界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
开发与工程 · SKILL.md

webui

webui

【开发与工程】Launch and troubleshoot DeTikZify's Gradio web UI, choose models and algorithms, and reason about runtime UI behavior.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: webui description: "Launch and troubleshoot DeTikZify's Gradio web UI, choose models and algorithms, and reason about runtime UI behavior." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Web UI

Use this sub-skill when the task is about python -m detikzify.webui, Gradio startup, model selection, share links, light mode, or the browser-facing synthesis experience.

Route away from this sub-skill when the task is mostly about programmatic inference, training, evaluation, or dataset internals.

Fast Path

  1. Check the CLI surface first:
    python scripts/webui_help.sh
    
  2. Confirm the package import and core API surface if the UI startup depends on the same environment:
    python scripts/api_smoke.py
    
  3. If the task mentions rendering or MCTS gallery behavior, also confirm the compile path works:
    python scripts/tikz_smoke.py
    

What This Sub-Skill Owns

  • python -m detikzify.webui
  • CLI options such as --model, --algorithm, --lock, --lock_reason, --share, --light, and --timeout
  • Gradio theme choices and the light-mode patching behavior
  • runtime behavior of the two UI algorithms: MCTS and sampling
  • model-selection choices and the browser-facing compile / gallery flow

Common Decisions

  • Use --light when you want a white UI background that matches scientific figures.
  • Use --share only when you want a public shareable Gradio link.
  • Use --lock when the model must not be changed interactively.
  • Choose mcts when you want multiple ranked TikZ candidates and compiled previews.
  • Choose sampling when you only want a single synthesized output image.
  • Treat the timeout as minutes, not seconds, in the web UI advanced settings.

Bundled References

Related Helpers

Guardrails

  • A successful --help run does not prove that the model can launch.
  • The UI still depends on the same CUDA and TeX-backed capabilities as the programmatic pipeline.
  • --light is a theme control, not a generation control.
我的收藏