prompt upgrades

app.py

@@ -100,7 +100,7 @@ DEFAULT_PROMPT = (
     "three views front side portrait"
 )
 DEFAULT_NEG = (
-    "blurry, dark skin, umber brown, muddy colors, soft edges, "
+    "blurry, desaturated, washed out, muddy colors, soft edges, "
     "no outlines, photorealistic, 3d render, CG, chibi, ugly, deformed, "
     "extra limbs, watermark, signature"
 )

spec/notebooklm-research-findings.md

@@ -0,0 +1,107 @@
+# NotebookLM Research: Improving African American Character Design in AI Art
+
+**Date:** 2026-03-04
+**Notebook:** Research: African American Character Design in AI Art
+**Notebook ID:** c39eab4d-4af3-4203-9d85-56633b54b3f5
+**Sources:** 44 deep web research sources + 3 manual sources (ControlNet model page, Ghibli LoRA page, CivitAI guide)
+
+---
+
+## Problem 1: Multiple Characters Being Generated
+
+### Root Cause
+- The prompt contains `"three views front side portrait"` — FLUX interprets this literally and generates 3 separate characters instead of 3 views of one character
+- FLUX is trained on natural language captions, not tag clouds. It reads `"three views"` as "draw three people"
+
+### Fixes
+- **Use singular language:** Replace vague terms with explicit singularity — `"a solo character"`, `"a single girl standing alone"`
+- **Word order matters:** Place character count and main subject at the **very beginning** of the prompt. FLUX weights the first tokens most heavily ("Prompt Pyramid" priority)
+- **Remove multi-view language** for single character output. Only use `"character design sheet, multiple views"` when intentionally generating a sheet
+- **Avoid descriptive overload:** Too much environmental detail can cause the model to add more subjects to "balance" the scene
+- **FLUX does not natively support negative prompts.** Including them can actually cause the model to generate the excluded element
+
+### Advanced: Regional Prompting (Future Enhancement)
+- **Region-Aware Attention Manipulation** (from InstantX research) allows defining exactly where a character should appear using masks
+- Use a Global Description (whole scene) + Regional Prompt (specific character in masked area) + Background Prompt (e.g., "empty white background") for non-masked regions
+- Tools: `Regional-Prompting-FLUX` (diffusers), `ComfyUI-FluxRegionAttention`
+- **Two-Pass Method:** Generate background first, then inpaint character into a masked region — most reliable "brute force" approach
+
+---
+
+## Problem 2: African American Hair Rendering
+
+### General FLUX Prompting Strategy
+- **Use natural language prose**, not comma-separated tags. Write as if describing the scene to a human artist
+- Instead of `"locs, anime style"` use `"A character with thick, well-defined locs rendered in a bold anime style"`
+- **Avoid negative prompts** for hair quality. Instead, describe the positive attribute: `"clean, sharp edges"`, `"smooth texture"`
+
+### Hairstyle-Specific Descriptive Language
+Use descriptive phrases that emphasize **volume, structure, and texture**:
+
+| Hairstyle | Descriptors |
+|-----------|-------------|
+| **Space Buns** | "two large, perfectly spherical space buns", "symmetrical rounded buns", "neatly wrapped buns on either side of the head" |
+| **Afro Puffs** | "two voluminous, cloud-like afro puffs", "bold rounded puffs with high texture", "large, fluffy puffs secured with invisible bands" |
+| **Braids** | "intricately woven box braids", "sharp, clean cornrows with visible parts", "long, flowing braids with a slight sheen" |
+| **Locs** | "thick, heavy, well-defined locs", "neatly maintained dreadlocks with visible twists", "structured locs gathered into a high ponytail" |
+
+### Achieving The Boondocks Rendering Style
+The Boondocks (LeSean Thomas, Aaron McGruder) is known for bold, high-contrast rendering of Black hair with defined shapes:
+
+- **Define the shape:** Use `"bold silhouettes"`, `"clean outlines"`, `"exaggerated hair volume"`
+- **Style references:** `"2D anime illustration"`, `"cel-shaded rendering"`, `"high-contrast animation style"`
+- **Geometric descriptors for sharp edges:** `"The hair has sharp, clean edges and a strong, defined outline against the background"`
+
+### Recommended Prompt Structure
+Use **Subject + Action + Style + Context** framework:
+
+**Template:**
+> "[Subject with Specific Hairstyle], [Action/Pose], in a [Specific Anime Style], [Lighting and Context Details]."
+
+**Example — Bold Afro Puffs:**
+> "A young African American woman with two large, perfectly rounded afro puffs, standing heroically, rendered in a bold 2D anime style with thick clean outlines and cel-shaded colors similar to urban animation, set against a vibrant city sunset with high-contrast shadows."
+
+**Example — Defined Locs:**
+> "A character with long, thick, well-defined locs tied back, looking intensely at the viewer, classic 2D animation style with sharp line art, soft ambient lighting in a modern server room, high detail on the hair texture."
+
+---
+
+## Problem 3: Optimal ControlNet Settings
+
+### Single Portrait vs. Multi-View Character Sheet
+
+| Setting | Single Character Portrait | Multi-View Character Sheet |
+|---------|--------------------------|---------------------------|
+| **Conditioning Scale** | 0.5 – 0.6 | 0.7 – 0.8 |
+| **Guidance Start** | 0.0 | 0.2 |
+| **Guidance End** | 0.8 | 0.8 – 1.0 |
+| **Control Strength** | 0.3 – 0.5 (avoid "burning" flat skin/hair) | 0.6 – 0.8 (maintain spatial arrangement) |
+
+### Canny Thresholds
+- **Standard recommendation:** Low **50**, High **200** — captures essential structural outlines without noise
+- For pencil sketches with hatching/shading: raise Low to **100-120** to filter noise
+
+### Resolution and Aspect Ratio
+
+| Use Case | Recommended |
+|----------|-------------|
+| **Single Portrait** | 9:16 (768x1344) or 3:4 (768x1024) — more vertical space for hairstyle |
+| **Character Sheet** | 1:1 (768x768) or 16:9 (1024x768) |
+| **Optimal FLUX resolution** | 1024x1024 or higher (up to 2.0 megapixels) |
+| **Canny ControlNet stability** | 768x768 (XLabs-AI recommendation) |
+
+---
+
+## Recommended App Changes
+
+### Immediate (Prompt & Settings)
+1. **Rewrite default prompt** in natural language prose with singular character language
+2. **Remove negative prompt default** — FLUX doesn't use it effectively; can cause unwanted artifacts
+3. **Switch to portrait aspect ratio** (768x1024) for single character mode
+4. **Lower ControlNet conditioning scale** to 0.5 for portraits
+
+### Future Enhancements
+5. **Add mode toggle:** Single Character vs. Character Sheet — with different default settings for each
+6. **Explore Regional Prompting** for precise character placement
+7. **Two-Pass generation:** Background first, then inpaint character
+8. **Portrait crop option:** Auto-crop to head/shoulders for hair-focused output