vrm-charsheet turns VRM avatars into 1536×2048 AI-ready character sheets

The tool grew out of a practical problem. A developer was making a music video for an original VTuber character, Shichisaki Shiho, using fal.ai's video generation models such as MiniMax H3 Max and Kling. Those models preserve a character's look when given a reference image, but the quality of the reference image shows directly in the output. The first attempt used VRoid Studio's own export, which produced images around 235px wide with faces of roughly 40px — too small for the AI to avoid guessing at facial detail, and below the 300px short-edge minimum Kling's Elements feature requires.

Manually framing the VRoid Studio camera for six shots at matching angles was tedious, so instead the tool loads a VRM file and saves the standard compositions at a chosen size. Under the hood it uses three.js 0.160.0 with @pixiv/three-vrm 2.1.0, loaded directly from jsDelivr via importmap with no build step, and derives each composition's gaze height and distance relative to the model's bounding box so avatars of different heights get roughly the same framing. The developer reports that after adding high-resolution front and three-quarter face images as references, eyes, bangs and contours in generated close-ups came much closer to the VRM model.

The current recommended reference set is four to five images: the start-frame image, front and three-quarter face shots, and a full-body front shot, with the two face images optional for wide shots. Planned work includes an English UI, shared JSON presets, and export with VRMA poses applied, which together with the developer's VRMA sequential player would keep video material creation inside the browser.