Add file_path as a tie-breaker for all sort modes in ModelCache, BaseModelService, LoraService, and RecipeCache to ensure deterministic ordering when primary keys are identical. Resolves issue #859.
Make preview file discovery case-insensitive so files with uppercase
extensions like .WEBP are found on case-sensitive filesystems. Also
explicitly list image/webp in the file picker accept attribute for
broader browser compatibility.
https://claude.ai/code/session_01SgT2pkisi27bEQELX5EeXZ
- Add LOG10(post_count) weighting to BM25 score for better relevance ranking
- Prioritize tag_name prefix matches above alias matches using CASE statement
- Remove frontend re-scoring logic to trust backend排序 results
- Fix pagination consistency: page N+1 scores <= page N minimum score
Key improvements:
- '1girl' (6M posts) now ranks #1 instead of #149 for search '1'
- tag_name prefix matches always appear before alias matches
- Popular tags rank higher than obscure ones with same prefix
- Consistent ordering across pagination boundaries
Test coverage:
- Add test_search_tag_name_prefix_match_priority
- Add test_search_ranks_popular_tags_higher
- Add test_search_pagination_ordering_consistency
- Add test_search_rank_score_includes_popularity_weight
- Update test data with 15 tags starting with '1'
Fixes issues with autocomplete dropdown showing inconsistent results
when scrolling through paginated search results.
- Add BatchImportService with concurrent execution using asyncio.gather
- Implement AdaptiveConcurrencyController with dynamic adjustment
- Add input validation for URLs and local paths
- Support duplicate detection via skip_duplicates parameter
- Add WebSocket progress broadcasting for real-time updates
- Create comprehensive unit tests for batch import functionality
- Update API handlers and route registrations
- Add i18n translation keys for batch import UI
Remove .safetensors/.ckpt/.pt/.bin extensions from model names in autocomplete
suggestions to improve UX and search relevance:
Frontend (web/comfyui/autocomplete.js):
- Add _getDisplayText() helper to strip extensions from model paths
- Update _matchItem() to match against filename without extension
- Update render() and createItemElement() to display clean names
Backend (py/services/base_model_service.py):
- Add _remove_model_extension() helper method
- Update _relative_path_matches_tokens() to ignore extensions in matching
- Update _relative_path_sort_key() to sort based on names without extensions
Tests (tests/services/test_relative_path_search.py):
- Add tests to verify 's' and 'safe' queries don't match all .safetensors files
Fixes issue where typing 's' would match all .safetensors files and cluttered
suggestions with redundant extension names.
Refactor force_refresh path to use thread pool execution instead of blocking
the event loop shared with ComfyUI. Key changes:
- Fix 1: Route force_refresh through _initialize_recipe_cache_sync() in thread pool
- Fix 2: Add GIL release points (time.sleep(0)) every 100 files in sync loops
- Fix 3: Move RecipeCache.resort() to thread pool via run_in_executor
- Fix 4: Persist cache automatically after force_refresh
- Fix 5: Increase yield frequency in _enrich_cache_metadata (every recipe)
This eliminates the ~5 minute freeze when rebuilding 30K recipe cache.
Fixes performance issue where ComfyUI became unresponsive during recipe
scanning due to shared Python event loop blocking.
- Allow empty sha256 when hash_status is 'pending' in cache entry validator
- Add on-demand hash calculation during bulk metadata refresh for checkpoints
with pending hash status
- Add comprehensive tests for both fixes
Fixes issue where checkpoints in extra paths were not visible in UI and
not processed during bulk metadata refresh due to empty sha256.
Checkpoints are typically large (10GB+). This change delays SHA256
hash calculation until metadata fetch from Civitai is requested,
significantly improving initial scan performance.
- Add hash_status field to BaseModelMetadata
- CheckpointScanner skips hash during initial scan
- On-demand hash calculation during Civitai fetch
- Background bulk hash calculation support
Introduce extra_folder_paths feature to allow users to add additional
model roots that are managed by LoRA Manager but not shared with ComfyUI.
Changes:
- Add extra_folder_paths support in SettingsManager (stored per library)
- Add extra path attributes in Config class (extra_loras_roots, etc.)
