Style Lora Research Pipeline

Style LoRA Research Pipeline

Use when: the user wants to mine external style LoRA libraries, curate a brand-aligned shortlist, or batch-test styles against real scene prompts.

Trigger conditions

• "curate style LoRAs", "test styles on FAL", "batch test LoRAs", "style library", "Loratlas", "Krea 2 LoRAs"
• User wants to find new visual styles aligned with a brand aesthetic before deciding whether to hybrid-fit locally or inference via API

Workflow

1. Discover: load the external library homepage and navigate to the indexed view (Library/All).
2. Filter by brand DNA: match palette, lighting, cinematic density, and genre overlap. Use subjective fit ranking, not like-counts alone.
3. Shortlist tiers rather than one flat list: core dramatic skin, secondary ornate skin, reserve anime/lighter skin, and mural/old-artwork fallback family.
4. Deep-dive each page: extract the trigger phrase, the Hugging Face weight path, and the developer-call prompt template. Pause before generating — read the page’s prompt guide first.
5. Author test prompts: brand-themed subjects, stripped to concrete nouns + minimal action/setting. Never redefine medium; let the LoRA carry it.
6. Batch-submit via fal.run endpoint fal-ai/krea-2/turbo/lora (or equivalent). Use async polling. See references/fal-lora-testing.md for the stdlib-only fallback and the exact poll loop, and references/cheonma-fal-batch-notes.md for the brand-local batch CLI and synchronous-response behavior.
7. Collect outputs: save image URLs + request IDs locally under ~/cheonma/loratlas_test_runs.json or brand-equivalent path.
8. Evaluate: rank tested styles by visual fidelity to brand prompts, not just novelty.
9. Formalize: save chosen skins as template files, subject prompts as JSON banks, and generate via a repeatable brand CLI instead of one-off scripts.

Brand-fit filter rubric

• Palette overlap with brand DNA
• Cinematic lighting density (chiaroscuro, backlight, golden hour, noir)
• Depth of field and texture grain behavior
• Consistency across varied subjects
• Scalability — can the style survive 1.0–1.3 scale without collapse

Pitfalls

• Don’t generate before reading the page-level prompt guide — each LoRA has a distinct ideal prompt shape.
• FAL’s image_generate tool is FLUX-only and cannot run custom LoRA weights. Use direct https://fal.run calls instead.
• requests is not guaranteed in the session venv. Use urllib.request + polling loop; see references/fal-lora-testing.md for the complete drop-in pattern.
• Some HF links reported broken; verify the v3b.fal.media/...safetensors URL is present on the style page before testing.
• This endpoint is sometimes synchronous: a completed response can appear as response.images[0].url while the top-level request_id is absent. Treat a missing request ID as “read the image from the response directly,” not as a hard failure requiring another submit.
• Async endpoints need polling — do not treat a 202/queued response as success.
• Markdown patch hygiene: when updating prompt template .md files, always inspect the fenced block afterward. Small replace patches can bisect code fences or duplicate headings.
• Subject bank sizing: 4–8 entries per archetype is too few for a recurring weekly feed. Formalize step should expand each bank to 12–15 concrete noun prompts before treating the template system production-ready.

Environment discovery for FAL auth

Preferred key lookup order:

1. ~/.hermes/.env
2. <brand root>/.env
3. profile-local ~/.hermes/profiles/<name>/.env only if explicitly configured
4. Do not assume ~/.env, /root/.env, or the active shell’s exports

The fallback script in references/fal-lora-testing.md hardcodes ~/.hermes/.env. Override that path only when brand setup explicitly differs.

Output expectations

• JSON array: one entry per style × subject.
• Fields: style, subject, prompt, request_id, final_status, image_url.
• Print summary counts by style at end.

Session package target

After testing, create one index doc next to references/fal-lora-testing.md, e.g. references/<brand>-style-shortlist.md, containing:

• Tiered shortlist
• Direct style-page URLs
• Trigger phrases
• Final style-URL pairs for future hybrid fitting