an RTX PRO 6000 or H200, about "~7 s per image all-in" with the optional detail pass, and memory of about 83 GB (76 GB with offload), so this is a workstation story. Outputs fall under the MiniMax H3 Community License, which "excludes use in the EU, UK, Korea and the US" without a separate license. The card warns that "composition drifts from baseline, graphic prompts can invent pseudo-text." Try it in the free Space.

Video

Krea published the full recipe for a one-minute film made from one phone photo. The Zoo York guide (November 5) shows Krea Agent turning a snapshot of a monkey with a laptop into twelve 5-second Seedance 2.5 clips. The whole recipe, including its failures, is in the workflows section below.

Pika H3 Flex now has a tested local workflow. Thelocallab's ComfyUI workflow (created November 5) takes "one to three pictures and a shot-by-shot prompt" and returns a scene with camera moves, dialogue, sound and music "in six sampling steps." It is second-day work on the October 29 Pika release, and the useful part is the author's timings on a 20 GB RTX A4500: 1 minute 43 seconds for a 5-second clip, about 11 minutes 15 seconds for 15 seconds with the 2x upscale. A 5-second clip fit under simulated 12 GB and 16 GB limits; 15 seconds peaked near 17 GB. Budget a 48 GB download and an NVIDIA driver 580 or newer. The MiniMax H3 Community License applies ("Read it before commercial use").

Luma moved into your chat window. Luma's MCP connector (November 3) brings Luma Agent, product shots, camera angles, background swap, Relight Studio and short promo clips into Claude and ChatGPT. It "uses your Luma credits"; the page lists no plan or price.

Audio and music

Stable Audio 3 runs on a Mac in about 5.5 GB of files. MusicForge posted an INT8 MLX build of the Medium model of Stability's May release on November 4 for Apple Silicon, alongside a 4-bit Small build. The card promises text-to-audio, "mask-based inpainting and continuation," and lengths "up to several minutes." It holds weights only; you run them with Stability's own MLX code. The Stability AI Community License allows commercial use "up to the revenue threshold defined by Stability AI." No speed or memory figures are given, and no demo exists.

Chatterbox voice cloning got faster on Apple Silicon. An 8-bit MLX build of Resemble's 23-language Chatterbox Multilingual V3 (created November 7, MIT) clones a voice from a short clip. On a base M5 Mac the author measured one sentence in 5.0 to 6.0 seconds against 8.0 to 10.9 for the full-size file, peak memory 3.5 GB, and warns timings "varied up to 2x" under memory pressure. The card's rule is the one to keep: "Use only your own voice or one you have permission to use."

Open and local

The local story this week is Apple Silicon. A music model and a cloning voice model both arrived as compressed MLX builds within three days, and vork's Mac path works on any Apple Silicon machine running macOS 26. All of these are community or preview builds with one author's numbers, so test on your own machine before you plan around them.

voyager-so/vork: an agent-driven 3D and VFX renderer that writes scenes as code and composites over real footage. Created November 5, first release November 6 (repo).
Thelocallab/Pika-H3-Flex-ComfyUI-Workflow: the local route to Pika's reference-to-video model with real timings on a 20 GB card. Created November 5 (card).
MusicForge/stable-audio-3-medium-8bit: Stable Audio 3 Medium compressed for Macs (about 5.5 GB), with a 4-bit Small sibling. Created November 4 (card).
roman220220/chatterbox-multilingual-v3-mlx-8bit: 23-language voice cloning on a Mac in 3.5 GB of memory. Created November 7 (card).
akhaliq/p-h3-image-pdmd-2step: 2-pass image add-on with a free Space, for workstation cards (about 83 GB, or 76 GB with offload). Created November 7 (card).

Creative workflows

1. One phone photo to a one-minute found-footage short, from Krea's Zoo York guide

The steps. Start from one real phone photo of your subject. Make 12 stills with GPT Image 2.5 at 16:9 and 2K, each ending in the same shared line: "cheap phone photo, taken from across the street, overcast winter dusk, muted colors, monkey small in the frame, empty New York street, no people." Use each still as the start frame of a 5-second Seedance 2.5 clip at 1080p, with a four-line prompt: "Shaky phone footage from across the street, 90s home-video feel." Then one gag. Then "Fast zoom in on the monkey." Then "Audio: real street sounds, no music, no voiceover." Order the clips, cut hard ("Don't use fades"), and finish with faded warm color, softer detail and grain.

How it works. The shared line fixes the camera, the light, the distance and the street, so the only thing changing from still to still is the gag. Keeping the subject small and the look cheap hides the drift that close-ups expose.

Why it is good. The whole image set came from five chat messages, and the recipe is complete enough to copy with your own subject.

Where it breaks. The first batch of stills "do not look consistent," and the fix was switching image models. "Two clips here, including the hammock swing, needed a second try." Ask for more than one action and "the model usually drops part of it." The post names no plan or price, and Seedance on Krea has needed a paid plan in earlier guides.