- Merge folder_paths with extra_folder_paths when applying library settings
- Update LoraScanner, CheckpointScanner, EmbeddingScanner to include
extra paths in their model roots
- Add comprehensive tests for the new functionality
This enables users to manage models from additional directories without
modifying ComfyUI's model folder configuration.
Replace _SYNC_KEYS (37 keys) with _NO_SYNC_KEYS (5 keys) in SettingsHandler.
New settings automatically sync to frontend unless explicitly excluded.
Changes:
- SettingsHandler now syncs all settings except those in _NO_SYNC_KEYS
- Added keys() method to SettingsManager for iteration
- Updated tests to use new behavior
Benefits:
- No more missing keys when adding new settings
- Reduced maintenance burden
- Explicit exclusions for sensitive/internal settings only
Fixes: #86
Add Early Access version support with filtering and improved UI:
Backend:
- Add is_early_access and early_access_ends_at fields to ModelVersionRecord
- Implement two-phase EA detection (bulk API + single API enrichment)
- Add hide_early_access_updates setting to filter EA updates
- Update has_update() and has_updates_bulk() to respect EA filter setting
- Add _enrich_early_access_details() for precise EA time fetching
- Fix setting propagation through base_model_service and model_update_service
Frontend:
- Add smart relative time display for EA (in Xh, in Xd, or date)
- Replace EA label with clock icon in metadata (fa-clock)
- Show Download button with bolt icon for EA versions (fa-bolt)
- Change EA badge color to #F59F00 (CivitAI Buzz theme)
- Fix toggle UI for hide_early_access_updates setting
- Add translation keys for EA time formatting
Tests:
- Update all tests to pass with new EA functionality
- Add test coverage for EA filtering logic
Closes#815
- Backend: Support limit=0 to return all tags in top-tags API
- Frontend: Remove tags limit setting and fetch all tags by default
- UI: Implement virtual scrolling in TagsModal for performance
- Initial display 200 tags, load more on scroll
- Show all results when searching
- Remove lora_pool_tags_limit setting to simplify UX
Fixes#819
Fix KeyError when 'hashes', 'name', or 'model' fields are missing from
Civitai API responses. Use .get() with defaults instead of direct dict
access in:
- LoraMetadata.from_civitai_info()
- CheckpointMetadata.from_civitai_info()
- EmbeddingMetadata.from_civitai_info()
- RecipeScanner._get_hash_from_civitai()
- DownloadManager._process_download()
Fixes#820
Add a segmented toggle in the Filter Panel to switch between 'Any' (OR)
and 'All' (AND) logic when filtering by multiple include tags.
Changes:
- Backend: Add tag_logic field to FilterCriteria and ModelFilterSet
- Backend: Parse tag_logic parameter in model handlers
- Frontend: Add segmented toggle UI in filter panel header
- Frontend: Add interaction logic and state management for tag logic
- Add translations for all supported languages
- Add comprehensive tests for the new feature
Closes#802
Add automatic cleanup of default values from settings.json to keep configuration files minimal and focused on user customizations. Introduces a threshold-based cleanup that only removes default values when the file contains a significant number of them (10+), preserving small template-based configurations while cleaning up legacy bloated files.
Key changes:
- Add DEFAULT_KEYS_CLEANUP_THRESHOLD constant to control cleanup aggressiveness
- Implement _cleanup_default_values_from_disk() method that removes default values from disk while keeping them available in memory
- Modify _ensure_default_settings() to only save when existing values are updated, not when defaults are inserted
- Update _serialize_settings_for_disk() to only persist settings that differ from defaults
- Add cleanup call during initialization for existing settings files
This reduces file size and noise in settings.json while maintaining full functionality at runtime.
- Use modelVersionId as fallback for all loras in fingerprint calculation (not just deleted)
- Add URL-based duplicate detection using source_path field
- Combine both fingerprint and URL-based duplicate detection in API response
- Fix _download_remote_media return type and unbound variable issue
When a deleted model is checked against the SQLite archive and not found, the `db_checked` flag was set in memory but never saved to disk. This occurred because the save operation was only triggered when `civitai_api_not_found` was True, which is not the case for deleted models (since the CivitAI API is not attempted). As a result, deleted models would be rechecked on every refresh instead of being skipped.
