Interactive animated view of Soraya

Your browser does not support the canvas element. A modern browser with WebGL2 is required to view the animated character.
1.0×

Loading character…

Every component of the result we’re presenting here, including the illustration, its segmentation into animatable parts, the rigging, and the engine rendering it, was produced with generative AI tooling, assembled by someone with no background, professional or otherwise, in animation or illustration. This post describes how, and what we believe it implies.

Why we tried this

Ever since Fable 5 came back after the export ban, we wondered whether its vision capabilities could support fully automated rigging for Live2D-style animation.

Traditionally, for any Live2D animation, an artist draws and segments an illustration into layered parts, and a rigger then defines deformers, parameters, and physics inside the Live2D Cubism editor. It is a multidisciplinary process that often involves several professionals.

An initial experiment, published on July 4th, tested the narrower question of whether Fable 5 could perform the rigging at all. Given only a pre-segmented practice illustration, it wrote a custom engine and rigged a working model in about thirty minutes:

The result answered one question and surfaced another. The rigging itself was no longer the bottleneck. The remaining problem was upstream: producing a properly segmented illustration is itself challenging work. Modern image models generate beautiful illustrations, but they do not, by themselves, easily produce transparency or separate a character into the layered parts this kind of rigging requires.

The missing piece

While looking for a solution, we found that one had been announced only a few months earlier. See-through, by a Japanese group identifying as Shitagaki Lab, segments a flat character illustration into a layered PSD, inpainting the occluded regions behind each part.

Upon this discovery, we realized that an incredibly cheap, quick, and fully automated pipeline for Live2D-style animation might already exist, and that the remaining work was to piece the existing building blocks together.

The workflow

We started experimenting and settled on the following process, which produced the demo at the top of this post:

  1. We generated a base image using OpenAI’s ChatGPT Images 2.0, from a textual prompt and a previous visual reference of Soraya.
  2. We manually edited the illustration to correct a few points; namely, in the original generation, the bags the character holds were cut off rather than fully visible, which is inadequate for animation. We then ran a few diffusion passes with a WAI-Illustrious-SDXL family model to homogenize the result.
  3. We used Shitagaki Lab’s See-through to segment the image into a layered PSD.
  4. We performed fine edits on the generated PSD, including separating parts the model had merged (e.g. the two arms) and reordering layers.
  5. We let GPT-5.6 Sol Ultra write a fully custom engine for animating and rendering the model with OpenGL, then iterated with textual feedback until the animation behaved as we judged fit.

While this process still requires human intervention at several points, compared to the traditional pipeline, which demands expertise in drawing, segmentation, and rigging, it is almost fully automated, and completely doable by someone with no expertise in animation, drawing, or adjacent disciplines.

A prediction for the future

We expect Live2D and the ecosystem around it — artists, riggers, and the commercial tooling — to change drastically over the coming months. Live2D itself, meaning the commercial platform rather than the style or the technique, might effectively die as a product, with the company either pivoting hard or following it down. The engine itself is no longer a business moat: anyone can produce one with modern LLMs, and we have seen several people arrive at this independently, some already publishing their projects to GitHub.

Live2D’s remaining runway comes from integration rather than technology. It is deeply embedded in existing tools such as VTube Studio, game engines, and other streaming workflows. That ecosystem does not evaporate because engines became easy to write.

The work that remains for the community, in our view, is standardization: a common format, a reference implementation, and skills and connectors so that LLMs can follow the specification and perform the rigging themselves. The alternative is thousands of similar-but-slightly-different implementations, each rigged model locked to the engine that happened to produce it. Everyone benefits more from a single ecosystem. We are looking forward to helping build one.

On running See-through

Shitagaki Lab provides a Hugging Face Spaces app for running See-through. Unfortunately, it doesn’t fit our workflow: we don’t use Spaces, and the free tier is limited enough that generating segmented images at a reasonable resolution isn’t practical. We have instead forked the project to add an Astro-based web UI, alongside a Docker container suitable for cloud providers such as RunPod. The fork is available on GitHub for anyone to use, modify, and redistribute.

Closing thoughts

None of the individual pieces or ideas described here are ours. Image generation, illustration segmentation, and LLM-written engines each existed independently; what changed in the last few weeks is that assembling them into a working pipeline stopped requiring specialized expertise at any step.

The title of this post deliberately uses the word “towards”: the pipeline is not yet fully automated, the manual interventions we described are still necessary, and our predictions about the ecosystem may prove wrong in timing or in substance. But the direction seems difficult to argue with; a discipline that was gated by the intersection of several skills is now widely accessible with one day of patience and a modest compute budget.

The next major milestone is coordination. If the community converges on shared formats and reference implementations rather than fragmenting into incompatible ones, it would be revolutionary for the vast majority of people who aspire to use Live2D-style animation resources, including game developers and VTubers, but were previously limited by time, skill, workforce, and finances.