Previously, identify_missing_models grouped missing entries only by filename,
which could merge unrelated models from different categories that happen to
share a filename. This change deduplicates by (category, filename) and tracks
all_node_refs with a reference_count so the frontend can preserve one
resolution per node reference instead of collapsing duplicates.
Frontend changes in resolve_download_methods.js iterate over all_node_refs,
and missing_browser_methods.js now shows '+ N refs' in the missing-model
label and clarifies the locate tooltip when multiple workflow references
share the same missing model.
Replace duplicated filename-to-query logic with a single
\uild_filename_search_queries\ helper in matcher/civarchive/huggingface,
standardize Civitai URL generation in the frontend via \getCivitaiModelUrl\,
unify browser user-agent and response file-size extraction in the downloader
path, and derive workflow node-type category hints from widget categories
rather than maintaining a large static mapping.
Replace duplicated SHA256 normalization, metadata sidecar discovery, trained-words
extraction, type coercion, and cache helpers into shared modules. Clarify that the
_civitai.py/_civarchive.py/_huggingface.py/_workflow_analyzer.py refactorings are
behavior-preserving so future contributors keep behavior aligned in one place.
Extract MODEL_EXTENSIONS and format_size_bytes into core/type_utils, removing duplicates from workflow_analyzer and downloader. Add getFilenameFromPath to render_format_methods and replace inline path-splitting logic across JS resolver actions, search, and views.
Replace global log_* wrapper functions with create_module_logger(__name__)
instances across the codebase, enabling proper module-level logger
namespacing. Simplifies core/log_system/log_funcs.py by removing legacy
_normalize_module_path and cleaning up _normalize_module_name. Bumps
log_system config to 2.0.0 and introduces PROJECT_ALIASES for cleaner
log module prefix stripping.
- Extend dynamic widget category hints to capture choice source (static/hybrid/folder_paths)
- Infer workflow schema model candidates from output category hints when widget values match known presets
- Surface auto_download_capable and auto_download_candidate flags in missing model data
- Merge these flags on duplicate missing model entries
- Add cloud-download badge UI with tooltip explaining auto-download context
- Update huggingface author index metadata timestamp
- Expand preferred categories to include clip_vision, ipadapter, sams, ultralytics
- Add clip_gguf alias for text_encoders in both backend and frontend
- Add explicit diffusion_models/text_encoders mappings in CATEGORY_MAP
- Refactor get_model_roots to skip custom_nodes/configs, use template key
aliases, accept new folder keys from known categories, and normalize
ultralytics paths (yolo/bbox, yolo/segm, yolo -> parent)
- Pass missing object through resolve_download_methods and search_panel
so category selection can use widget hints
- Refactor frontend download target resolution:
- getMissingSupportedDownloadCategories respects widget/type hints
- getNodeTypeDownloadCategory splits exact-node map and adds
normalizeNodeCategoryToken / getNodeTypeDownloadCategoryFromTokens
for GGUF and future node families
- getSourceResultDownloadCategory and _sourceSupportLevel accept
missing for unified logic
- Prioritize dynamic widget category hints in workflow analyzer and
include category_hints/folder_key_hints on all model refs
- Add folder count badges and path normalization in missing browser
(normalizeMissingFolderRootPath handles category suffixes / bbox/segm)
- Add get_comfy_root_path() to core/path_utils.py and use it in __init__.py and core/downloader.py instead of inline logic
- Add select_primary_model_file(), parse_size_to_bytes(), get_version_sort_key(), and check_credential_http() to core/type_utils.py
- Replace duplicated credential-check helpers in civitai.py, huggingface.py with check_credential_http()
- Replace duplicated size parsing and version sorting in civarchive.py, civitai.py with shared utilities
- Replace duplicated primary-file selection in civarchive.py, civitai.py, lora_manager_archive.py with select_primary_model_file()
- Replace inline tempfile-based JSON writes in popular.py with write_json_atomic()
- Remove dead reload_model_list() from model_list.py
- Fix workflow_analyzer.py hint prioritization for diffusion_models vs specific output hints
Move CATEGORY_MAP, normalize_download_category, URN_TYPE_MAP,
CIVITAI_API_TYPE_MAP, and CIVARCHIVE_API_TYPE_MAP into type_utils.py
so they are defined once and imported by settings, civitai, civarchive,
and workflow_analyzer instead of being duplicated across modules.
- core/workflow_analyzer.py: add dynamic widget category hint extraction via threaded cache and sentinel-prefix matching; get_model_widget_category_hints now returns all inferred categories including dynamic ones, improving model folder detection for workflows with encoded category metadata.
- web: add ? help icon with context-aware tooltip to download folder selector explaining why a category was auto-selected (preserved manual choice, node type hint, workflow-detected category, default fallback, or best available). Tooltip refreshes live as the user changes the folder selection.
- web/css/resolver-main.css: style the new category help badge.
- metadata/huggingface-author-index.json: update index entry.
