FervorCreative AI
Live Latest 18.11.26 · morning 401 tools tracked 992 workflows indexed 626 topics Hot: Gaussian splatting, ComfyUI, ZONOS3-Mini

The week's best animation tools start from one finished drawing and build the moving version around it, so the artist's own pixels stay the master and AI only paints what the drawing never showed.

image-to-psdspine-heroline-art-animationAdobe MAX SneaksRecraft Vectorizer ProStable Audio 3Toyxyz Blender Retargetercreative-workflowsimage-gendesign-toolsaudio-genlocal-creative-aiopen-weights

Creative AI Briefing: Thursday, November 12, 2026

You can now hand a coding agent one finished anime character illustration and get back a Photoshop file in 19 standard layers, with bangs, back hair, eye whites, irises and lashes each on their own layer and the hidden parts behind them filled in, ready to rig for 2.5D animation. A day and a half after that skill appeared, a game developer published the other half of the job: a method for turning one painted hero into a skeleton-driven character whose moves are written as code and checked by a script before a human sees them. The pattern across the open releases this week is that animation now starts from one finished picture and grows outward from it. Your drawing stays the master file. The AI is there to paint what the drawing never showed: the forehead under the bangs, the far arm, a second weapon. That keeps the style yours. It also moves the hard part of the job from drawing frames to checking them.

New models

Nothing qualifying shipped from a major lab in the window. The one model item is second-day work: Stability's Stable Audio 3 now has a community Mac version (see Audio and music). Its original weights date to May 2026.

Image

Adobe's MAX Sneaks showed where the editing tools go next, as research only. Adobe's Sneaks post (November 11) lists ten projects. The ones that matter most for illustrators and retouchers: Project Deep Brush, where brushstrokes follow the surfaces in a photo and paint light onto them; Project Rinse Repeat, which copies an edit from one part of an image to another and keeps it editable with normal Photoshop tools; Project Hands On, which puts on-canvas controls on scene elements; Project Sight Span, which turns one image into an environment you can move through beyond the frame; and Project Link Up, which drags design elements while their perspective, scale and style adapt, producing editable vectors. None of these is a product. Adobe is opening some of them as "early research previews" in the Adobe Labs Playground, and says the research "powers the next generation of Adobe tools." Plan nothing around a Sneak. Watch which ones land in a beta.

Recraft's Vectorizer Pro fixes the three things auto-trace always got wrong. Recraft's post (November 11; the tool has been live since October 29) says it keeps "straight lines stay straight, corners stay sharp, and curves stay smooth," and shows the standard vectorizer bending parallel hatching, rounding icon tips into blobs, and adding bumps where curves meet straight edges. It is "available to everyone in Recraft Studio, on every plan," takes PNG, JPG or WebP, and exports SVG, PDF or Lottie. No credit cost is stated. Photos and heavily textured art remain a weaker fit, per the post, so simplify those first.

Video

Four of the ten Sneaks were video. Project Scene Sweep applies one edit across every cut and camera angle by recognizing characters, locations and scenes. Project Live Set changes a video's subject, environment or camera direction while it plays. Project On Cue controls the timing of actions in generated video with an editable timing signal, which is the control generated video has always lacked. Project Screen Test builds a character's look, voice and personality so you can talk to it. Creative Bloq's roundup (headlined "5") walks through six of them. Same caveat: research, no ship date.

A ComfyUI pack for MiniMax H3 claims HD video on a 6 GB card. alinuo666/ComfyUI-Banzhang-All (created November 12, GPL-3.0) is a Chinese-language node pack with multi-segment continuation, a "director's desk," memory management and a music-video mode that lays four continuation segments against an audio waveform. Its headline is a three-pass sampling workflow and the line "6GB 显存也能跑高清视频" (HD video even on 6 GB of graphics memory). The README gives no resolution, frame count or timing behind that claim, the model files come through a single Quark cloud-drive link, and some nodes stay inactive without proprietary NVIDIA or Topaz engines. MiniMax H3's own license carries regional and revenue limits, as covered in earlier briefings.

Audio and music

Stable Audio 3 runs on a Mac now. mlx-community/stable-audio-3-medium went up on November 11 (API createdAt 2026-11-11T20:33:23Z). It is a conversion of Stability's own Apple-format export in stabilityai/stable-audio-3-optimized (created May 18, 2026). The card says it "generates stereo 44.1 kHz music and sound effects from a text prompt," up to 380 seconds, in a 5.79 GB download. Stability's card says the model runs in "less than a few seconds on a MacBook Pro M4." Neither card gives a memory figure. The catch: the mlx-audio support it needs "is in review and ships in its next release," so it does not run out of the box today. The license is the Stability AI Community License; you must keep the NOTICE file and display "Powered by Stability AI," and commercial terms live on stability.ai/license, which we did not read this run.