Changes:
- Introduce a `needs_save` flag to track when metadata state is updated
- Save metadata whenever `db_checked` is set to True, regardless of API status
- Ensure `last_checked_at` is set for SQLite-only attempts
- Add regression test to verify the fix
Refactor recipe lookup logic to improve efficiency from O(n²) to O(n + m):
- Build recipe_by_id dictionary for O(1) recipe ID lookups
- Simplify persisted_by_path construction using recipe_id extraction
- Add fallback lookup by recipe_id when path lookup fails
- Maintain same functionality while reducing computational complexity
When enable_metadata_archive_db=True, the previous filter logic would
repeatedly try to fetch metadata for models that were already confirmed
to not exist on CivitAI (from_civitai=False, civitai_deleted=True).
The fix adds a skip condition to exclude models that:
1. Are confirmed not from CivitAI (from_civitai=False)
2. Are marked as deleted/not found on CivitAI (civitai_deleted=True)
3. Either have no archive DB enabled, or have already been checked (db_checked=True)
This prevents unnecessary API calls to CivArchive for user-trained models
or models from non-CivitAI sources.
Fixes repeated "Error fetching version of CivArchive model by hash" logs
for models that will never be found on CivitAI/CivArchive.
- Add cache entry validator service for data integrity checks
- Add cache health monitor service for periodic health checks
- Enhance model cache and scanner with validation support
- Update websocket manager for health status broadcasting
- Add initialization banner service for cache health alerts
- Add comprehensive test coverage for new services
- Update translations across all locales
- Refactor sync translation keys script
- Remove backward compatibility code for `model_type` in `ModelScanner._build_cache_entry()`
- Update `CheckpointScanner` to only handle `sub_type` in `adjust_metadata()` and `adjust_cached_entry()`
- Delete deprecated aliases `resolve_civitai_model_type` and `normalize_civitai_model_type` from `model_query.py`
- Update frontend components (`RecipeModal.js`, `ModelCard.js`, etc.) to use `sub_type` instead of `model_type`
- Update API response format to return only `sub_type`, removing `model_type` from service responses
- Revise technical documentation to mark Phase 5 as completed and remove outdated TODO items
All cleanup tasks for the model type refactoring are now complete, ensuring consistent use of `sub_type` across the codebase.
This commit resolves the semantic confusion around the model_type field by
clearly distinguishing between:
- scanner_type: architecture-level (lora/checkpoint/embedding)
- sub_type: business-level subtype (lora/locon/dora/checkpoint/diffusion_model/embedding)
Backend Changes:
- Rename model_type to sub_type in CheckpointMetadata and EmbeddingMetadata
- Add resolve_sub_type() and normalize_sub_type() in model_query.py
- Update checkpoint_scanner to use _resolve_sub_type()
- Update service format_response to include both sub_type and model_type
- Add VALID_*_SUB_TYPES constants with backward compatible aliases
Frontend Changes:
- Add MODEL_SUBTYPE_DISPLAY_NAMES constants
- Keep MODEL_TYPE_DISPLAY_NAMES as backward compatible alias
Testing:
- Add 43 new tests covering sub_type resolution and API response
Documentation:
- Add refactoring todo document to docs/technical/
BREAKING CHANGE: None - full backward compatibility maintained
Moves onboarding_completed and dismissed_banners from localStorage
to backend settings (settings.json) to survive incognito/private
browser modes.
Fixes#786
- Add aliases column to tags table to store comma-separated alias lists
- Update FTS schema to version 2 with searchable_text field containing tag names and aliases
- Implement schema migration to rebuild index when upgrading from old schema
- Modify search logic to match aliases and return canonical tag with matched alias info
- Update index building to include aliases in searchable text for FTS matching
This enables users to search for tag aliases (e.g., "miku") and get results for the canonical tag (e.g., "hatsune_miku") with indication of which alias was matched.