2. Real footage plus rendered 3D, with vork's VFX example

The steps. Install with the README's script (install.sh on Linux, install.ps1 on Windows), then run vork --version and vork help. Give your coding assistant the README's starting prompt: "Read vork help. Make a small scene, render a preview and inspect the image." Open examples/first-scene.json to see the format, then work from examples/vfx-artifact/. Read vork help limits before you commit to a shot.

How it works. The assistant writes a scene file, renders a still, looks at it and edits the file. For the plate shot, a tracked camera places the object, occlusion mattes let real things pass in front of it, and stand-in geometry catches its light.

Why it is good. Every revision is a code edit you can read and undo, the plate stays real, and the preview license allows commercial work.

Where it breaks. No fog or volumes, failures on large or very transparent scenes, multi-minute compiles for new programs, a broken Blender helper on 5.2.1, and license terms that may change after this version. The README does not say what camera-track format the example expects.

3. Local Pika H3 Flex, short clips first, with Thelocallab's workflow

The steps. Load Pika-H3-Flex-References-SemanticBridge.json in ComfyUI, add one to three reference pictures and a shot-by-shot prompt. "Test your prompt on a short clip first": 5 seconds at 0.6 MP. Only then render 15 seconds, and upscale last.

Why it is good. No per-second fees, and the author published real timings and memory peaks instead of promises.

Where it breaks. 0.8 MP at 15 seconds ran out of memory on 20 GB; three references at 1344 x 768 needed 24 GB. The license excludes several territories. The card has no limitations section, so expect to find them yourself.

Worth testing

P-H3-Image 2-step Space: large images in two passes, free in the browser. Tradeoff: the author warns of invented text in graphic prompts and drifting people, and the territory exclusions apply to outputs.
GPT Image 2.6 Sunburst in Recraft: free access to OpenAI's newest image model. Tradeoff: free-tier images are Recraft's and not for commercial use.
vork 0.1.0: free agent-driven rendering, commercial use allowed for now. Tradeoff: research preview, no volumes, ray-tracing NVIDIA card or macOS 26 required.

What actually matters from today's signal

The agent is moving from the prompt box to the operator's chair. vork, Krea Agent, the Luma connector and yesterday's Kapwing connector all ask you to direct by notes while an assistant drives a real tool. The difference that matters is what you get back. vork hands you a scene file. Krea hands you twelve clips plus the prompts that made them. A chat that only hands you a finished MP4 leaves you back in the reroll lottery the moment a client asks for one change.

That makes "can I open it and edit it" the first question to ask of any agent tool. It also changes the cost math. Agent connectors spend your existing credits, and an assistant that renders, checks and renders again spends them faster than you would by hand. Kapwing said as much yesterday. Watch the meter on the first project.

The counter-signal is licensing. vork is free for commercial work only for "this version." Recraft's free tier gives Recraft the rights. Pika and P-H3 outputs carry MiniMax's territory exclusions. Read terms before you build a pipeline on any of them, because the cheapest tool this week may not be the cheapest tool in March.

Source access notes: Blocked or unreadable this run: openai.com/news (client error), midjourney.com/updates and seed.bytedance.com (JavaScript only), blog.comfy.org (JavaScript shell), fal.ai/models (robots.txt), several Hugging Face model pages rate-limited mid-run (Pika-H3-Flex-Turbo-LoRA, Kandinsky-S2V mirrors, Zakyrie's H3 add-ons were not read and are omitted). Hacker News Algolia returned a stale August feed, so Show HN coverage is missing. The shell had no direct network access; Hugging Face API, GitHub API, shields.io and raw files were read through a fetch tool. No new posts in the window from Runway, BFL, Stability, ElevenLabs, Replicate, fal, Pika, HeyGen or Ideogram (undated); Luma's November 5 and 6 posts were buyer's guides. Suno's "Start Songs with MIDI" (November 4) body was again unreadable. Writing and Design sections omitted. Stable Audio 3's original release date (May 2026) comes from secondary reporting and the card's paper date.

An adversarial fact-check pass ran on this draft and caught nine problems, all fixed: the vork speed range (stills and movies now split, low end 1.14x), the Krea five-message claim (it covers the stills, not the clips), Stable Audio 3's file size (5.5 GB, not 4.5), a 4-bit claim now pinned to the Small build, a weekday error, an 80 GB-card claim that contradicted the 83 GB figure, a stale star count, an unverifiable Phonon conversion count, and an overbroad "labs were quiet" line. Unverified after the pass: vork's GitHub creation timestamp and API star count (API failed for the checker), some vork example details, and the full Krea shared prompt line beyond the checker's truncated fetch (taken from an earlier full read). vork's release year is inferred from the repo creation date.