Live2D 流程因 Astra 迎来加速可能

R.K. Snow @RKM_Snow原推发布于 2026.09.14

在 X 上打开原推

So obviously @Leaflit has been developing the process for some incredible work on live 2D models and her work is going to be infinitely better than what I've got here, but I wanted to drop some info on how I've managed develop my own live 2D model with zero experience. It's not excellent, but I like it so far and I think that those who want to get started and have GPT-6 Astra you can follow this simple guide to start producing your own VTuber avatar and live 2D model. I supplied the character direction and tested the avatar, while GPT Astra helped create the artwork, prepare the layers, and work through the rigging and export. It took several iterations rather than one prompt. Here’s a practical guide others can follow: Start with a character you like. Provide a portrait or reference sheet and describe the features you want preserved. We established our wizard’s brown hair, beard, glasses, navy-and-gold clothing, and subtle half-elf ears before refining the presentation. Explore the style before committing to a rig. Ask Astra to generate variations, using references to explain proportions, shading, and overall appearance. We explored different directions and settled on a rounded, cel-shaded look. Getting the design right here saves rebuilding animation later. Turn the illustration into layered artwork. A Live2D model needs separate movable pieces: eyes, eyelids, eyebrows, mouth shapes, glasses, head, and body. Ask Astra to help prepare a layered PSD for Cubism. Hidden areas also need artwork—for example, the neck must extend behind the beard so moving the head doesn’t reveal a gap. A finished portrait alone isn’t a finished Live2D asset. Build the rig gradually in Live2D Cubism. Import the PSD, create meshes and deformers, and connect them to parameters for blinking, mouth opening, gaze, and head movement. We started with simpler expressions and movement, then tested more ambitious motion. Astra helped with scripts, file preparation, and supported desktop control; you still need to review the visual results. Export to VTube Studio and test with your own face. Export the Cubism runtime model, load it in VTube Studio, and connect webcam tracking to its parameters. Adjust the input ranges and smoothing while blinking and speaking naturally. Tracking calibration and rigging are separate: good tracking cannot compensate for a head or mouth that deforms poorly. Use real testing to guide revisions. Test in OBS and review a recording. Tell Astra exactly what looks wrong—or provide a screenshot. Our testing exposed mouth-response issues, facial warping, and a visible throat separation. We kept versioned backups, returned to an earlier rig when needed, and improved the artwork’s overlap. Our current result is a working streaming avatar with eye and mouth controls and modest head tilt. Natural left/right turns and looking up/down still need further rigging. A useful starting prompt is: “Help me turn this character into a Live2D avatar for VTube Studio and OBS. Let’s approve the design first, prepare properly separated artwork with hidden overlap, then build and test the rig in stages. Keep versioned backups and explain what you can automate with the tools available.” The biggest contribution you make is art direction and testing. Astra can help carry out the production work, but clear feedback is what makes each revision better.