Trying 3D Animation with GPT-6 Astra x Blender (VRoid to Storyboard to PV)
I wanted a short promotional video (PV) without standing up a whole studio crew. Character, background, storyboard, 3D CG, then an anime-look finish. GPT-6 Astra can drive most of that stack if you keep the brief tight and you're honest about which model does the final cel pass. TLDR: VRoid for the…
I wanted a short promotional video (PV) without standing up a whole studio crew. Character, background, storyboard, 3D CG, then an anime-look finish. GPT-6 Astra can drive most of that stack if you keep the brief tight and you're honest about which model does the final cel pass. TLDR: VRoid for the character, Blender CLI + MCP for the classroom and motion, a 15-second / 4-cut board as the bridge, then Omni 1.1 Flash (not Astra) for the 2D anime look. Treat this as a medium-inference, one-shot style experiment. Times below are what this kind of pass typically reports (VRM in about 8 minutes). Expect retries on faces and timing. If you still need procedural envs before you animate, I covered Geometry Nodes with Astra. When the next step is playable engines, see Astra for game creation. Why this is useful if you're solo Traditional PV work splits across character modelers, environment artists, storyboard artists, animators, and compositors. This stack collapses those roles into: VRoid Studio - faithful character from a still Blender CLI + Blender MCP - scene + motion under Astra Image / storyboard pass - 15-second, 4-cut PV plan Optional anime renderer - Omni for cel-look finish You still need good refs and a clear brief. What you don't need on day one is a five-person crew just to prove a trailer concept for an itch page or client pitch. Pipeline overview girl.png -> VRoid (Astra) -> girl.vrm / girl_standing.png classroom.png -> Blender CLI + MCP (Astra) -> classroom.blend / classroom_final.png refs -> 15s / 4-cut storyboard (16:9, stacked) storyboard.png + girl.vrm + classroom.blend -> 3D CG video (Astra x Blender) 3D video (previs) + character/bg refs -> 2D anime look (Omni 1.1 Flash) + end line Inference level for this writeup: medium. Treat results as "what one medium one-shot pass produced," not a Max-tuned production bible. 1. Character creation (VRoid Studio -> VRM) Prompt Using the attached image as reference, recreate the character as faithfully as possible in VRoid Studio. Input girl.png (character reference) Output Astra drives VRoid Studio and can produce a VRM in about 8 minutes on a clean pass. Tip: export both the .vrm (for Blender / engines) and a clean standing still (girl_standing.png) for storyboard and later anime conversion refs. 2. Background creation (Blender CLI + MCP) Prompt Using the attached image as reference, recreate the scene as faithfully as possible with Blender CLI and Blender MCP. Input classroom.png (environment reference) Output Astra rebuilds the classroom in Blender via CLI + MCP, not GUI-only clicking. Why CLI/MCP matters: same pattern as other Astra 3D demos. Headless-friendly control beats fighting the full GUI for agent loops. Save classroom.blend and a beauty still (classroom_final.png) for the storyboard stage. 3. Storyboard (15 seconds, 4 cuts) Prompt Create a 15-second, 4-cut PV storyboard starring the girl in girl_standing.png (VRM) set in the location from classroom_final.png (Blender). Storyboard frames: 16:9 aspect. Stack the panels vertically. Inputs girl_standing.png classroom_final.png Output A vertical strip of four 16:9 panels. That's the previs language for the next Blender pass. Keep cuts readable (who moves, camera intent, timing) so Astra isn't inventing blocking from vibes alone. 15 seconds / 4 cuts is a nice length: enough drama for a hook, short enough to regenerate when a cut fails. 4. 3D CG animation (storyboard -> Blender video) Prompt Using Blender CLI and Blender MCP, generate a video that follows storyboard.png as faithfully as possible. Use model girl.vrm and background classroom.blend. Inputs storyboard.png (or your board file; some pipelines name it ekonte / 絵コンテ.png) girl.vrm classroom.blend Output Storyboard-to-3D pass under Astra x Blender x VRoid assets. This is the core Astra segment: motion + camera + timing constrained by the board and existing assets, not a from-scratch mesh fantasy. 5. Convert to 2D anime look (Omni 1.1 Flash, not Astra) Important: the cel finish in this pipeline uses Omni 1.1 Flash, not GPT-6 Astra. Keep the model names honest in your own notes and client pitches. Prompt summary (Omni 1.1 Flash) Use the 3D animation as previs. Match motion, acting, camera, composition, and timing as closely as possible. Apply character + background reference designs faithfully. Target a hand-drawn Japanese anime look: cel-style lines and shading anime fill soft light depth of field glow, bloom chromatic aberration film grain Suppress 3DCG feel; stabilize face / costume / line / color across frames. End with a short line if you want one. The demo pass used 「きみを待ってた」 ("I've been waiting for you"). Refs 3D animation video (previs) girl_standing.png (character design lock) Why split the models Treat 3D as blocking previs and Omni as the look pass. That keeps Astra in tool use + spatial build, and puts stylization on a model tuned for video/anime conversion. Practical workflow checklist Stage Tool Model Artifact to keep 1 Character VRoid Studio GPT-6 Astra girl.vrm, standing still 2 Background Blender CLI + MCP GPT-6 Astra classroom.blend, beauty still 3 Storyboard Image / board gen (per your stack) 4x16:9 vertical strip 4 3D CG Blender CLI + MCP GPT-6 Astra short PV video 5 Anime look Omni video pass Omni 1.1 Flash final stylized clip + end line Try this next (same afternoon) Lock two refs only - one character still, one env still - before any motion prompt. Force a 4-cut board before Blender animation; reject boards that don't name camera moves. Label the Omni pass clearly in your README so nobody thinks Astra did the cel finish. Caveats Medium one-shot: expect retries on face stability, cloth, and cut timing. Faithful reproduction depends on reference quality; muddy refs become muddy VRM / classroom. Don't claim studio-ready TV anime from a single medium pass. Claim a solo PV pipeline you can iterate.