Remove all autocomplete.txt parsing logic and fallback code, simplifying
the service to use only TagFTSIndex for Danbooru/e621 tag search
with category filtering.
- Remove WordEntry dataclass and _words_cache, _file_path attributes
- Remove _determine_file_path(), get_file_path(), load_words(), save_words(),
get_content(), _parse_csv_content() methods
- Simplify search_words() to only use TagFTSIndex, always returning
enriched results with {tag_name, category, post_count}
- Remove GET/POST /api/lm/custom-words endpoints (unused)
- Keep GET /api/lm/custom-words/search for frontend autocomplete
- Rewrite tests to focus on TagFTSIndex integration
This reduces code by 446 lines and removes untested pysssss plugin
integration. Feature is unreleased so no backward compatibility needed.
- Centralize cache path resolution in new py/utils/cache_paths.py module
- Migrate legacy cache files to organized structure: {settings_dir}/cache/{model|recipe|fts|symlink}/
- Automatically clean up legacy files after successful migration with integrity verification
- Update Config symlink cache to use new path and migrate from old location
- Simplify service classes (PersistentModelCache, PersistentRecipeCache, RecipeFTSIndex, TagFTSIndex) to use centralized migration logic
- Add comprehensive test coverage for cache paths and automatic cleanup
- Add TagFTSIndex service for fast SQLite FTS5-based tag search (221k+ tags)
- Implement command-mode autocomplete: /char, /artist, /general, /meta, etc.
- Support category filtering via category IDs or names
- Return enriched results with post counts and category badges
- Add UI styling for category badges and command list dropdown
Adds custom words autocomplete functionality similar to comfyui-custom-scripts,
with the following features:
Backend (Python):
- Create CustomWordsService for CSV parsing and priority-based search
- Add API endpoints: GET/POST /api/lm/custom-words and
GET /api/lm/custom-words/search
- Share storage with pysssss plugin (checks for their user/autocomplete.txt first)
- Fallback to Lora Manager's user directory for storage
Frontend (JavaScript/Vue):
- Add 'custom_words' and 'prompt' model types to autocomplete system
- Prompt node now supports dual-mode autocomplete:
* Type 'emb:' prefix → search embeddings
* Type normally → search custom words (no prefix required)
- Add AUTOCOMPLETE_TEXT_PROMPT widget type
- Update Vue component and composable types
Key Features:
- CSV format: word[,priority] compatible with danbooru-tags.txt
- Priority-based sorting: 20% top priority + prefix + include matches
- Preview tooltip for embeddings (not for custom words)
- Dynamic endpoint switching based on prefix detection
Breaking Changes:
- Prompt (LoraManager) node widget type changed from
AUTOCOMPLETE_TEXT_EMBEDDINGS to AUTOCOMPLETE_TEXT_PROMPT
- Removed standalone web/comfyui/prompt.js (integrated into main widgets)
Fixes comfy_dir path calculation by prioritizing folder_paths.base_path
from ComfyUI when available, with fallback to computed path.
Introduce a new PersistentRecipeCache service that stores recipe metadata in an SQLite database to significantly reduce application startup time. The cache eliminates the need to walk directories and parse JSON files on each launch by persisting recipe data between sessions.
Key features:
- Thread-safe singleton implementation with library-specific instances
- Automatic schema initialization and migration support
- JSON serialization for complex recipe fields (LoRAs, checkpoints, generation parameters, tags)
- File system monitoring with mtime/size validation for cache invalidation
- Environment variable toggle (LORA_MANAGER_DISABLE_PERSISTENT_CACHE) for debugging
- Comprehensive test suite covering save/load cycles, cache invalidation, and edge cases
The cache improves user experience by enabling near-instantaneous recipe loading after the initial cache population, while maintaining data consistency through file change detection.
Add Lora Cycler node that cycles through LoRAs sequentially from a filtered pool. Supports configurable sort order, strength settings, and persists cycle progress across workflow save/load.
