Creative AI Briefing November 3, 2026 morning edition 11 min read

With no big lab launch this week, the useful releases all hand creators more control over cheap drafts, from a three-tier video price list to an animation pipeline where every frame is written in code and can be fixed by hand.

claude-animated-shorts Overtone MiniMax H3 Lite Qwen-Image-2.1 Lightning v2 RL Anima Control Pose HeartMuLa video-gen image-gen creative-workflows local-creative-ai comfyui music-gen

Creative AI Briefing: Tuesday, November 3, 2026

A history channel can now make a 67-second animated short about a collapsing 1831 bridge without touching a video model: a free Claude skill writes the script, draws the cartoon townsfolk in code, builds the 3D town, scores the music and renders the whole thing on your own PC. That project, published November 1, sits at the center of a quiet week for the big labs and a busy one for control. MiniMax now sells the same video model at three speeds and prices, a four-step image model landed for fast drafting, an anime model learned to follow a stick-figure pose, and a caption tool cuts one transcript into every screen shape. The common thread is that a cheap draft is only worth something if you can steer what comes next.

New models

MiniMax H3 Lite turns one video model into a price ladder. MiniMax now offers H3 in four API tiers (H3, Highspeed, Lite and Max), and the two newest reached Replicate with versions dated October 29. H3 Lite does the same jobs as full H3 (text to video, first and last frame, and reference clips built from images, video and up to three audio files) at 480p or 768p for 4 to 15 seconds. It costs $0.025 per second at 480p and $0.04 per second at 768p, so a 10-second 768p draft runs 40 cents. H3 Highspeed sits in the middle at $0.06 per second, 768p only, and full H3 is $0.08 per second at 768p with 2K on top. For a creator the consequence is a budget plan: rough out shots on Lite at half the 768p price of H3, and only pay for full quality on the takes that earned it. The catch: Replicate does not state a license on these pages, Lite drops 2K and the "adaptive" aspect ratio, and nothing published says a Lite draft and an H3 final will look alike from the same prompt.

Image

Qwen-Image-2.1 in four passes instead of forty. lightx2v/Qwen-Image-2.1-Lightning-v2-RL (created November 2, 11:54 UTC) is a fast version of Qwen-Image-2.1 that makes, edits and combines images in four passes. The authors refined it with a reward-based training step after the speed-up, and the card shows side by sides against the original at 40 passes for text to image, single-photo edits and multi-photo edits. There are no published scores, only those examples. It is a full replacement for the original's picture model, so you still need the original's other parts, and it only reproduces the authors' results with the timing schedule in its inference_config.json. The card names Diffusers (installed from GitHub) as the documented route, with Draw Things and DiffusionBee listed as local apps. License: the Qwen Research License, which the card says carries non-commercial terms. Treat it as a sketching tool, not a client deliverable.

An anime model that holds a pose. GaBeng/Anima-Control-Pose (created November 2, 10:25 UTC) is a preview pose-guide add-on for the Anima anime model. Feed it a reference photo, its bundled detector draws the skeleton, and the character follows it. It ships two ComfyUI node folders, four example workflows (start with pose_control_demo.json) and a roughly 316 MB detector. At 1024 pixels the author reports pose accuracy rising from about 0.37 without the add-on to about 0.83 with it (the author's own measure; gains are smaller at 512 and 768). The honest list from the card: running, jumping and sitting poses are the least reliable, hands still fuse, and the detector was built for photos, so it reads anime art poorly. The weights are non-commercial; the card says the images you make are not restricted.

Video

Kling 4.0 is now callable on Replicate. A week after a report (secondary) that Kling's own developer page listed the 4.0 API as "not yet available", kwaivgi/kling-v4.0 runs as an official Replicate model with a latest version dated October 26. It takes text, start, middle and end frames, up to 10 reference images or a video to edit, and returns 3 to 30 seconds at 720p or 1080p. Price: $0.157 per second at 720p and $0.235 at 1080p, dropping to $0.094 and $0.141 when you feed it a video, billed on whichever is longer, your input or the output. A 30-second 1080p shot costs about $7, which makes the H3 Lite ladder above look very cheap for drafting.

