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

spine-hero's real lesson for animators is the order of judgment: a measuring script rejects bad motion before any AI reviewer or human sees it, and every defect a person catches becomes a written rule, so quality compounds instead of resetting with every clip.

spine-heroSpinecreative-workflowsimage-genvideo-genai-editing

Animate a 2D Game Character From One Painting: The spine-hero Workflow Where a Script Gets the First Veto

A game team threw out 403 frames of AI-generated animation and rebuilt its hero from a single painting on a skeleton. The method they published is less about AI than about who gets to say no, and in what order.

There is a cheap way to animate a 2D game character right now. Paint the hero once. Feed the painting to an image-to-video model with a prompt like "swings hammer." Cut the frames out of the clip. Repeat for the walk, the jump, the taunt. Before long you have a full move set.

Then the art director wants a different hammer.

That is the moment the cheap way falls apart, and it is the opening scene of spine-hero, a method its author posted to GitHub on November 11. Their hero started as exactly that flipbook: ten clips, 403 frames cut from generated videos, with the weapon, shield and armor baked into every frame. The README's caption on that slide says it all: "Change the hammer? Redo all 10 clips."

The fix they landed on is older than AI: cut the painting into parts, put the parts on a skeleton, and animate the skeleton. What is new is how they used AI around that skeleton, and in particular the rule that decides who gets to reject a clip first. That rule is the part worth stealing, even if you never touch their code.

The idea in one paragraph

Spine, from Esoteric Software, is a 2D animation program where you attach cut-out pieces of a drawing to a skeleton of bones and animate the bones instead of redrawing frames. It is popular for 2D game characters. spine-hero keeps that structure and hands parts of the work to AI models, each with a narrow job. An image model paints what the painting never showed. A video model shows timing. A coding model writes the motion. Then a plain script, which the method calls the gate, measures the result and rejects it if the feet slide, before any model or person looks at it. A second, independent model checks close-ups against a written list of past mistakes. A human has the final say, and every mistake the human catches becomes a new line on the list or a new check in the script.

Why a skeleton beats a flipbook here

The payoff is shown in one of the README's comparison clips: the same leap played with a hammer, an axe and a mace, and the same shield move with two different shields. Because each piece of gear rides on its own bone, swapping it changes one drawing, and every move still plays.

It also shrinks the problem generated frames always have. A video model redraws the character on every frame, so the face, the armor trim and the brush texture drift a little each time. On a skeleton, the visible pixels are the painting's own, moved. Drift can only creep in through the pieces an image model paints, which is exactly why the method checks those pieces so hard.

The cost is setup, and spine-hero is honest that it is large.

What you need before you start

The README lists what you must bring, and it is worth reading as a budget:

  • One finished painting of the hero, side view, standing at rest.
  • Spine or a similar skeletal animation tool. The original used Spine 4.2. Check the edition: Esoteric's purchase page listed Spine Essential at $69 and Spine Professional at $379 when I checked (shown as reduced from $99 and $449), and states that Essential "is unable to save or export projects containing Spine Professional features, such as meshes or IK." This method depends on both. Meshes are what let a coat bend between hip and hem instead of moving like cardboard. IK (inverse kinematics) is what keeps a foot pinned to the floor while the body moves over it. So if you build or edit the rig in the Spine editor, budget for Professional. (The original skipped the editor and wrote the skeleton file with a script, which only a programmer will want to copy.) There is a free trial, but per the same page, buying is what "enables saving projects and exporting."
  • A way to cut the painting into parts and build the skeleton. Not included. The author's scripts were written for their own game. The rig guide describes the method, or you can build it by hand in Spine.
  • A way to capture frames from your game engine and write the measurements to a file. Also not included, and the README says to build it before the first clip, because "nothing can be gated or reviewed without it."
  • An image model that can take your painting as a reference, an image-to-video model if you have one, and a coding assistant (an AI that writes and runs code on your computer, such as Claude Code, Codex or Cursor) to run the skill, which is a folder of written instructions the assistant follows.

If that list made you wince, good. The skill and the gate script are free. The rig builder and the capture are most of the work, and they are yours.

Put this into practice: the loop for one move

Here is the sequence for a single move, using the walk the README walks through.

1. Approve the rest pose (the hero standing still) before anything moves. Render the rigged hero standing still next to the original painting. The rig guide says to stop here and get a human sign-off: no visible seams, cuts or color shifts. Every clip inherits this pose, so a flaw here repeats in every move.

