This repo compares two prompt formats, structured JSON and natural-language text, across multiple FAL image models while holding the semantic content as closely matched as possible.

Public demo: https://ecsplendid.github.io/prompt-format-json-vs-text/

Current prompt set: v2-semantic-match

• A generator that submits both prompt formats to each configured FAL model.
• A local/static viewer that groups results by model and shows JSON vs text side by side.
• Downloaded local image outputs under web/generated/<model-slug>/<run-id>/.
• A static manifest at web/data/experiment.json that the GitHub Pages build can serve directly.

The current rerun uses a prompt pair labeled v2-semantic-match. The earlier pilot prompts were directionally similar but not tightly mirrored, so this rerun aligns camera, composition, lighting, grading, typography, and exclusions much more closely. The format is intended to be the main variable.

• fal-ai/nano-banana-2
• fal-ai/phota
• fal-ai/flux/schnell
• fal-ai/z-image/turbo
• fal-ai/bytedance/seedream/v4.5/text-to-image
• fal-ai/ideogram/v3
• fal-ai/imagen4/preview

• Nano Banana 2: complete
• FLUX Schnell: complete
• Z Image Turbo: complete
• Seedream v4.5 Text to Image: complete
• Ideogram V3: complete
• Imagen 4 Preview: complete
• Phota: partial on the fair rerun because the provider returned repeated downstream 500 errors for 4 natural-language variations

npm run generate
npm run dev

• npm run generate expects FAL_KEY in the environment.
• npm run dev serves the viewer locally with a dynamic /api/experiment endpoint.
• The static GitHub Pages site serves the committed snapshot from web/data/experiment.json.
• Rerunning with the same RUN_ID resumes from files already present on disk and only submits missing variation slots.

• MODELS=fal-ai/nano-banana-2,fal-ai/phota runs a subset of models.
• RUN_ID=2026-03-27T09-40-17-299Z MODELS=fal-ai/phota npm run generate resumes only missing outputs for a partial run.
• VARIATION_COUNT=15 changes the number of prompt variations per model.
• GROUP_CONCURRENCY=5 controls how many jobs each model/prompt group runs in parallel.
• MAX_VARIATION_ATTEMPTS=5 increases retry depth for flaky providers.

• No API keys are stored in this repo.
• .env*, node_modules/, local junk, and archived scratch outputs are ignored.
• The public snapshot is pinned to the fair rerun 2026-03-27T09-40-17-299Z; superseded scratch reruns are kept out of the public repo.
• The generator reads FAL_KEY from the environment at runtime.