Dafeiyu character assets
DSH Dafeiyu Assets: Transparent Motion & Manifests
- Author
- DeepSeekAgent.io Editorial Team
- Published
- Updated
Pet code can be reused, but character quality depends on disciplined assets. Inconsistent anchors, broken loop endpoints, and green transparent edges make a character look like a video pasted onto the screen rather than a resident of the desktop.
The community dsh-pet project publishes a complete route from prompts and green-screen video to transparent animation, plugin configuration, and release. This guide turns that process into a market-ready character-pack workflow.
Define a minimum motion set
Begin with six semantic states:
| State | Suggested motion | Source |
|---|---|---|
IDLE | breathing, blinking, glancing | no active Session |
THINKING | pondering, gentle sway | model generation |
WORKING | sweeping, typing, swimming | Tool Call running |
WAITING | looking at the user, sign | input or approval required |
SUCCESS | cheering, spin, jump | successful turn |
ERROR | upset, dizzy, lying down | tool or turn failure |
Click, head pat, drag, and feeding are temporary interaction layers. They must not replace authoritative Agent state; after playing, return to the latest Host snapshot.
Keep the character consistent at generation time
Every generated motion should share a template: fixed aspect ratio, camera, character size, and foot baseline; a background distinct from the character; no motion outside the frame; similar start and end poses; consistent clothing, tail, headpiece, and colors; preserved prompts, seeds, source video, and generator version.
For a bottom-edge pet, foot anchoring matters more than visual centering. A changing baseline makes the character float during transitions.
Convert green screen to transparent animation
A practical pipeline is:
- key every frame;
- remove green spill and repair semi-transparent edges;
- normalize canvas size or record per-frame anchors;
- remove duplicate frames and wasteful bitrate;
- export alpha WebM, PNG sequences, or atlases;
- inspect edges over light, dark, and complex backgrounds.
WebM is simple to package; atlases provide precise frames and transition control; Live2D is more expressive but introduces separate model, texture, motion, and physics licenses.
Character manifest
A pack should identify itself, renderer, anchor, version, and state assets:
{
"id": "my-dafeiyu",
"name": "My Dafeiyu",
"version": 1,
"renderer": "webm",
"anchor": { "x": 0.5, "y": 1 },
"states": {
"idle": ["idle.webm"],
"thinking": ["thinking.webm"],
"working": ["working.webm"],
"waiting": ["waiting.webm"],
"success": ["success.webm"],
"error": ["error.webm"]
}
}
Use the real schema of the target runtime. Also describe loops, enter/exit segments, weights, cooldowns, audio, subtitles, and asset licenses.
Performance budget
Pause rendering when hidden, avoid high-rate idle repaints, load each state once, defer large audio and video, provide reduced-motion and static alternatives, and allow low-end devices to disable particles, shadows, and live blur.
A market page should disclose package size and typical CPU/GPU and memory behavior, not only a showcase video.
License every layer
MIT or Apache-2.0 code does not grant rights to character art, generated video, voice, font, or music. Dafeiyu derivatives may also carry attribution, noncommercial, and share-alike obligations from earlier designs.
List code license, character origin, derivative and generation notes, voice/music/font terms, and whether commercial use, redistribution, reskinning, and training are allowed.
Next connect the motions to the DSH Session and Tool Call state machine, or start with the build-your-own Dafeiyu guide.