I copied every animation I saw: 18 GitHub repositories and a Fable 5.2 director's handbook
AI-generated animation is no longer difficult; what's difficult is making it not look like AI. Florence's 9 directing rules summarized from 40 works, plus 18 GitHub repositories that can copy, measure, and control animation, form a complete AI animation workflow.
AI-generated animation has become so accessible that it is now a matter of “one sentence.” Anthropic's Fable 5.2 can generate a complete animated film from a prompt. But based on data from 30 hours after Fable 5.2's release, most films people make with it will be forgotten a day later.
The problem is not prompt length. Florence shared a list on X, saying she “copied every animation I saw,” relying on 18 GitHub repositories. She also published a long article summarizing 9 “directing” lessons from 40 completed works. Put the two together, and they exactly illustrate the next bottleneck for AI animation: generating is easy; controlling is hard.
Fable 5.2: It can generate a film, but it needs a director
Fable 5.2 is the Claude family model focused on 3D, hardware, and motion design. Use /model fable to switch to it in Claude Code; give it a description, and it can write a complete set of HTML/JavaScript code that renders as animation in the browser. It can even drive Blender, synthesize audio, and generate subtitles. Florence's observation is very accurate: most films generated with Fable 5.2 will be forgotten. The model itself is not weak; what's missing is a director.
What exactly does “director” mean here? Florence answers this question with 9 lessons, all centered on “making AI understand your intent precisely.” For example:
- Use numbers instead of adjectives. Fable 5.2 can understand roughness 0.38, but not “premium.” Replace “smooth” with roughness 0.38, clearcoat 0.35; replace “cool” with color temperature 3200 K; replace “fluid” with cubic-bezier(0.16, 1, 0.3, 1).
- Spring damping ratio determines emotion. “bouncy” is an emotion; 0.5 is an instruction. A damping ratio of 1.0 suits UI text, 0.8 suits cards, and 0.5 suits characters. Write numbers, not adjectives.
- Still-frame checks can be deceptive. A model checks 6 still frames and all pass, but the film is frozen. Use freezedetect and frame-to-frame difference maps to quantify motion.
- Write the timeline first. Use a timing.json as the master clock; all jumps, transitions, and sound effects read from it. Then actions and music won't be out of sync.
- High beats Max. Although Max thinks longer, it costs 3x as much and may not produce better results. In Florence's test, High completed the task for $9.08, while Max needed $26.54 and scored only 3 points higher. More ironically, Max once produced nothing at all after thinking for 128,000 tokens.
These rules are very concrete and can even be pasted directly into prompts for use.
18 repositories: from copying to control
If the 9 lessons are the “software,” then the 18 repositories are the “arsenal.” They solve the same problem: how to make what AI generates not only “watchable” but also “controllable.” These repositories can roughly be divided into the following categories:
Reverse engineering and copying
- reanimate: reverse-engineer animations from websites, videos, GIFs, or Lottie files. It first records a reference, then lets your own code rebuild it, then records again and compares until there is a frame-level match. The idea is very engineering-minded: don't judge by eye; rely on data.
Motion design and style
- motion-design-skill: teaches AI motion design principles (timing, easing, choreography). Suitable for all animation systems.
- mg-styles-15: 15 motion design styles, each with a complete prompt and source code. All code was written by Claude Opus 5.5, and includes music and sound effects.
- huashu-art-motion: brings paintings to life and supports 35 art styles.
Generation and effects
- anidoodle: draw doodles and looping animations with code.
- cinetic: create films directly from code.
- motionmaxxing: turn “AI slop” into designer-level motion.
- Manim: smooth animation of mathematical formulas.
- blender-skills: give AI 28 skills for using Blender.
- three.quarks: particle effects.
- react-postprocessing: make every frame cinematic.
Control and compositing
- Tweakpane: real-time parameter sliders for easier debugging.
- EffectCraft: a fast compositing tool written in Rust.
Audio and subtitles
- Openverse: free sounds and music.
- librosa: detect beats and sync motion to them.
- WhisperX: millisecond-accurate subtitles.
Output and rendering
- Remotion Skills: make videos with React, official skills.
- FFmpeg: the final video encoding tool.
The common thread among these repositories is that they all “measure” and “verify.” reanimate compares frame-level differences, motion-design-skill has a quality checklist, and mg-styles-15 has AI judges scoring. This is exactly what Florence emphasizes: the person checking cannot be the person doing the work. A reviewer who is not involved in the build can truly spot problems.
Cost: don't kill the goose
Florence also raised an important issue: Fable 5.2 burns through quota very quickly. One user used up a 5-hour allowance in 20 minutes. Her advice is: iterate repeatedly on Medium, and only use High to build at the last moment. It's like raising a goose: one golden egg a day is enough; don't cut open its belly.
Conclusion
“Copy any animation” sounds like plagiarism, but it is actually understanding. These repositories and rules let you take apart an animation: look at its rhythm, elasticity, and transitions, then reassemble them into your own work. AI executes; you decide.
As Florence said in a reply: “A prompt can get you something watchable, but the angle and the reason to buy still come from a human.” Technology lets people do more, but it cannot replace taste.
发布时间: 2026-11-20 02:01