EricRollei-Local_LLM_Prompt_Enhancer/docs/BUGFIX_V161.md

Text-to-Image Node Bug Fixes - v1.6.1

Issues Fixed

1. Settings Appearing in Prompt Output ✅ FIXED

Problem: The node was appending settings information to the prompt:

...golden hour sunlight | Settings: camera angle: extreme close-up; lighting: volumetric...

Solution:

• Updated system prompt to explicitly instruct LLM not to include settings
• Changed user prompt to say "Required settings to incorporate" instead of just "Settings"
• Added regex cleanup to strip any "| Settings:" or trailing settings that slip through
• Settings are now incorporated naturally into the description, not appended

Result: Prompts now look like:

masterpiece, best quality, detailed portrait shot from extreme close-up angle 
with volumetric golden hour sunlight...

---

2. Emphasis Syntax Being Broken ✅ FIXED

Problem: User input like (dark skin:1.5) was being converted to dark skin (1.5), losing the emphasis weight.

Solution:

• Added _preserve_emphasis_syntax() function that protects emphasis patterns before LLM processing
• Emphasis patterns like (keyword:1.5) are temporarily replaced with placeholders __EMPHASIS_0__
• After LLM processing, _restore_emphasis_syntax() puts them back
• Supports any format: (text:number) where number can be decimal

Supported Emphasis:

(dark skin:1.5) - Increases weight to 1.5x
(background:0.5) - Decreases weight to 0.5x
(red hair:2.0) - Doubles the weight

---

3. Alternation Syntax Added ✅ NEW FEATURE

Problem: No way to have random selection from options in the prompt.

Solution:

• Added _process_alternations() function
• Supports {option1|option2|option3} syntax
• Randomly picks one option each time
• Processes before LLM to ensure clean input

Supported Alternations:

{apple|banana|orange} - Picks one randomly
{red|blue|green} dress - Color chosen randomly
a {cat|dog|bird} in the garden - Animal chosen randomly

Nested/Combined:

{red|blue|green} (dress:1.2) - Random color with emphasis
{tall|short} woman with {blonde|brunette|red} hair - Multiple alternations

---

4. Tooltip Updated ✅ ADDED

New Tooltip for text_prompt field:

Describe the image you want to generate

Supports:
- Emphasis: (keyword:1.5) to increase weight
- De-emphasis: (keyword:0.5) to decrease weight
- Alternation: {apple|banana|orange} picks one randomly
- Nested: {red|blue|green} (dress:1.2) works too

Users now see usage instructions directly in the GUI!

---

Technical Details

Functions Added

1. _process_alternations(text: str) -> str

   • Uses regex to find {option1|option2|option3} patterns
   • Splits by | and randomly chooses one
   • Handles nested alternations (up to 10 levels)
   • Runs BEFORE LLM processing

2. _preserve_emphasis_syntax(text: str) -> str

   • Finds patterns like (keyword:1.5)
   • Replaces with __EMPHASIS_0__, __EMPHASIS_1__, etc.
   • Stores originals in self._emphasis_store
   • Protects from LLM modification

3. _restore_emphasis_syntax(text: str) -> str

   • Replaces __EMPHASIS_N__ placeholders back to original
   • Runs AFTER LLM processing
   • Clears the emphasis store

System Prompt Updates

Added explicit instructions:

CRITICAL OUTPUT REQUIREMENTS:
- Output ONLY the final image prompt text
- DO NOT include any settings information in your output
- DO NOT append "Settings:" or list the camera/lighting/etc values
- Incorporate settings naturally INTO the description

With examples of correct vs wrong output.

Response Cleanup

Added regex patterns to strip:

# Remove "| Settings: ..." at end
cleaned = re.sub(r'\s*\|\s*[Ss]ettings:.*$', '', cleaned)
# Remove standalone "Settings: ..." at end
cleaned = re.sub(r'\s*[Ss]ettings:.*$', '', cleaned)
# Clean up trailing pipes
cleaned = cleaned.rstrip('|').strip()

---

Usage Examples

Example 1: Emphasis

Input: "a woman with (dark skin:1.5) and (blue eyes:1.3)"
Platform: flux

Output: "masterpiece, best quality, detailed portrait of a woman 
with (dark skin:1.5) and (blue eyes:1.3), photorealistic..."

✅ Emphasis preserved!

Example 2: Alternation

Input: "a {cat|dog|rabbit} playing in the garden"
Platform: flux

Run 1: "...a cat playing in the garden..."
Run 2: "...a dog playing in the garden..."
Run 3: "...a rabbit playing in the garden..."

✅ Random selection works!

Example 3: Combined

Input: "{red|blue|green} (dress:1.2) on a {tall|short} woman"
Platform: pony

Output: "score_9, score_8_up, score_7_up, 1girl, blue (dress:1.2), 
short, detailed_face..."

✅ Both features work together!

Example 4: No Settings Leak

Input: "woman in garden"
Settings: camera_angle=low angle, lighting=golden hour

OLD OUTPUT (WRONG):
"...woman in garden with golden hour lighting | Settings: camera angle: low angle..."

NEW OUTPUT (CORRECT):
"masterpiece, best quality, woman in garden shot from low angle, 
golden hour lighting, professional photography..."

✅ Settings incorporated naturally, not appended!

---

Testing Checklist

• Emphasis syntax (keyword:1.5) preserved
• Alternation syntax {a|b|c} works
• No "| Settings:" in output
• Settings incorporated naturally
• Tooltip shows usage instructions
• Code compiles without errors
• Multiple emphasis in one prompt works
• Multiple alternations in one prompt works
• Nested alternations work
• Combined emphasis + alternation works

---

Files Modified

text_to_image_node.py:

• Updated text_prompt INPUT_TYPES with new tooltip
• Added _process_alternations() function
• Added _preserve_emphasis_syntax() function
• Added _restore_emphasis_syntax() function
• Updated enhance_prompt() to call alternation and emphasis functions
• Updated _build_system_prompt() with explicit anti-settings instructions
• Updated _build_user_prompt() to clarify settings are to be incorporated
• Updated _parse_llm_response() to strip any stray settings info
• Added regex cleanup for "| Settings:" patterns

Lines Changed: ~100 lines modified/added

---

Version

Version: 1.6.1 (Bug Fix Update) Date: October 15, 2025 Status: ✅ Ready to use

---

Next Steps

1. Restart ComfyUI (if running)
2. Test emphasis syntax: (keyword:1.5)
3. Test alternations: {option1|option2}
4. Verify no settings appear in output
5. Check tooltip displays correctly

All bugs fixed and features added! 🎉