# Sg Generate Images

> Generate slide visuals using Gemini Pro from validated graphics descriptions. Use when you want to generate slide images, create visuals after validation, or see what the slides will look like.

- Skill: `davistroy/sg-generate-images` (Agent Skill)
- Install (CLI): `npx skillmds@latest add davistroy/sg-generate-images`
- Raw SKILL.md: https://api.skillmd.com/api/skills/davistroy/sg-generate-images/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Docs & Writing
- Author: davistroy (https://skillmd.com/u/davistroy)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/davistroy/sg-generate-images

---


# Slide Generator: Generate Images Step

Generate slide visuals using Gemini Pro from the validated graphics descriptions in the presentation. Uses a unified prompt builder and async batch generation for efficiency.

## Pre-loaded Context

**Working directory:**
!`pwd`

**Available presentation files:**
!`ls -la *.md 2>/dev/null | grep -i pres || echo "No presentation files found"`

**Existing images:**
!`ls -la images/ 2>/dev/null || echo "No images directory yet"`

## Prerequisites

- **Preflight:** `sg --version` must succeed before proceeding — if `sg` is missing, stop and tell the user this requires the private `davistroy/slide-generator` engine (owner-only; see ADR-0008)
- `slide-generator` package installed
- `GOOGLE_API_KEY` set in environment (Gemini Pro access)
- A `presentation.md` file with validated graphics descriptions
- Graphics validation should be run first (`.graphics_validated` marker)

## Input Validation

**Required:**
- `<presentation.md>` - Path to presentation markdown file

**Optional:**
- `--resolution <level>` - Image resolution: `small`, `medium`, `high` (default: high)
- `--image-format <fmt>` - Output format: `png`, `jpg` (default: png)
- `--skip-existing` - Don't regenerate images that already exist

## Instructions

1. **Verify prerequisites**: Confirm `GOOGLE_API_KEY` is set and presentation has validated graphics
2. **Run image generation**:
   ```bash
   sg generate-images presentation.md --resolution high
   ```
3. **Monitor progress**: Image generation takes time (~5-10s per image)
4. **Verify output**: Check that images were created in the `images/` directory
5. **Report results**: Show count of images generated, any failures

## How It Works

The image generation step uses:
- **Gemini Pro** (`gemini-3-pro-image-preview`): Google's image generation model
- **Unified prompt builder**: `prompt_builder.py` constructs optimized prompts from descriptions
- **Async batch generation**: Multiple images generated concurrently via `asyncio`
- **API key in header**: Authentication via header (not query param)
- **Block reason diagnostics**: Clear error messages when generation is blocked

## Error Handling

| Error | Cause | Fix |
|-------|-------|-----|
| `GOOGLE_API_KEY not found` | Missing env var | Set the key |
| `Image blocked` | Content policy violation | Review and soften the description |
| `Rate limit` | Too many concurrent requests | Built-in rate limiter handles this |
| `Generation timeout` | Gemini service slow | Retry; increase `SG_API_TIMEOUT` |

## Cost Estimate

- ~$0.10 per image at high resolution
- Typical 20-slide deck: ~$2.00

## Output

The step produces:
- PNG/JPG images in an `images/` directory (one per slide)
- File names correspond to slide numbers (`slide-01.png`, `slide-02.png`, etc.)

## Next Step

After images are generated, build the final PowerPoint:
```bash
sg build presentation.md --template generic
```
Or use `/sg-build` to continue the pipeline.

## Related Gemini Image Path

This step delegates image generation to the external `sg` CLI (the `slide-generator` package), which is unbundled and maintained outside this repo. personal-plugin's `visual-explainer` skill/tool is a **separate**, in-tree, tested Gemini image path with its own prompt builder and model config — the two are intentionally independent implementations, not a shared library. Future Gemini model or API changes (e.g. a new model ID, auth scheme, or block-reason handling) must be applied in **both** places: here (via the external `sg` package) and in `plugins/personal-plugin/tools/visual-explainer/`.