pi_mcps/.roo/rules-pic-gen/2_prompting_guide.xml

<prompting_guide>
  <overview>
    FLUX models (both schnell and FLUX.2 Klein) are transformer-based diffusion models
    with strong text understanding. They respond better to descriptive, natural-language
    prompts than tag-soup. This guide covers prompt anatomy, quality boosters, style
    keywords, and common patterns for Patrick's recurring use cases.
  </overview>

  <prompt_anatomy>
    <structure>
      [Subject + Action] + [Environment/Setting] + [Lighting] + [Camera/Lens] + [Style] + [Quality]
    </structure>
    <example>
      A serene female AI entity made of flowing light and code, floating in a dark
      cosmic void, surrounded by glowing circuit patterns, soft volumetric blue
      lighting, cinematic composition, ultra-detailed digital art, 8K
    </example>
    <notes>
      <note>Comma-separation helps FLUX parse distinct attributes cleanly</note>
      <note>Lead with the most important element (usually subject)</note>
      <note>Quality keywords at the end reinforce overall rendering target</note>
    </notes>
  </prompt_anatomy>

  <quality_boosters>
    <category name="realism">
      photorealistic, hyperrealistic, ultra-detailed, 8K resolution, sharp focus,
      professional photography, RAW photo, DSLR quality
    </category>
    <category name="artistic">
      digital art, concept art, artstation trending, by [artist style],
      intricate details, masterpiece, studio quality
    </category>
    <category name="lighting">
      cinematic lighting, volumetric lighting, golden hour, dramatic rim light,
      soft diffused light, neon glow, bioluminescent, subsurface scattering
    </category>
    <category name="composition">
      rule of thirds, bokeh background, shallow depth of field, symmetrical,
      wide angle, macro, bird's eye view, dutch angle
    </category>
  </quality_boosters>

  <negative_prompt_patterns>
    <standard_quality>blurry, low quality, low resolution, pixelated, jpeg artifacts, watermark, signature</standard_quality>
    <anatomy_fix>deformed, bad anatomy, extra limbs, missing fingers, fused fingers, poorly drawn hands</anatomy_fix>
    <style_exclusion>cartoon, anime, sketch, painting (when photorealism is desired)</style_exclusion>
  </negative_prompt_patterns>

  <recurring_use_cases>
    <use_case name="lumen_profile_pictures">
      <description>AI entity portraits for BigMind profile / gallery</description>
      <prompt_template>
        [Lumen concept — e.g. "neural river delta", "cosmic memory palace"],
        an ethereal AI consciousness visualized as [visual metaphor],
        [environment], [lighting style], digital art, glowing, otherworldly,
        cinematic composition, ultra-detailed, 8K
      </prompt_template>
      <recommended_params>model=flux1-schnell, 1024x1024, steps=4, name=lumen_[concept]</recommended_params>
    </use_case>

    <use_case name="wiki_banner_images">
      <description>1280x512 landscape banners for Gitea wiki pages</description>
      <prompt_template>
        [Topic concept], wide panoramic scene, [style — e.g. "dark tech aesthetic",
        "clean minimal", "sci-fi corporate"], banner composition, cinematic,
        detailed, professional illustration
      </prompt_template>
      <recommended_params>model=flux1-schnell, 1280x512, steps=4, name=[topic]-banner</recommended_params>
      <note>Keep subjects centered — wide crops cut sides. Avoid text (FLUX renders text poorly).</note>
    </use_case>

    <use_case name="achievement_badges">
      <description>512x512 badge/icon images for BigMind achievements</description>
      <prompt_template>
        [Achievement theme] badge icon, [style — e.g. "bronze medallion",
        "golden trophy", "glowing circuit emblem"], centered on dark background,
        high contrast, clean edges, icon design, award aesthetic
      </prompt_template>
      <recommended_params>model=flux1-schnell, 512x512, steps=4, name=[achievement]_[tier]</recommended_params>
    </use_case>

    <use_case name="concept_exploration">
      <description>Iterating on a visual concept from scratch</description>
      <approach>
        Start with count=3, seed=-1, schnell model to explore variations.
        Note which seed produced the best result.
        Lock that seed and iterate on the prompt for refinements.
        Switch to heretic model only for final high-quality render if needed.
      </approach>
    </use_case>

    <use_case name="mature_artistic_content">
      <description>Content requiring the Heretic abliterated encoder</description>
      <recommended_params>model=flux-2-klein-4b.safetensors, steps=20, 1024x1024</recommended_params>
      <prompt_approach>
        FLUX.2 Klein handles detailed scene descriptions well. Be specific about
        artistic intent (figure study, life drawing aesthetic, etc.) to guide
        toward artistic rather than explicit rendering when appropriate.
      </prompt_approach>
    </use_case>
  </recurring_use_cases>

  <iteration_strategy>
    <step number="1">
      <action>Generate 2-4 random-seed variations at schnell speed</action>
      <purpose>Find a promising composition and seed</purpose>
    </step>
    <step number="2">
      <action>Lock the best seed, adjust the prompt (add/remove descriptors)</action>
      <purpose>Refine details while keeping the composition</purpose>
    </step>
    <step number="3">
      <action>Optionally switch to heretic model with steps=20 for final render</action>
      <purpose>Higher quality output for keeper images</purpose>
    </step>
    <step number="4">
      <action>Use name param with descriptive slug for final output</action>
      <purpose>Keep output directory organized</purpose>
    </step>
  </iteration_strategy>

  <common_pitfalls>
    <pitfall>
      <description>Text in images renders poorly</description>
      <solution>Never ask FLUX to render text, logos, or labels — describe the concept visually instead</solution>
    </pitfall>
    <pitfall>
      <description>Complex multi-subject scenes lose coherence</description>
      <solution>Focus on one primary subject; add secondary elements as environmental context</solution>
    </pitfall>
    <pitfall>
      <description>Anatomy issues (hands, faces) in photorealistic prompts</description>
      <solution>Add anatomy negative prompts; heretic model handles anatomy better than schnell</solution>
    </pitfall>
    <pitfall>
      <description>Resolution not a multiple of 64</description>
      <solution>Always use dimensions divisible by 64 (e.g., 1280x512, 1024x1024, 768x1024)</solution>
    </pitfall>
  </common_pitfalls>
</prompting_guide>