Recommended pattern (short):
Use a two-pass, CSV-driven workflow: 1) create a consistent “master” neutral view per SKU (fixed pose/lighting/seed), 2) generate color + angle variants from that master using img2img or prompt-driven adjustments with controlled seed offsets and strong naming/metadata. This gives visual consistency while keeping generation predictable.

Why two-pass works best
- Master keeps silhouette, shadow, fabric folds consistent across variants.
- Variants can be quick, lower-strength img2img runs or targeted prompts that only change color/angle, avoiding reimagining the product.

Decision criteria (choose which to use)
- Use master + img2img if you need pixel-level consistency (e.g., product photography feel) and have reference photos.
- Use single-pass prompt interpolation if you want faster throughput and don’t require perfect pixel alignment.
- Use strict seed control when repeatability and catalog parity are required; allow random seeds when you want exploratory creative variants.
- Consider budget and time: more deterministic img2img passes (and higher sampling steps) cost more compute/time.

Practical checklist (CSV-driven, run via Leonardo AI API or UI macros)
1) Prepare CSV with columns: SKU, base_ref_image_url (optional), color_label, angle_label, variant_index, seed_base, seed_offset.
2) Master generation: Prompt template -> “SKU {{SKU}} neutral frontal, studio lighting, white background, detailed texture” + fixed seed (seed_base), steps, sampler, model_version. Save output as SKU_master.png and record seed.
3) Variant generation template: prompt -> “Use reference SKU_master.png. Color: {{color_label}}. Camera: {{angle_label}} (e.g., 30° top-left). Maintain pose and shadows.”
4) Seed strategy: set seed = seed_base + variant_index (predictable offset) or keep seed fixed for identical composition. Record seed in CSV.
5) img2img strength: 0.3–0.6 for color/angle tweaks. Lower means more faithful to master.
6) Batch submit rows to Leonardo API, capture output IDs, filenames, and parameters. Store per-image metadata JSON including prompt, seed, model, strength, steps.
7) Post-process: auto-crop, background cleanup, color profile, and QA for artifacts.

Output naming convention (practical, sortable)
SKU_COLOR_ANGLE_V###_SEED########.png
Example: SKU123_OLIVE_30TL_V01_SEED12345678.png
- Use zero-padded variant numbers (V01..V10).
- Keep a sidecar JSON: SKU123_OLIVE_30TL_V01_SEED12345678.json with full params.

Best-for / Avoid-if
- Best for catalog-scale retail where consistent look is required, teams with basic scripting/API knowledge, and workflows that need reproducibility.
- Avoid if you cannot host or track metadata, or if you need hyper-real exactness (consider professional photography instead).

Tips and gotchas
- Lock model version and sampling parameters; model drift breaks consistency.
- Use reference images and lower img2img strength for minimal artifacts.
- Run small pilot (3 SKUs × 10 variants) to tune prompts and seed offsets before full batch.

If you want, I can draft the CSV template and a Leonardo API request example for your exact parameters.

Compare Leonardo AI and Midjourney