One cosplay photo → Nano Banana Pro T pose → Tripo 3D → a 12-second MMD dance.
One original photo · a generated T pose · 104 rig bones · 12 seconds in motion
Your character already has a look. CosMMD carries that look into motion: the hair, outfit, cuffs, fingers, shoes and small details that make the character yours. Nano Banana Pro, accessed through Tripo, prepares a neutral T-pose image. Tripo turns that reviewed image into the 3D starting point. Blender supplies the rig, motion and rendering. CosMMD packages the glue, the repair recipes, and the checks we wish we had on day one.
Like the transformation? Star CosMMD and help build the next chapter of single-image character animation.
- Photo-to-T-pose preparation with Nano Banana Pro through Tripo API: remove pose/scene clutter while preserving the character, then review before 3D generation.
- A Tripo API workflow for the prepared T pose → textured GLB → humanoid auto-rig, with task checkpoints and no blind paid retries.
- Appearance-preserving rig tools: transfer weights to the original surface, fit finger chains, calibrate the neutral pose, and constrain knee bends.
- An MMD import bridge for a locally supplied, licensed VMD through external MMD Tools.
- A robust rendering path: a fresh Blender process per frame, resumable output, every-frame scans, and GIF generation.
- The actual reference-character recipes, plus a detailed failure notebook covering swapped legs, misplaced elbows, unnecessary reconstruction, and black render artifacts.
- A companion agent skill that guides the full process and keeps geometry-specific decisions explicit.
This is an early, agent-guided toolkit, not a universal one-click auto-rigger. The only character input is one original photo, ideally showing the whole outfit; the T pose is an intermediate output, not a second photo you must supply. You also need a Tripo API key, Blender, and a motion you are allowed to use. Calibration is still necessary for complex garments and generated topology. The input image does not generate the dance choreography.
Python 3.10+; the Blender workers were exercised with Blender 5.2.1 on Apple Silicon. CPU rendering is the portable default. Other operating systems/GPU backends are configurable but have not been verified in this showcase.
git clone https://github.com/CyclopsRay/CosMMD.git
cd CosMMD
python -m venv .venv
# Activate the environment for your shell, then:
python -m pip install .
# Dry run: inspect the Nano Banana Pro T-pose plan without uploading or spending.
cosmmd prepare --image private/photo.jpg --run private/my-character
# Set TRIPO_API_KEY privately, then generate the T-pose image:
cosmmd prepare --image private/photo.jpg --run private/my-character --execute
# Review t_pose.png against the original, then create the 3D model and auto-rig:
cosmmd generate --image private/my-character/t_pose.png \
--run private/my-character/model --executeRead photo → T pose → 3D for the API payload, preservation
prompt, review checklist and resume behavior. An existing clean T-pose image can go
straight to generate.
Then follow the workflow to inspect the generated mesh, calibrate the skeleton, import a licensed motion, and export your clip. For an agent-led session, install 2DCosplayerToMMD. There is deliberately no automatic “repair everything” step that silently replaces your character's hands or costume.
Already have a model in Tripo Studio? Export the source and rigged GLBs locally and start at inspection. A Studio asset identifier is not necessarily an API task ID. See API and reproducibility notes.
| Original photo / 原始照片 | T pose / 建模中间图 | 3D + MMD / 12 秒舞蹈 |
|---|---|---|
 |
 |
 |
The GIF contains 180 freshly rendered frames / 12 seconds / approximately 15 fps. It is an actual Blender render, not an AI-generated video. The first panel is the owner's original photo; the middle panel reuses the prepared T pose used for the existing Tripo Studio model. The newly packaged Nano Banana Pro preprocessing and 3D API stages have offline tests; this update did not buy a fresh end-to-end run. A face close-up was available during the original inspection, but is not required by the packaged single-image workflow. See provenance and validation.