- workflow_analyzer: refactor widget category hint resolution and add widget name candidate extraction across widgets, inputs, proxyWidgets, and localized names; expose new helpers get_widget_name_candidates/get_widget_name_hint
- workflow_analyzer: add get_model_widget_category_hint and get_effective_model_category_hint to unify category fallback logic
- workflow_analyzer: include widget_name on model refs emitted by get_node_model_info
- workflow_analyzer: introduce promoted-widget context resolution (_build_promoted_widget_contexts, _find_promoted_widget_targets, etc.) and wire it into analyze_workflow_models so instance widgets and promoted widget locators are tracked with subgraph scope
- lifecycle_graph_methods: refactor locateNodeInGraph into subgraph-aware resolver using findLocateTarget/getGraphNodeById/findSubgraphById/findNodeInGraphTree; support targeting nodes inside specific subgraphs with proper canvas focus and notifications
- missing_browser_methods: use unified getMissingLocateTarget for locate button/display fields; pass subgraphId and isTopLevel to locateNodeInGraph; add promoted-widget-aware tooltips and canLocate handling
Add support for additional ComfyUI model types including diffusers, gligen,
style models (latent/regular upscale), audio encoders, background removal,
frame interpolation, geometry estimation, detection, model patches/photomaker,
and optical flow models.
Backend changes:
- Extend CATEGORY_MAP, URN_TYPE_MAP, and node type hints in workflow analyzer
to cover dozens of new loader node types and singular/plural aliases.
- Add MODEL_OUTPUT_TYPE_TO_CATEGORY and get_node_output_category_hint() so
loader output types can contribute category hints.
- Make widget name matching normalize underscores, hyphens, and spaces.
- Improve directory discovery to enumerate folders from folder_paths while
filtering unsupported categories.
- Include folders as model entries in directory scanning (with a diffusers
guard requiring model_index.json).
Frontend changes:
- Update download target category maps, display names, token name maps, and
per-node widget index lists to match the expanded backend coverage.
- Add model_title match type for CivArchive and CivitAI sources, resolving
extensionless workflow values by title similarity (threshold 82.0)
- Recognize new node types and widget indices across model categories
(GGUF, CLIP, ControlNet, StyleModel, Hypernetwork, etc.)
- Scan known model combo widget indices even when the stored value has no
file extension, treating non-placeholder strings as model references
- Strip known model extensions during workflow URL lookup so filename stems
match regardless of extension
- Update HuggingFace author index metadata
- README.md: updated badge labelColor from 111827 to 1a1a1a for consistency
- core/workflow_analyzer.py: add LoaderGGUF and gguf_name mapping to diffusion_models
- web/resolver/model_resolver.js: rename panel title from 'Model Resolver' to 'Resolver'
Pass available_models through to the workflow analyzer, enabling
case-sensitive relative path lookups against scanned model data.
Normalize paths for comparison while preserving letter casing.
Skip irrelevant categories (custom_nodes, configs) upfront.
Fall back gracefully when no scanner data is available.
Also ignore the tests/ directory in .gitignore.
Replace all link dialog terminology with resolver terminology
throughout the codebase, including module names, CSS classes,
IDs, JS variables, SVG gradient IDs, and log module names.
Remove the old web/linker/ tree and introduce the new web/resolver/
tree with parallel structure. Also replace core/linker.py with
core/resolver.py.
- Add 'node_title' field to model info dictionaries in get_node_model_info
- Display custom node titles in missing models dialog when available
- Show 'NodeLabel #id · CustomTitle' format for better model identification
Performance and correctness improvements across model scanning, matching, workflow analysis, and CivitAI URL handling:
- Add short-lived in-memory cache for get_model_files (with force_rescan and invalidate_model_files_cache) to reduce frequent directory scans.
- Pass scanned available_models into workflow analysis to avoid rescanning and build category-indexed lists plus an ordered-candidates cache for faster candidate ordering.
- Improve analyze_and_find_matches to special-case urn:air: URNs (use filename suffix when present) and skip fuzzy matching when appropriate.
- Rewrite find_matches to maintain only the top N results via heapq, cache normalized candidate path/filename values on candidate dicts, and early-exit on max_results<=0 to reduce allocations and CPU.
- Add helper functions in CivitAI source to extract/build stable image page URLs (_extract_civitai_image_id, _build_civitai_image_url) and use them when populating image metadata.
- Propagate available_models into get_node_model_info and analyze_workflow_models so subroutines reuse the same scanned list.
- Guard frontend and backend folder lookups to return early for unknown categories to avoid unnecessary work.
These changes aim to reduce IO/CPU during frequent matching operations and produce more accurate/efficient matching for URNs and local models.
Replace direct uses of the logging module and module-level loggers with the centralized log_system.log_funcs (log_debug/log_info/log_warn/log_error/log_exception) across multiple modules. Updated imports and removed unused logging.getLogger() instances in downloader, scanner, sources (civitai, huggingface, model_list, popular), workflow_analyzer, workflow_updater, and __init__.py. Minor message/text tweaks and import formatting applied; behavior unchanged aside from routing logs through the shared logging helpers.