ElevenLabs' NYC Summit was for businesses, not creators. The summit post (November 11) covers Scribe v3 Medical, agent memory and routines, a partner program and a marketplace "starting next year." Nothing new for voice, music, sound effects or dubbing.

Open and local

The local story this week is character animation that starts from a finished drawing. link-ding/image-to-psd splits one illustration into a layered PSD. enoch3712/spine-hero-skill documents how a game team turned one painting into a rig and wrote six moves as code. JDhruv14/line-art-animation skips images entirely and draws ink figures live in code. All three are agent skills, folders of instructions a coding agent such as Claude Code or Codex follows. Stars below are GitHub API counts at run time (shields.io was unreachable). A note on noise: GitHub's new-repo search for November 10 is flooded with near-identical "Windows software" repos at exactly 102 stars, created within minutes of each other, plus a Blender "texture forge" with the same download-a-ZIP pattern. We excluded all of them.

  • JDhruv14/line-art-animation: a skill and small library for ink-on-paper SVG figures that walk, ride and react, with a day-to-night palette and frame-exact video capture. Created November 10, 516 stars, MIT (repo).
  • link-ding/image-to-psd: one character illustration in, a PSD in 19 standard layers with separate pixels and an editable mask per part out. Created November 10, 73 stars, no license file (repo).
  • enoch3712/spine-hero-skill: animate a 2D game hero from one painting on a Spine skeleton, with a script that fails sliding feet before a person looks. Created November 11, 31 stars, no license file (repo).
  • toyxyz/toyxyz-blender-retargeter: moves humanoid motion between rigs with different leg lengths in Blender 5.0, adjusting stride and foot contact instead of copying rotations. Created November 12, 20 stars, GPL-3.0 (repo).
  • op7418/guizang-game-skill: a skill that builds a playable 3D action game in the browser from scratch, with a demo. Created November 11, 165 stars, AGPL-3.0 (repo).
  • QQTangBa/cangshu-koubo-video-skills: two Claude Code skills that cut flubs and pauses from talking-head footage into 9:16 or 3:4 edits, macOS only. Created November 10, 41 stars, Apache-2.0 (repo).
  • alinuo666/ComfyUI-Banzhang-All: the low-memory MiniMax H3 pack above. Created November 12, 34 stars (repo).

Creative workflows

1. One illustration to a layered PSD for 2.5D animation (link-ding/image-to-psd, skills/image-to-psd/SKILL.md, example examples/blue_maid.psd)

The steps: Install with npx skills add link-ding/image-to-psd --skill image-to-psd -a codex -g -y (swap codex for your agent), then pip install pillow numpy psd-tools (add cairosvg for SVG input). Give the agent the character image and say "Use $image-to-psd to split this character artwork into a layered PSD." Open the result in Photoshop, Krita or Clip Studio Paint.

How it works: The agent studies the art and marks where each part ends and what covers what. It generates part reference images with an image tool (ImageGen in Codex), cuts the visible pixels from your original, and fills only the hidden areas: the face under the bangs, the hair behind the head, the eye whites under the irises. It exports each of 19 named parts as its own layer with its own mask, then reopens the file and checks it layer by layer for bleed, gaps and stray fragments.

Why it is good: The visible pixels are yours. Style, colors, pose and canvas size stay unchanged. Parts that are missing stay as empty layers, so every character gets the same layer order, and the export script also works alone if you already have part PNGs.

Where it breaks: The README says it "does not promise one-click, production-grade rigging quality," and the hidden areas are guesses the agent tells you about. It keeps only the mouth as drawn and does not add vowel mouth shapes for lip sync. The instructions are in Chinese (the agent follows them in any language). It depends on an image generator, so expect that tool's cost. There is no license file, so ask the author before building a product on it.

2. One painted hero to a gated, skeleton-driven move set (enoch3712/spine-hero-skill, skills/spine-hero/references/rig.md, scripts/gate.mjs)

The steps: Cut the painting into parts on a Spine skeleton (the original used Spine 4.2; you write the rig builder or build it by hand). Get a human to approve the rest pose. Write a capture script that measures each clip. Then, per move: generate a reference video of the hero for timing only, have an image model paint missing pieces, let a coding model write the move as keys on the rig, run node gate.mjs on the capture, have a second model check close-ups against defects.md, and show the result to a human.

How it works: The gate measures sliding feet, ground speed, cracks and a move that does not return to rest. No model judges those numbers. The defect list holds 13 problems a human caught that the checks missed, such as a far boot as bright as the near one.