2. Get a timing reference, not frames. Have a video model animate the hero's own painting through the move. You use the clip only to read beats: when each foot lands, how long a step takes, how high the swing goes. The README is firm that the clip "is never traced or used as frames." If you have no video model, you write the beats by hand.

3. Paint the parts the painting never showed. A side-view painting hides the far arm, the inside of the far leg, the back of the coat. Have an image model paint those as separate pieces. The rig guide has a sharp warning here: a hidden far limb gets its own paint, "never a toned copy of the near one," because a copy reads instantly as a clone.

4. Let the coding assistant write the move as motion on the skeleton. Not frames. Keys on bones, with the feet held in place by IK. It builds the clip, captures it, and looks at its own result.

5. Run the gate. This is the step that makes the method work. A small script, gate.mjs, reads the measurements and checks for sliding feet, wrong ground speed, cracks between parts, a move that does not return to rest, and anything drawn outside the character's box. The README's example shows a walk failing on a single foot slide: "FAIL example-hero walk: 7/8 checks." No AI model judges these numbers. They either pass or they don't.

6. Check close-ups against the defect list. A second model, separate from the one that wrote the move, goes through defects.md item by item on enlarged crops of the feet and head. The list holds 13 problems a human caught that the checks missed, each with how to spot it. One example: a far boot painted as bright as the near one, when the far side should sit in shadow.

7. You decide. And anything you catch becomes line 14.

You can see step 5 work before you build anything. The repo ships a sample measurement file, and node --test skills/spine-hero/scripts/gate.test.mjs runs the gate's own tests. That costs nothing and shows you exactly what the script considers a failure.

The rule worth stealing

Strip away Spine and the specific models and one idea is left: put the cheapest, most objective judge first.

A script cannot be talked into anything. It does not get tired at clip nine, and it does not decide a slide is "basically fine." So it goes first and kills the obvious failures for almost nothing. The AI reviewer goes second, and it is pointed at a written list rather than asked whether the clip "looks good." The human goes last, with the easy problems already gone, and spends their attention on the things only a person sees.

The written list is the other half. Most review notes vanish into a chat thread. Here every catch becomes a permanent check, so the next clip gets reviewed by everything you have learned so far. The defect list reads like a senior animator's notebook: rigid coats that should trail the legs, a generated far boot whose sole matched the near one but covered only 67 percent of its area when planted and 48 percent mid-swing, which "the human saw at once," parts floating free of their limbs.

You can adopt this without spine-hero. If you animate by hand, write down every note your lead gives you, in the form "what went wrong, and how to see it." Check each new clip against the list before you send it. That is the method, without the software.

Where it breaks

It was built for one game. Side-view painted heroes, a web game, Spine 4.2. The author says it has not been tried with other engines or art styles. The gate's pixel limits were set at one capture size, and the gate guide explains which limits carry over and which you must set again.

Most of the setup is not in the box. The rig builder and the frame capture are yours to write. Those are the hard parts, and if you are not comfortable with scripting, you will need someone who is.

It costs AI model usage, and the bill can grow. The author's own arithmetic, which they label as arithmetic and not a measurement: a clip that passes on its first round uses about half a million tokens of model usage, the unit AI providers bill by. Each failed round adds up to about 0.4 million, and with six rounds allowed one clip can reach about 2.5 million. An idle, a walk and one move come to about 1.5 million if everything passes first time, and "several times that" if not. The image and video models are billed on top.

There is no license file. GitHub shows none on the repository, which means the author has not formally granted reuse rights to the scripts. Read it, learn from it, and ask before you ship the code inside a product.

The human is not optional. The README says it directly: "The checks miss things: the defect list exists because they did." This is a method for making a person's judgment go further, not for replacing it.

What to do this week

You do not need to adopt the whole pipeline to get something from it.

If you already animate in Spine, read defects.md and moves.md tonight. They are short, and the move guide's advice on when to animate a pose on the skeleton versus when to use a few held drawings for a fast blow is good craft on its own.

If you have been cutting frames out of generated video, count your clips and ask what one gear change would cost you. If the answer is "redo everything," the painting-on-a-skeleton route is worth the setup, with or without AI.

And whatever you animate with, start your own defect list today. Put the first line in it the next time something looks wrong and you cannot say why at first. Your future reviews will thank you, and so will the art director who wants a different hammer.


From the briefing