Audio and music

A full open song model on a 24 GB Mac. The mlx-community published HeartMuLa-oss-3B-8bit (created November 2, 04:57 UTC) alongside separate HeartCodec repos, Mac conversions of the Apache 2.0 HeartMuLa song model, which takes style tags and lyrics and sings them. This is second-day work on a model that shipped months ago. The card says the default setup fits a 24 GB Mac and runs through a Swift package rather than a one-line command. Songs can run four minutes, with a default cap of 120 seconds, and the 8-bit version hit that cap on 10 of 24 test songs versus 5 of 24 for the larger version, so expect some songs to stop at the cap instead of finishing.

Writing

Captions that fit every screen shape from one transcript. Overtone (repo created October 28, MIT) stores each spoken word once with its timing and speaker, then cuts separate caption layouts for 16:9, 9:16 or any ratio you name. It follows reading rules: lines avoid ending on "the" or "of", names and numbers stay together, pauses become breaks. The recut command reshapes subtitles you already have, offline and with no account. Transcribing from scratch uses paid services; the author's benchmark puts it at about 2.3 to 2.5 cents per audio minute at list prices (2.9 cents with ElevenLabs Scribe as a second listener), and a --local flag swaps in whisper.cpp and Ollama. There is a browser demo. Full line rules exist only for English and Dutch, and there is no visual editor.

Open and local

The local story today is pipelines rather than models: a skill that animates without a video model, a skill pack that drives desktop 3D software, and conversions that move existing models onto Macs and new formats.

mromidev/claude-animated-shorts: a Claude skill that makes 55 to 70 second vertical cartoon shorts with three.js and Remotion, every frame drawn in code. Created November 1, about 880 stars on shields.io (repo).
scenario-labs/skills: about 80 skills for making images, video, audio and 3D through Scenario, plus about 87 helpers for Blender, Maya, ZBrush, Unity, Unreal, Godot and Roblox. Trending on Trendshift today, 1.8k stars (repo).
lonthq/overtone-captions: one transcript, captions cut for each aspect ratio. Show HN on October 29 (repo).
GaBeng/Anima-Control-Pose: pose guide for Anima with ComfyUI nodes and workflows. Created November 2 (card).
lightx2v/Qwen-Image-2.1-Lightning-v2-RL: four-pass Qwen-Image-2.1 for drafts, research license. Created November 2 (card).
mlx-community/HeartMuLa-oss-3B-8bit: open song model converted for Apple Silicon. Created November 2 (card).
OzzyGT/motionbricks_soma_diffusers: NVIDIA's MotionBricks character-motion transition model repacked for Diffusers; give it the end of one move and the start of another and it fills the gap at 30 fps. Created November 2; NVIDIA Open Model Agreement weights, and it runs downloaded code, so read it first (card).

Creative workflows

1. Make an animated explainer short where every frame is code, with claude-animated-shorts (SKILL.md, templates/bridge/, story.ts, scripts/short_qa.py)

The steps. Copy the folder into ~/.claude/skills/cartoon-3d-short/ and install Remotion with @remotion/three, Node, Python 3 and ffmpeg. Ask Claude Code for a short on one surprising true story with a visible cause and effect. Approve the script (about 3.7 words per second, hook, setting, escalation, snap, explanation by everyday analogy, callback). Record or generate voice.wav, make words.json word timings and fix them by hand. Plan shots in story.ts, check still frames, lock the picture, then build the sound from the event table. Render with npx remotion render, mux the audio, and run short_qa.py to catch black, frozen or popping frames and loudness problems.

How it works. No video model is involved. Characters are drawn by a small drawing library on a canvas, the town is built from recoloured free Kenney props, and lighting, ink lines and lens blur are code. Every reaction is timed to land just before its word in the voice track.

Why it is good. Anything you dislike is a line you can edit, not a reroll. The proven example ran 67 seconds and rendered in about 10 minutes at about 4 frames per second.

Where it breaks. The skill does not name a voice tool, so narration quality is on you. Each render bundle copies about 1.3 GB into your temp folder and never deletes it. The look is one house style; a different style means rebuilding the world and characters. And the bundled reference frames are for private study only.

2. Cut one interview into wide and vertical captions with Overtone (npx overtone-captions recut <file>)

The steps. Start with an existing SRT or VTT. Run recut with Node 22.12 or newer; it needs no key and no ffmpeg. Take the 16:9 and 9:16 files it writes and drop them into each edit, or load the .overtone.json into the video.js plugin to let the player pick the right cut.