Still visible: some hair deformation and skirt/thigh intersections on larger moves; no facial morphs, lip sync, cloth physics, or audio. The reference recipes produce editable Blender scenes, not a verified native PMX release.
| Failure | What changed |
|---|---|
| Left/right leg assignment and knee folds were wrong | Inspect anatomical sides, fit actual joints, use a tested hinge axis and preferred bend |
| Elbows and hands missed the source anatomy | Fit the joints to the mesh; calibrate T-pose → MMD neutral pose before importing rotations |
| Rebuilt hands and shoes lost identity | Restore the original surface, UVs and materials; repair only verified missing/fused areas |
| Fingers looked fused in one view | Inspect topology and multiple views; the source already had separate finger shapes |
| A valid MP4 contained a black character for 16 frames | Scan every frame, compare PNG masters, and rerender from a fresh scene/process |
| Inner-shoe black patches were blamed on seams | Same-model rerenders removed them; separate rendering failures from geometry defects |
The full failure notebook includes the failed assumptions, concrete checks, and remaining limitations. These are operational lessons, not a claim that all future models will behave the same way.
The next wins are better hair/garment separation, automatic landmark proposals, repeatable multi-character evaluation, native PMX export validation, and fewer manual calibration steps. Contributions with a reproducible before/after are welcome. See CONTRIBUTING.
The technical roadmap / 可复用技术路线 specifies the next cloth, lighting and scene stages, their technology choices, reusable contracts and acceptance gates. These are planned capabilities, not additions to the current demo. The part-pipeline investigation compares existing-mesh separation, segmented assets and regional generation, with an inspected face-alignment failure. Earlier publication ideas are archived and paused.
Code: MIT. Showcase media and third-party assets: excluded from that license.
No source motion, model, baked animation, audio, or .blend file is distributed.
Motion: つん;
choreography: 足太ぺんた;
song: TOKOTOKO / 西沢さんP.
Download at the author's page ·
Author terms · Full credits & media policy.
给它一张图，让角色走出画面，跳起 MMD。
CosMMD 把原始照片整理成 T pose、Tripo 单图建模、Blender 骨骼校准、MMD 动作导入和渲染整理成一套 可维护的工具与 agent 工作流。重点是保留角色原有的辨识度：发型、衣服、手型、 指甲、袖口、鞋面与袜子细节。
你只需要提供 一张原始角色照片，另需 Tripo API 密钥。流程先通过 Tripo API 中的 Nano Banana Pro 生成 T pose 图片，检查角色细节后，再把这张中间图交给 Tripo 生成 3D 模型。 T pose 是流程产物，不需要你额外提供。舞蹈动作需要从 原作者处另行取得并符合使用许可；它不是从图片中生成的。当前版本是需要 agent 或人工校准的早期工具，复杂模型仍不能保证一键完成。
已经提供： Nano Banana Pro T pose 预处理、API 任务断点恢复、原网格权重转移、手指骨骼工具、膝关节约束、 MMD 导入桥接、逐帧独立渲染、黑帧检查、GIF 制作，以及这次角色的完整修复配方。 示例达到 104 根骨骼、30 根手指骨骼与 12 秒舞蹈；头发拉伸、裙摆穿插、表情和 物理模拟仍是下一阶段的改进重点。
下一阶段的技术栈、布料代理、原图场景与灯光、缓存恢复和跨角色验收见 可复用技术路线。文档明确区分已有能力和待实现模块。 角色分件、局部生成与脸部错位的选型依据见 分件路线调查； 论文相关讨论已 记录并暂停。
先用 prepare 从原图生成 T pose；检查脸、手、鞋和左右服装细节后，再用 generate
从 T pose 生成 3D 模型。两步都默认只预览计划，加 --execute 才调用付费任务。
完整参数与提示词见 原图到 T pose 再到 3D。模型生成后请按 完整流程 校准，或使用配套
2DCosplayerToMMD skill。
展示已换成原始照片，并展示当前模型实际使用的 T pose。已有模型与舞蹈保持对应；
本次新增 API 预处理做了离线测试，没有重新付费生成整套案例。
仓库只发布代码、文档与审核过的展示图片/GIF。模型、动作包、带动作工程、贴图、
音频、API 密钥和运行记录都留在本地。除了 .gitignore，还会检查 Git 暂存区和历史。
原作者链接与致谢见 素材说明。
如果你也想让一张图里的角色动起来，欢迎 Star CosMMD，分享你有权公开的前后对比， 一起把这个工作流做得更稳、更好用。