Why it is good: The original replaced a flipbook of 403 frames cut from generated videos with one painting on a skeleton, so swapping a hammer for an axe no longer means redoing ten clips.

Where it breaks: Built on one game, one art style and one runtime. The rig builder and capture are not included and are most of the setup. The author's own arithmetic puts one clip at about 0.5 million tokens if it passes first time and up to about 2.5 million if it does not.

3. Line-art figures drawn live in code, rendered frame-exact (JDhruv14/line-art-animation, examples/walker/, scripts/capture.ts)

The steps: git clone https://github.com/JDhruv14/line-art-animation ~/.claude/skills/line-art-animation, then ask Claude for a scene ("a bakery at dawn: a baker carrying trays, steam from the oven"). It shows one figure for approval before building the rest. Render with node scripts/capture.ts and ffmpeg.

How it works: The walk is solved, not keyed: the planted foot slides back at exactly the travel speed, so feet never skate, and the body rises and dips with each step.

Why it is good: Vector output stays sharp at any size, and the capture steps one frame at a time, so a slow machine still renders a smooth video.

Where it breaks: One house style, ink on paper. Needs Node 22.18 or newer and Chrome. Revisions go through the agent, not a timeline.

Worth testing

  • Recraft Vectorizer Pro on a logo or hatched illustration you already trace by hand. Tradeoff: no stated credit cost, and textured art needs simplifying first.
  • image-to-psd on one character you own. Tradeoff: hidden areas are invented, and you still rig it yourself.
  • The spine-hero gate alone: node --test skills/spine-hero/scripts/gate.test.mjs shows what it measures before you commit. Tradeoff: its limits were calibrated on one game at one capture size.
  • Toyxyz Blender Retargeter on a mocap clip and a stylized short-legged rig. Tradeoff: Blender 5.0 or later only, and the README says hand contact is "currently limited."
  • Stable Audio 3 on a Mac once mlx-audio's next release ships. Tradeoff: not runnable from the stock release today, and read the commercial license first.

What actually matters from today's signal

The drawing is becoming the source file again. For two years the cheap way to animate a character was to generate frames and live with the drift. This week's best releases do the opposite: keep the artist's picture, cut it into parts, and use AI only for the pixels nobody drew. That is good news for illustrators, because the thing that makes a character look like yours, the line and the paint, never passes through a model.

The cost moved to checking. image-to-psd reviews its own file layer by layer. spine-hero runs a script, then a second model, then a person, and keeps a written list of every defect the machines missed. Neither claims one-click results. If you adopt this way of working, the skill you need is a precise eye for seams, sliding feet and borrowed brightness, plus the patience to write down what you catch.

The risk is the paperwork. Two of the three headline repos have no license file, and the Stable Audio port needs a release that has not shipped. Free and open are not the same. Read the license before a client sees the result.


Source access notes: Hugging Face's createdAt-sorted lists for text-to-video, text-to-image, text-to-speech and image-to-video again returned stale entries (newest September 2 or November 4), so discovery leaned on image-to-image and text-to-audio lists, the trending list and GitHub search. No creative model in the Hugging Face trending top 40 was created on or after November 9. MiniMax H3-Image, cited in The Neuron's November 11 digest, predates the window (PrunaAI's H3-Image repackage dates to October 27). GitHub data came from the GitHub search API and raw.githubusercontent.com; shields.io, hn.algolia.com and the Hugging Face API were blocked from the main workspace, so Hugging Face reads went through a separate web reader. openai.com/news failed; seed.bytedance.com and midjourney.com returned only metadata; blog.comfy.org showed a shell; Ideogram, PetaPixel, RedShark and MusicTech show no dates; runwayml.com/news showed no articles. BFL (November 8), Stability (August 25), Luma (October 15), Suno (November 4), Replicate (November 3), fal (October 20), Krea (November 3 to 5), Pika (October 29) and ComfyUI releases (v0.43.0) had nothing in the window. DeepMind and blog.google showed no dated November items. Adobe MAX launch items were covered yesterday and are not repeated. Spotify's new AI DJ voices are listening features, omitted. Video and image sections hold research previews only; New models reduced to a note. Fact-check pass: an adversarial pass checked the draft against primary sources and caught four problems, all fixed: the Sneaks were not "mostly video" (four of ten); the Playground holds only some Sneaks, and the two Adobe quotes come from separate sentences; Creative Bloq's "5" roundup walks through six projects; the gap between the image-to-psd and spine-hero repos is about a day and a half, not two days. It also suggested "19 standard layers," since an optional 20-layer layout exists.

Articles from this briefing