How it works. Words are stored once with times, and each layout is a separate cut through the same words under fixed rules, so the same input always gives the same captions.

Why it is good. Vertical reels stop inheriting wide-screen line breaks. The author's small public test found 0.3% of breaks landing after a weak word with no pause, against 16.1% for ElevenLabs' SRT export.

Where it breaks. Recut estimates word timing from your existing cues, so it is only as precise as they were. Full rules cover English and Dutch only, and media is capped at two hours.

3. Draft on H3 Lite, finish on H3, using MiniMax's tiers on Replicate

The steps. Write the shot prompt and attach your references. Run three or four 480p Lite takes at $0.025 per second. Pick the take whose staging works, tighten the prompt, and run one final on full H3 at 768p ($0.08 per second) or 2K.

How it works. Replicate's own page pitches Lite for drafts and high-volume work and H3 for the highest quality; this just uses that split deliberately.

Why it is good. Four 10-second Lite drafts at 480p cost a dollar, less than two full-quality 768p takes.

Where it breaks. Nothing documents that a Lite take predicts the H3 result for the same prompt, so treat drafts as tests of the idea, not of the final frame.

Worth testing

Overtone demo: switch caption styles on 16:9 and 9:16 players in the browser. Tradeoff: the demo shows finished files; transcribing your own footage needs paid keys or a local setup.
H3 Lite on Replicate: the cheapest way into H3 this week. Tradeoff: 768p maximum and no stated license on the page.
claude-animated-shorts example: watch the bridge short before you install anything. Tradeoff: setup is heavy and the style is fixed.
Anima Control Pose: runs in your own ComfyUI. Tradeoff: preview quality, weak on action poses, non-commercial weights.
Adobe for Startups (November 2): a free year of Firefly Pro Plus for companies under five years old that raised under $15 million. Tradeoff: an application and country limits apply.

What actually matters from today's signal

The price of a draft keeps falling, and the vendors now say so out loud. MiniMax sells one model at three published prices on Replicate and a fourth tier in its own API, and lightx2v ships a four-pass image model built for speed. That changes how a working creator should spend: generate many rough versions where they are cheap, decide fast, and save the money for the shot you will deliver. If you still render every idea at full quality, you are paying for decisions you have not made yet.

The more interesting signal is the animated-shorts skill. It passed 880 stars in two days by refusing to use a video model at all. Everything in its frames is code an agent wrote, so a wrong expression or a late camera move is a fix rather than a lottery ticket. For explainer, history and education channels, that trade (one fixed house style in exchange for total editability) is worth more than photoreal motion. Expect more of these: language models writing scenes in tools like Remotion and three.js, with generative video kept for the shots that truly need it.

The counter-signal is licensing. The fast Qwen model and the Anima pose add-on are both non-commercial, Replicate does not state a license for the H3 tiers, and the animation skill's reference frames are study-only. Cheap and controllable is not the same as safe to sell. Read the card before the client sees anything.

Source access notes: Thin day for primary labs: no new creative releases found on blog.adobe.com (beyond Adobe for Startups), elevenlabs.io/blog, blog.fal.ai, replicate.com/blog, Krea news or Recraft's blog in the window. openai.com/news (client error), blog.google (no post list rendered), runway.com/news (no dated list), fal.ai/explore and cdm.link (robots), Hugging Face trending API and the dated papers page (client error), and the GitHub API for ComfyUI releases and scenario-labs/skills (client error) were unavailable; ComfyUI's docs changelog showed nothing newer than v0.43.1 (October 28). github.com/trending served a stale page and was ignored; Trendshift used instead. Midjourney, BFL, Stability, Luma, Suno, Pika, HeyGen and Udio checked via search only; no new release found. H3 Lite and Highspeed dates come from Replicate version timestamps (October 29), not a MiniMax announcement. All Hugging Face dates from the API createdAt field. Adversarial fact-check pass ran (Sonnet subagent): no wrong dates, prices or licenses found; it corrected the Anima pose baseline (0.37 at 1024, not a mixed 0.33 to 0.38 range), added Overtone's 2.9-cent Scribe option and Kling's video-input billing rule, clarified that HeartMuLa's codec lives in separate repos, and flagged a GitHub API star count (171) that disagreed with shields.io (884), so the star figure is now hedged. Bucket D (PetaPixel, Creative Bloq, No Film School) not reached this run.