PromptWizard (FastAPI + HTMX)

PromptWizard is a local, browser-based prompt engineering tool that analyzes reference images and generates high-quality prompts for image/video diffusion models. It ports the original React UI into a FastAPI + HTMX + Tailwind workflow while keeping the same smooth, single-page experience.

What it is built for

• Quickly craft strong prompts from 1–6 reference images.
• Target specific model families/checkpoints (SD 1.5, SDXL, SD3, Flux, Wan, etc.).
• Tune creative direction using focused aspects, creativity level, and auxiliary pipeline add-ons.
• Use Gemini, OpenAI, or Grok as the captioning engine with per-provider API keys.

Install (uv only)

Requires Python 3.11+.

uv sync

Run

uv run uvicorn main:app --reload

Then open http://127.0.0.1:8000.

How to use the UI

The interface is split into three columns:

1. Visuals (left)

• Upload or drag and drop up to 6 images.
• These images are sent to the selected provider for analysis.

2. Wizard Lab (center)

• Target Architecture: choose the model family to tailor prompt structure.
• Specific Checkpoint: select the exact checkpoint within the family.
• Focus Areas: emphasize aspects like Subject, Outfit, Style, Lighting, etc.
• Creativity: adjust how loosely or strictly the model follows the reference.
• Pipeline Add-ons: optional upscaler, face fix, or ControlNet hints.
• Context / Notes: add custom instructions (e.g., setting, color shifts).
• NSFW (Grok only): UI toggle only for now (no backend logic yet).

3. Transmutation (right)

• Displays the generated prompt and optional negative prompt.
• Includes copy-to-clipboard buttons.

Top-right key bar:

• Choose Gemini/OpenAI/Grok and paste your API key.
• Keys are stored in LocalStorage for the session and persisted to .env.
• On refresh/restart, the UI loads the key from .env if present.

How UI changes affect the prompt

• Model family and checkpoint determine prompt structure and guidance.
  • Example: Flux expects natural language, SD 1.5 expects tags, video models expect motion descriptions.
• Focus Areas add explicit detail requests for selected aspects.
• Creativity adjusts model temperature: temperature = 0.2 + (creativity_level * 0.6).
• Auxiliary add-ons add specialized guidance:
  • ControlNet: emphasizes pose/structure clarity.
  • Upscaler: pushes high-frequency detail.
  • Face Fix: reinforces facial identity.
• Additional Context is appended as direct instruction for the prompt.

Notes

• Keys persist in .env and are reloaded into the UI on refresh.
• LocalStorage is reset per app instance to avoid stale key reuse.
• All prompt generation happens server-side through the selected provider API.