Introduce a new logging subsystem under core/log_system (logger.py, config.py, log_funcs.py) implementing AzLogs (LogLevel enum, colored console output, optional file rotation) and a lightweight wrapper (log_debug/info/warn/error/exception) with safe fallback when AzLogs is unavailable. Integrate these log functions across many core modules (downloader, linker, matcher, scanner, workflow_analyzer, workflow_updater, sources and related files) and export logger and helper log functions from core.__init__. Replace several direct logging calls in linker with the new log_* helpers and update core.__init__ exports. Also include small code style and robustness tweaks (path/filename normalization, JSON/file handling, minor refactors and formatting fixes) to improve matching and model-list/popular lookups.
Add the "Lora Stacker (LoraManager)" node type to NODE_TYPE_TO_CATEGORY_HINTS and to the list of text-based Lora loader types so the analyzer treats the LoraManager Stacker as a lora loader and applies the same model-info extraction logic.
Add support for the gguf model format across the codebase and improve handling of nested/dict-type widget model references. Changes include:
- Add ".gguf" to model extension lists in core/linker.py, core/scanner.py, core/sources/huggingface.py, and core/workflow_analyzer.py so gguf files are detected and suggested.
- Update HuggingFace search/repo listing to consider gguf files and improve some request/response handling and logging.
- Improve scanner implementation: formatting/PEP8 cleanup, robust extension checks, normalized MODEL_EXTENSIONS set, safer path handling, and clearer logging.
- Add NESTED_MODEL_KEYS and logic in workflow_analyzer to detect model filenames stored inside dict-type widget values (e.g. rgthree Power Lora Loader) and record a nested_key for resolution.
- Propagate nested_key through apply_resolution mapping and update update_model_path to write resolved paths back into nested dict keys when present (instead of only replacing the top-level widget value).
- Update web UI (web/linker.js) regexes to strip/display gguf extensions in model name formatting.
These changes allow gguf models to be discovered, resolved and updated in workflows, and enable support for nodes that store model references inside nested dict fields.
Filter out LoraManager (LoraLoaderV2) entries that already exist locally so they don't appear as missing, and skip matching for those entries. Integrate scanner data to detect local lora files (full_path/exists) in workflow analysis. Add special handling in update_model_path/update_workflow_nodes to update LoraManager lora lists and the formatted lora text widget (widgets_values[2] and [1]) when a new lora is resolved, using mapping fields is_lora_v2 and original_lora_name. Improve convert_to_relative_path and get_base_directory_for_model to prefer folder_paths lookup and preserve exact ComfyUI filename formats. In the web UI, make resolve button IDs unique by including a missingIndex, avoid duplicate listeners, forward is_lora_v2/original_lora_name to the resolve API, and add extra logging for debugging.
Add support for text-based Lora loaders (LoraLoaderV2 / LoraManager) and mark model refs with activity/connection state, plus UI improvements.
- __init__.py: clarify comment and propagate new flags from refs (is_lora_v2, active, connected).
- core/workflow_analyzer.py: reformatting and small style changes; add LoraLoaderV2 and LoraManager node hints; detect node connected/bypassed state; parse text-based Lora widget lists (name/strength/active) and emit model refs with strength, is_lora_v2, active and connected flags; include connected flag for URN and file refs; keep existing URN/file resolution and dedup logic but annotate refs with subgraph/context info.
- web/linker.js: group loaded models into active/inactive per-category, add UI filter buttons (All / Active / Inactive), separate copy buttons for active/inactive/all, update display to indicate active vs inactive models and copy behavior, and add global helper functions for filtering and clipboard operations.
These changes enable accurate detection of LoraV2 loader states and provide a clearer UI for copying model spec strings while preserving existing behavior for standard loaders.
Detect ComfyUI-AIR URNs in workflows, resolve them to expected filenames via the CivitAI API, and prefer those filenames for matching.
Key changes:
- core/workflow_analyzer.py: added URN_REGEX, URN_TYPE_MAP, is_model_filename() now recognizes URNs, and get_node_model_info() emits structured URN info (is_urn, urn fields, category, exists=False).
- core/sources/civitai.py: added resolve_urn(model_id, version_id) with caching, API fetch of model/version and primary file, and result shaping (model_name, version_name, expected_filename, files).
- core/linker.py: import resolve_urn and call it for missing URNs, attach civitai_info/expected_filename, and prefer expected_filename as matching target; cleaned up deduplication logic.
- web/linker.js: UI shows URN original plus resolved CivitAI model/version labels where available.
- TODO.md: added plan/notes for URN detection and next steps.
This improves detection and fuzzy matching for models referenced by URNs so missing model resolution is more accurate. Further matcher and updater enhancements remain TODOs.
Dariusz L