Backend:
- New LoraCyclerNode with cycle() method
- New /api/lm/loras/cycler-list endpoint
- LoraService.get_cycler_list() for filtered/sorted list
Frontend:
- LoraCyclerWidget with Vue.js component
- useLoraCyclerState composable
- LoraCyclerSettingsView for UI display
Fixes a critical bug in FTS query building where multi-word searches
with field restrictions incorrectly used OR between all word+field
combinations instead of requiring ALL words to match within at least
one field.
Example: searching "cute cat" in {title, tags} previously produced:
title:cute* OR title:cat* OR tags:cute* OR tags:cat*
Which matched recipes with ANY word in ANY field.
Now produces:
(title:cute* title:cat*) OR (tags:cute* tags:cat*)
Which requires ALL words to match within at least one field.
Also adds fallback to fuzzy search when FTS returns empty results,
improving search reliability.
Co-Authored-By: Claude <noreply@anthropic.com>
- Add execution_seed and next_seed parameters to support deterministic randomization across batch executions
- Separate UI display generation from execution stack generation to maintain consistency in batch queues
- Update LoraService to accept optional seed parameter for reproducible randomization
- Ensure each execution with a different seed produces unique results without affecting global random state
Introduce a new RecipeFTSIndex class that provides fast prefix-based search across recipe fields (title, tags, LoRA names/models, prompts) using SQLite's FTS5 extension. The implementation supports sub-100ms search times for large datasets (20k+ recipes) and includes asynchronous indexing, incremental updates, and comprehensive unit tests.
CivitAI does not distinguish between checkpoint and diffusion model types -
both are labeled as "checkpoint". For certain base model types like
"ZImageTurbo", all models are actually diffusion models and should be
saved to the unet/diffusion model folder instead of the checkpoint folder.
- Add DIFFUSION_MODEL_BASE_MODELS constant for known diffusion model types
- Add default_unet_root setting with auto-set logic
- Route downloads to unet folder when baseModel matches known diffusion types
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When users paste CivArchive URLs, the system now fetches metadata from
CivArchive API first instead of Civitai. This prevents download failures
when a model has been deleted from Civitai but remains available on
CivArchive with alternative mirrors.
Changes:
- Source-aware metadata fetching: Uses CivArchive API when source='civarchive'
- URL prioritization: Prefers non-Civitai mirrors for CivArchive downloads
- Fallback mechanism: Falls back to default provider if CivArchive fails
Fixes#769
- Add folder_include parameter support in backend API handlers
- Add folder_include to FilterCriteria and implement multi-folder filtering logic
- Update frontend to send all include folders instead of only the first
- Add tests for single/multiple include folders, include with exclude, and non-recursive filtering
Add support for respecting recommended strength values from LoRA usage_tips
when randomizing LoRA selection.
Features:
- New toggle setting to enable/disable recommended strength respect (default off)
- Scale range slider (0-2, default 0.5-1.0) to adjust recommended values
- Uses recommended strength × random(scale) when feature enabled
- Fallbacks to original Model/Clip Strength range when no recommendation exists
- Clip strength recommendations only apply when using Custom Range mode
Backend changes:
- Parse usage_tips JSON string to extract strength/clipStrength
- Apply scale factor to recommended values during randomization
- Pass new parameters through API route and node
Frontend changes:
- Update RandomizerConfig type with new properties
- Add new UI section with toggle and dual-range slider
- Wire up state management and event handlers
- No layout shift (removed description text)
Tests:
- Add tests for enabled/disabled recommended strength in API routes
- Add test verifying config passed to service
- All existing tests pass
Build: Include compiled Vue widgets
- Add blank line after module docstring for better PEP 8 compliance
- Reformat long lines to adhere to 88-character limit using Black-style formatting
- Improve string consistency by using double quotes consistently
- Enhance readability of complex list comprehensions and method calls
- Maintain all existing functionality while improving code structure
- Add `_preprocess_loras_input` method to handle different widget input formats
- Move core randomization logic to `LoraService` for better separation of concerns
- Update `_select_loras` method to use new service-based approach
- Add comprehensive test fixtures for license filtering scenarios
- Include debug print statement for pool config inspection during development
This refactor improves code organization by centralizing business logic in the service layer while maintaining backward compatibility with existing widget inputs.
