Recommendation
Use Leonardo AI’s batch-generation (API or scripted UI automation) with consistent seeds and a small set of prompt templates (white-background and lifestyle). Store every output plus a tiny metadata JSON per image, then push originals + thumbnails to S3 in a strict path scheme so downstream services can consume them reliably.

Decision criteria (pick approach)
- Budget: full AI generation (cheapest) vs hybrid composite (photography + AI composite) — choose photography for premium SKUs or where reflections/materials must be exact. AI-only is best-for mid-range catalogs and rapid scale.
- Skill & team size: one dev can run a scripted pipeline; bigger teams should add a QA step and asset manager. If no dev, use Leonardo’s web batch features and manual S3 upload.
- Output quality: if you must match an existing brand shoot, prefer hybrid; if flexible, pure AI with deterministic seeds is fine.

Prompt templates (use variables: {SKU}, {product_description}, {style_refs})
1) White background (clean e‑commerce)
Prompt: "Product: {product_description}. Photorealistic studio shot on seamless pure white background (#FFFFFF). Centered, 3/4 angle, soft shadows, even lighting, no props, high detail texture, realistic material rendering. Shot like a catalog product image. --seed {seed}"
Negative: "no watermark, no text, no logos, no extra objects, no background texture"

2) Lifestyle background (contextual)
Prompt: "Product: {product_description}. Lifestyle scene: modern minimal interior, soft natural light, neutral tones. Product in use or on surface, visible depth of field (bokeh), warm color grade, realistic shadows and reflections. Include subtle context props from style_refs: {style_refs}. --seed {seed}"
Negative: "no watermark, no brand logos, avoid overly stylized filters, no text"

Practical parameters to lock (consistent across batch)
- Resolution: 2048–3072px long edge (adjust per SKU). Aspect ratio: 1:1 for thumbnails, 4:5 or 3:2 for hero.
- Determinism: fix seeds per SKU to allow re-generation. Fix “guidance/strength” values to maintain consistency.
- Variants: generate N variants (N=3) and pick best via QA script.

Naming convention & metadata
File name: {sku}__{bgtype}__v{version}__s{seed}.png
Example: SKU12345__white__v1__s98765.png
Metadata JSON per image (same folder): SKU12345__white__v1__s98765.json with keys: sku, bg_type, seed, prompt, negative_prompt, model, resolution, created_at, checksum, selected(boolean).
S3 path convention
s3://your-bucket/product_images/{sku}/{bg_type}/{filename}
s3://your-bucket/product_images/{sku}/{bg_type}/thumbs/{filename}_thumb.jpg

Export steps & S3 integration (practical checklist)
1. Prepare CSV with SKU, product_description, style_refs, preferred_aspect, seed(optional).
2. Run a small dry-run (10 SKUs): generate, review, adjust prompts.
3. For production: script reads CSV → composes prompt → calls Leonardo API or web batch → receives image + seed → save to local with metadata JSON → create 200px and 800px thumbnails → validate (dimensions, bg white pixel ratio) → upload images + metadata + thumbs to S3 using boto3 or awscli.
4. Run automated QA: check for logos/watermarks, white-bg purity (>=98% white pixels), and seed consistency.
5. Approve and mark final version in metadata.

Automation pseudocode (high level)
- for row in CSV: build prompt, call Leonardo API → save image & meta → run QA → upload to S3 → log result

Best-for / Avoid-if
- Best-for: rapid scaling, consistent catalog look, budget-conscious teams.
- Avoid-if: you require spot-on physical accuracy (high gloss, subsurface scattering), or legal guarantees on likeness/materials—use pro photography or hybrid compositing.

Notes
Use ChatGPT to iterate and optimize prompt variations before large runs. Start small, lock seeds/params, and keep a human QA pass for the first 50–100 SKUs.

Quality score: 92

Compare Leonardo AI and Midjourney