- Tag.get_or_create now returns None for invalid tag names instead of creating error tags
- Add null checks before using tag.id in all callers
- Adjust tag validation: non-Chinese tags now limited to 8 words or 40 characters
- Change Tag.get_or_create to return None for invalid tag names
- Add None checks at all call sites to prevent AttributeError
- Protects against invalid Chinese names (>16 chars) and other language names (>8 words)
- Ensures system stability when encountering problematic tag data
Important performance optimization for handling large tag datasets:
- When total tags > 8192, limit 'pos' type tags to top 4096 by count
- Return all non-pos tags without limitation
- Significantly reduces memory usage and load time for large databases
- Maintains functionality while improving performance with massive tag counts
- Add tag name validation in Tag.get_or_create method
- Chinese names: max 16 characters
- Other languages: max 8 words
- Return error tags with uppercase English names when validation fails
- Ensure error tags are created through get_or_create to maintain database consistency
- Clean up database: removed 487,938 non-custom tags, keeping only 12 custom tags
- Add IIB_COMFYUI_EXTRACT_ALL_PROMPTS environment variable
- When enabled, extract all CLIPTextEncode prompts instead of following KSampler chain
- Note: All prompts will be merged into positive prompt field
- Add keyword count limit (default 600) for Tag layer to improve performance
- Implement irregular node layout with random perturbations
- Add layer spacing based on adjacent layer widths for consistent gaps
- Add cache result banner with collapse/expand functionality
- Add streaming response utility for LLM API calls
- Update i18n translations for new cache banner texts
- Optimize tag graph: increase max tags to 500, add level-2 tiered requirements
- Fix keyword deduplication and null handling in tag graph
- Add get_all_keywords_frequency method to TopicTitleCache
- Initialize keyword frequency from historical cached cluster keywords
- Prioritize top 100 high-frequency keywords when generating new keywords
- Update LLM prompt to prefer existing keywords from frequency list
- Reduce duplicate/similar keyword generation across clusters
- Add streaming support for tag_graph LLM requests
- Increase LLM timeout and retry limits for better reliability
Features:
- Tag-based relationship graph using force-directed layout
- Hybrid weight calculation (frequency + TF-IDF)
- Automatic community detection using Louvain algorithm
- Interactive visualization with ECharts
- Support for Tag and Cluster nodes
- Click to search images by tag
- Top-N filtering to avoid visual clutter
Backend:
- New API endpoint: POST /db/cluster_tag_graph
- Tag weight calculation with configurable alpha
- Category inference (character/style/scene/object)
- Community detection (requires networkx, optional)
Frontend:
- New component: TagRelationGraph.vue
- Switch between Cluster Cards and Tag Graph views
- Interactive graph with drag/zoom/pan
- Node details panel with search functionality
- Stats panel showing tags/clusters/images count
Documentation:
- Added TAG_GRAPH_README.md with usage guide
Co-Authored-By: Claude <noreply@anthropic.com>
- Lock requests to >=2.0.0,<3.0.0 (current: 2.32.5)
- Lock numpy to >=2.0.0,<3.0.0 (current: 2.4.0)
- Lock hnswlib to >=0.0.0,<1.0.0 (current: 0.8.0)
- Add runtime version info for requests, numpy, hnswlib in /version endpoint
- Add audio file recognition in backend (tool.py) and frontend (file.ts)
- Support audio formats: mp3, wav, ogg, flac, m4a, aac, wma
- Add audio player in TikTok view with starfield background animation
- Add audio modal similar to video modal
- Update file type filter to multi-select checkboxes (image/video/audio)
- Add filename display in TikTok view with Douyin-style bottom gradient
- Add long press to hide/show controls in TikTok view
- Fix audio resource cleanup when closing TikTok view
- Add Audio media type tag in database indexing
- Add i18n translations for audio related strings
Backend optimizations:
- Refactor get_random_images() using SQLite ORDER BY RANDOM()
- Simplify from 60+ lines to ~20 lines with better performance
- Remove complex multi-cycle index calculations
- Maintain file existence check and auto-cleanup
ComfyUI support:
- Add fallback to 'parameters' field in PNG metadata
- Improve detection for ComfyUI-generated images
Frontend improvements:
- Smart prompt display: auto-detect tag vs natural language format
- Add getTextLength() with Chinese character weight (1 char = 3 letters)
- Refactor isTagStylePrompt() with clear rules:
* Any tag length > 50 natural language format
* Average tag length > 30 natural language format
- Natural language prompts shown as paragraphs
- Tag-style prompts shown as badges
- Better support for Chinese/English mixed content
Build:
- Update frontend asset references
- Add AutoTagMatcher singleton class for automatic tag application
- Support filtering by multiple fields (prompt, model, sampler, etc.)
- Add frontend UI for managing auto-tag rules in global settings
- Support AND logic for multiple conditions within a rule
- Integrate with image indexing process
- Add natural language vs tag-style prompt detection and rendering
- Include rule descriptions and usage guidance in UI
- Auto-reload rules when updated via API
- Added video tag support for enhanced media file handling:
* Video files can now use tag-based search functionality
* Added support for reading video generation info from txt files
* Enhanced EXIF data handling for video media types
- Implemented random sort feature for image grid:
* Added random sort toggle button with visual indicators (🎲/📅)
* Supports both random and date-based sorting modes
* Optimized pagination for random sorting with offset-based cursors
- Complete internationalization support:
* Added translations for random sort and date sort in all languages
* Supported languages: zh-hans, en, de, zh-hant
* Improved UI consistency across language variants
- Backend improvements:
* Enhanced database queries to support random sorting
* Added random_sort parameter to API endpoints
* Improved error handling for video file processing
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
zanllp
wuqinchuan
bupro
Hugs288
younyokel