7.7 KiB
7.7 KiB
LLM Studio Prompt Logging
Debug logging for LLM Studio API calls in the Media Describe node.
Overview
When using LLM Studio for media description, it's crucial to verify that prompts are being sent correctly. The verbose logging system provides detailed insight into:
- System prompts being constructed
- User prompts being sent
- Model configuration options
- Media type and model type selection
Enabling Verbose Logging
Set verbose=True in the LLM Studio - Options node to enable detailed prompt logging.
Log Output Format
Prompt Construction Log
Shows the initial configuration after prompts are built:
============================================================
🔧 LLM Studio Prompt Construction:
============================================================
Media Type: image
Model Type: Text2Image
Describe Clothing: True
Change Clothing Color: False
Describe Hair Style: True
Describe Bokeh: True
Describe Subject: True
============================================================
Frame Processing Log (Videos)
Shows prompts when processing video with all frames sent in a single request:
📹 Extracting 5 frames from video...
✅ Extracted frame 1/5
✅ Extracted frame 2/5
✅ Extracted frame 3/5
✅ Extracted frame 4/5
✅ Extracted frame 5/5
============================================================
🔍 LLM Studio Video Prompt Debug:
============================================================
📝 System Prompt:
You are an expert assistant specialized in analyzing and verbalizing input videos for cinematic-quality video transformation using the Wan 2.2 + VACE workflow.
DECISIVENESS REQUIREMENT: Always provide definitive, certain descriptions...
Return **only** a single valid JSON object with **exactly six** string fields in this exact order:
1. "subject" - Detailed description of the main subject
2. "clothing" - Clothing and style details
3. "movement" - Actions, motion, and choreography across frames
4. "scene" - Setting, environment, and background elements
5. "cinematic_aesthetic_control" - Lighting, camera details, and rendering cues
6. "stylization_tone" - Mood/genre descriptors
Each field's value is one fully formed paragraph (a single string) for that category.
📝 User Prompt:
Please analyze this video and provide a detailed description in the JSON format specified in the system prompt.
📊 Sending 5 frames in single request
============================================================
✅ Video analysis complete: {"subject": "The video begins with a close-up of...
Image Processing Log
Shows prompts for single image analysis:
============================================================
🔍 LLM Studio Image Prompt Debug:
============================================================
📝 System Prompt:
Generate a Wan 2.2 optimized text to image prompt. You are an expert assistant specialized in analyzing and verbalizing input media for instagram-quality posts using the Wan 2.2 Text to Image workflow.
DECISIVENESS REQUIREMENT: Always provide definitive, certain descriptions...
Return **only** a single valid JSON object with **exactly six** string fields in this exact order:
1. "subject" - Detailed description of the main subject
2. "clothing" - Clothing and style details
3. "movement" - Pose, gesture, or implied motion
4. "scene" - Setting, environment, and background elements
5. "cinematic_aesthetic" - Lighting, camera details, and rendering cues
6. "stylization_tone" - Mood/genre descriptors
Each field's value is one fully formed paragraph (a single string) for that category.
📝 User Prompt:
Please analyze this image and provide a detailed description in the JSON format specified in the system prompt.
============================================================
Prompt Variations
Text2Image Mode (Images)
System Prompt Pattern:
- Specifies Wan 2.2 Text to Image workflow
- Requires JSON output with 6 fields
- Uses "cinematic_aesthetic" field name
- Emphasizes decisiveness
User Prompt:
- Simple analysis request
- References JSON format from system prompt
ImageEdit Mode (Images)
System Prompt Pattern:
- Focuses on Qwen-Image-Edit instructions
- Requires single-sentence output
- No JSON structure
- Emphasizes decisiveness
User Prompt:
- Requests single-sentence instruction
- References guidelines from system prompt
Text2Image Mode (Videos)
System Prompt Pattern:
- Specifies Wan 2.2 + VACE workflow
- Requires JSON output with 6 fields
- Uses "cinematic_aesthetic_control" field name (different from images!)
- Emphasizes movement and choreography
- Emphasizes decisiveness
User Prompt:
- Video-specific analysis request
- References JSON format from system prompt
Common Issues Debugged
1. Empty or Generic Responses
Symptom: LLM returns generic captions instead of structured JSON
Debug Check:
- Verify system prompt includes JSON structure definition
- Check that model_type is correctly set (Text2Image vs ImageEdit)
- Ensure media_type matches actual media (image vs video)
2. Wrong Field Names
Symptom: JSON response uses different field names than expected
Debug Check:
- Video prompts use "cinematic_aesthetic_control"
- Image prompts use "cinematic_aesthetic"
- Verify prompt construction matches media type
3. Uncertain Language
Symptom: Descriptions include "appears to be", "seems to be", etc.
Debug Check:
- Verify DECISIVENESS REQUIREMENT appears in system prompt
- Check that prompt includes forbidden phrases list
4. Non-JSON Responses
Symptom: Response is paragraph format instead of JSON
Debug Check:
- Verify "Return only a single valid JSON object" appears in system prompt
- Check max_tokens is sufficient (1000 for videos, 500 for images)
- Ensure model supports vision input
- For videos, verify frames are being sent correctly in content array
5. Poor Video Understanding (FIXED)
Symptom: Video descriptions lack temporal coherence or motion tracking
Solution: As of the latest update, all video frames are sent in a single API request, allowing the model to:
- Understand temporal relationships between frames
- Track motion and action progression across the sequence
- Generate more coherent descriptions
- Reduce network overhead (1 request instead of N requests)
- Better leverage Qwen3-VL's video understanding capabilities
6. JSON Not Being Parsed (FIXED)
Symptom: Control Panel shows entire JSON string in "Subject" field instead of parsed fields
Root Cause: Code was ignoring the JSON response and creating a simplified structure
Solution: Updated code to:
- Parse JSON response from LLM Studio
- Normalize field names (
cinematic_aesthetic_control→cinematic_aesthetic) - Ensure all required fields exist
- Fall back to simple structure only if JSON parsing fails
Debug Output:
============================================================
🔍 Parsing LLM Response:
============================================================
Raw response:
{"subject": "A close-up of a person's mouth...", "clothing": "...", ...}
============================================================
✅ Successfully parsed JSON response
Fields found: ['subject', 'clothing', 'movement', 'scene', 'cinematic_aesthetic', 'stylization_tone']
Code Location
Implementation: nodes/media_describe/media_describe.py
Key methods:
_process_with_llm_studio()- Lines 1773-2080 (updated)- System prompt construction - Lines 1825-1858
- Video frame extraction - Lines 1888-1930 (single request approach)
- Image logging - Lines 1978-1986