chaithanya-m/image-generation

Universal image generation

This is a universal image-generation capability. Never assume the supplied content is finance-related. Determine the subject and visual style from the actual input.

When to invoke

Use the image agent whenever the user asks for visual assets, illustrations, carousel art, or images based on supplied content. It is not coupled to a research or publishing workflow.

Supported input

• Plain text and Markdown.
• JSON objects and arrays, including serialized DialogueScript and InstagramPost content.
• Individual slides/scenes and multi-section documents.
• Existing visual_prompt fields, which should be reused as context.

Plan and generate

Run commands from the image-generation-agent project directory. Always preview a plan first when requested or when generation cost/volume warrants it:

python -m src.image_agent.cli --input content.json --dry-run

The dry run analyzes the actual topic with the configured LLM and prints the domain, sections, style, aspect ratio, and full prompts; it does not contact the image provider or save images. To generate:

python -m src.image_agent.cli --input content.json
python -m src.image_agent.cli --input content.md --aspect-ratio 4:5
python -m src.image_agent.cli --input content.json --style cinematic \
  --images-per-section 2 --output output/custom

# Retry failed images with their saved prompts, without re-analysis
python -m src.image_agent.cli \
  --retry output/images/generated/topic/generation_manifest.json

The plan determines a content-appropriate number of images, bounded by the configured maximum. For slides/scenes, keep ordering and a consistent style across the series. Use existing visual prompts as source material. Do not add facts beyond the input. Exact text should be overlaid with Pillow, never delegated to the image model.

Report and failures

Images and generation_manifest.json are saved under the requested output directory, in a unique content-slug folder. Report the exact output path, number of successful/failed images, and manifest path. Check the manifest before reporting completion.

If analysis fails, report the configured LLM API error. If image generation fails, report the provider error and failed image number; do not imply an image was produced or silently switch providers. Cloudflare generation requires CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN. Previously generated files are never overwritten. Retry failures with --retry <generation_manifest.json> so successful images are skipped and the original prompts are reused.