Hand an AI one PSD, and 1 and a half hours later it moves: StandRig and "the Live2D for AI"

StandRig is an open-source tool built by the head of a Japanese VTuber agency. It reads a PSD with properly layered parts, and lets Claude or GPT directly edit meshes, deformers and physics through MCP, producing a 2D character that moves and can be streamed live. It gained 206 stars in 58 days, almost entirely from the author's own demo videos on X. The most recent jump came from riding a free model that was trending that day. Below: how it took off, how crowded this path is now, and where to start if you want a project that picks up stars the same way.

A free DeepSeek girl, and one prompt

In 2026 Oct 30, kamiya_ryoko, a modeler who takes Live2D commissions, posted a free "DeepSeek girl" Live2D model on X and put it on VGen, original PSD included. The post got 2602 likes and 126k views (fxtwitter, 11-06).

A few hours later, sayaka_aiart quoted the post:

The psd file was released too, so I used StandRig to have an AI do the modeling. It took 1 and a half hours. The prompt I gave was: "Based on this psd with separated parts, use StandRig to build a highly movable 2D model with physics simulation."

It came with a video of the modeling process and a GitHub link. The post got 1037 likes, about 50k views, and 819 bookmarks. Over the next two days, StandRig's stars went from 92 to 130 and then to 189, with a single-day peak of +59 and 15 more forks (GitHub, 10-30 to 11-01). After that it gained only 1 to 5 a day, and by 11-06 it had stalled at 206.

What it actually is

One line: a 3D modeling kernel built specifically for AI to operate, 2D.

The author explains the origin clearly in a post from Sep 10: Live2D Cubism, the industry-standard software for 2D character modeling, has no complete editing API, so AI can only drive it through Computer Use, simulating mouse clicks. It's "very slow, and burns through tokens." So she decided to build a modeling tool optimized for AI from scratch. By her estimate, for modeling alone, the tool already covers about 80% of Live2D's features.

Specifically:

• A local service reads layered PSDs, and all operations go through MCP or HTTP;
• Safety workflow for agents: dry-run first, numeric checks, confirmation, checkpoint rollback;
• Deformation actions designed for AI (smooth, relax, inflate, pinch, bend);
• Playback directly in the browser. The same author also built Connect (camera face tracking, audio lip sync, OBS output) and Bridge, which connects to the Live2D Cubism API.

It's also worth being clear about what it doesn't do: it doesn't export Live2D's moc3 format, so the models can't go straight into VTube Studio, and it doesn't auto-separate layers, so you have to split the PSD yourself.

Where the stars came from

Almost all of the promotion was the author's own posts. She is a Japanese AI illustrator and head of the VTuber agency LiberAIte, with 39k followers on X. Her usual work includes live-stream illustrations, Live2D modeling, editing, and mixing. Each step StandRig took lines up with one of her posts (data from fxtwitter, 11-06):

The September wave brought roughly 80 stars (10-09 had 82 in total, per GitHub), and the 10-30 post brought about 100. Beyond that, the spread was small: one Chinese blog post, one Korean technical summary, a few MCP directory listings, and a portfolio site that collects "Opus 5.5 videos." I couldn't find any big account or media outlet that wrote about it.

One number deserves special attention: the September 10 post got 187k views and 1000 bookmarks but turned into only about 80 stars. The people who viewed her posts were VTubers, illustrators, and IRIAM streamers. They'll bookmark a video, but most don't go to GitHub. The traffic was decent, but very little of it converted to stars.

Why it took off comes down to three things:

1. The demo is the draw. "Hand a file to a frontier model, and in X minutes you get a professional result" is the kind of video people love to watch, especially now that models like Opus 5.5 and GPT-6 Astra can run long tasks.
2. The author is an insider with a built-in audience. The demo model looks good, and the prompt is posted so people can copy it.
3. It rode the right trend. DeepSeek girl was also trending in Chinese developer circles those days, and GitHub had several DeepSeek-girl stickers and a Live2D library for a whale girl appear around the same time. Chinese accounts replied under the 10-30 post, but how many stars came from the Chinese community can't be measured (my inference is that some did).

Who it's competing with

"Having AI rig a 2D character" is something that around 2026 9, October suddenly saw many projects crowding in. StandRig is not the biggest (star counts come from GitHub, 11-06):

Upstream, See-through (909 stars) does AI auto-layering, and these tools can be chained together.

One project launched a month after StandRig with a lower barrier to entry has 9 times its stars; another that exports the native format has 4 times as many. StandRig's "full API, built for AI" design made the demo videos work, but it lost on onboarding difficulty and ecosystem compatibility to contemporaries. Its stars came mainly from the author's distribution power, not from a product lead.

What to learn from it

• The topic-selection method can be reused: find a field where professional software has no API and AI can only brute-force through with Computer Use, build an AI-facing core, and expose it via MCP. The positioning is clear in one sentence, and it also explains why it had to be a new tool rather than a plugin.
• The distribution method can be reused: short videos of "frontier model + one file + X minutes = finished product," with a prompt people can copy. When a free asset is trending that day, make a same-topic demo with it.
• Pick the right audience: posting to illustrators and streamers gets lots of bookmarks but few stars. To earn stars, post where developers look, or ship a one-click Windows package and support mainstream formats so non-developers are willing to hit the star button.

If it were you

C. You can learn from it, but cloning it on the 2D character-rigging path isn't worth it. Since 9-October, at least 6 open-source projects have crowded into this path, and the top 1800 have many stars. StandRig itself has only 206 and has flattened out. It took off because of the author's 39k-follower industry account, which followers can't replicate.

Who it's for: people who already have an audience in a specialized creative community (3D, audio, typesetting, CAD and video color grading all count) and who can code, and who want to carry the "kernel for AI + MCP" approach into their own field.

Who it doesn't suit: anyone who wants to build yet another tool on the Live2D or VTuber puppet path, or anyone without their own distribution channel who hopes the project will be discovered on its own.

How to build it: Pick a niche where the pro software has no editing API. Start with the smallest possible core, so Claude or GPT can produce a shippable result from one input file through MCP. Copy the agent safety flow as-is: dry-run, validation, rollback. If you still want to do 2D characters, don't go head-on against the modelers. Fill the gaps nobody has done well yet: one-click export to a format VTube Studio can use, and a full pipeline from a single image to a stream-ready model.

How to get users: post screen-recorded videos one at a time, each showing a "file in, product out" process, with the number of minutes it took and the prompt included. Post separately in the Chinese and English developer communities, not only in user communities. If a free asset goes viral that day, use it for a same-topic demo.

When to move early: in the few weeks after a new model launches, videos of "have it do a professional grunt task" get the most traction. Have demos ready before the next generation of models comes out.

Signs you've misread the trend: if leading projects like mesh-avatar-studio and psd2live also flatten out within a month or two and no one ships anything new, the heat on this path was just a gust of wind from 9-October, and copying the playbook into another field should be discounted. Conversely, if Live2D officially releases a full editing API, the reason for "build a separate AI-facing core" disappears on its own.