Changelog

All notable changes to ComfyUI Eclipse are documented in this file.

A quick note regarding the repackaging allegations: Eclipse does not contain a single repackaged node. Anyone who claims to have looked under the hood and found someone else's code is lying—and is only proving that they have no idea what they're talking about.

Entries follow conventional commit prefixes:

2026-07-11

Version: 4.1.13

Fix
- Image Transform & FX Nodes: Enabled native list-wise output support (is_output_list=True in schemas) for Upscale With Model (v1), Upscale w/wo Model (v2), Image Rescale, Color Match, Filter Adjustments, Filter Adjustments Advanced, Inset & Crop, Soften, and Image Crop by Mask. Returning processed frames as a list rather than stacking them to a single batch tensor preserves list mapping structures downstream. This prevents downstream nodes (such as KSampler) from executing with batched latents (which caused tensor mismatch errors with models like Krea2).
- Image Loader Nodes: Enabled native list-wise output support (is_output_list=True in schemas) for Load Batch From Folder and Load Batch From Folder (Step/Advanced). Returning loaded images/masks as lists directly rather than stacking them into batches avoids huge contiguous RAM/VRAM allocation overheads and prevents list-flattening bugs.

Version: 4.1.12

Fix
- KSampler (Pipe) & KSampler (Kargim): Implemented automatic batch-splitting support to process batched image/latent inputs sequentially. This prevents tensor dimension mismatch exceptions (Sizes of tensors must match except in dimension 1) during sampling initialization with models that use custom attention wrappers (such as Krea2 and HunyuanVideo).

Version: 4.1.11

Fix
- Save Prompt: Fixed AttributeError exception when processing None/empty text inputs passed from upstream switch nodes or empty sliced text outputs in the IO Slicer node.
- Save Prompt: Fixed base folder resolution fallback when using filename_opt directly (without a connected pipe), enabling %source_base_folder to resolve correctly for nested folder structures.

Version: 4.1.10

Feat
- Pipe Slicer: Introduced a new node IO Slicer (Slice & Dice) (RvPipe_IO_SliceDice.py) that slices all connected inputs directly using a provided indices input (supporting lists, strings, and tensors) and passes the indices downstream.
- Image Selector: Changed second output from pipe to indices (type LIST) so users can connect selection indices directly to the new IO Slicer node.
- Switch Nodes & Router Nodes: Migrated router and switch nodes (Any Multi-Switch, Any Multi-Switch Purge, Any Dual-Switch, Any Dual-Switch Purge, Any Multi-Switch Lazy, Any Multi-Switch Lazy Purge) to support list-wise execution using is_input_list=True and is_output_list=True. This prevents ComfyUI from looping/broadcasting switch nodes and resolves slot merging/duplication when mismatching list lengths are connected.
- Pass-Through Nodes: Migrated Any Passer and Any Passer Purge to use is_input_list=True and is_output_list=True to execute list transfers efficiently in a single pass.
- List Fallbacks: Updated all modified switches and passer nodes to return [None] when no inputs are connected/active, resolving type iteration merging errors in the ComfyUI executor.
Fix
- Pipe Selector: Fixed batch-slicing logic for latents, conditioning, and image lists. The selector now correctly slices batched tensors down to the actual selected indices count (e.g. 1 instead of 4) even when the list length M == 1 matches the indices length.
- Pipe Selector: Resolved conditioning tuple unpacking error in single-tuple batch slicing.
- Pipe Selector: Added automatic total batch size detection and expansion factor calculation to correctly align and repeat upstream elements (e.g. 4 prompts) with downstream generated batches (e.g. 12 images/latents) when slicing selected subsets.
- Pipe Selector: Implemented a direct pass-through bypass for image and image_list when resolved directly from the upstream pipe (where they have already been sliced), resolving double-slicing duplication issues.
- Pipe Selector & Pipe Slicer: Implemented robust recursive list/batch flattening for image and image_list inputs to resolve unpacking/concatenation exceptions when inputs are wrapped in outer lists (e.g. from folder/batch loaders).
- Image Selector: Fixed state discard behaviors during selection toggles. Deselecting all items or clicking "Discard" now performs a selection reset without deleting cached preview files on disk, avoiding image reloading/flashing on the next run.
Breaking
- Pipe Selector: Deprecated and deleted the legacy IO Selector node (RvPipe_IO_Selector.py). Users should use the new IO Slicer (Slice & Dice) node instead.

Deprecated nodes:

IO Selector

Version: 4.1.9

Feat
- Convert Primitive: Implemented automatic complete recursive flattening for list and tuple inputs.
- Convert Primitive: Implemented element-wise type conversions for list/tuple inputs when targeting scalar types (STRING, INT, FLOAT, BOOLEAN).
- Convert Primitive: Implemented boolean-fallback logic for INT and FLOAT target conversions if inputs are not directly convertible to numbers (returns 1/1.0 for truthy/non-empty values and 0/0.0 for falsy/empty values).
- Convert to List: Added string-to-list splitting behavior (prioritizing newlines over commas) so that connecting a string splits it into list items instead of wrapping it as a single-element list.
- Smart Model Loader: Changed Prompt Variations task output format to use newline-only separation (no hyphens) for easier list conversion.
Fix
- Convert Primitive: Resolved RuntimeError boolean ambiguity crash when converting multi-element PyTorch tensors to BOOLEAN by using the tensor's .any() state.
- Convert Primitive: Catch general Exceptions (including RuntimeErrors) in the execution block to provide graceful fallbacks instead of crashing the pipeline.

2026-07-10

Version: 4.1.8

Fix
- IO Selector: Resolved a bug in slice_text where single text prompt strings containing commas were incorrectly split and sliced, resulting in truncated prompts. Slicing now only applies when the input is an actual list of strings.

Version: 4.1.7

Feat
- Smart Model Loader: Added an option when deleting a loader template to also permanently delete its associated core model file (such as checkpoints, UNets, GGUFs, or Nunchaku files, excluding CLIP, VAE, or LoRA files) from local disk, prompting the user with a confirmation list of the specific files to be removed.
Fix
- Image Selector: Resolved an infinite resize and layout flickering loop when using fixed images-per-row preview modes (such as "2 Images per Row") by calculating cell dimensions based on the stable container.clientWidth instead of the dynamically-shifting grid.clientWidth, and updating ResizeObserver to observe the container while managing its lifecycle on the node.
- Image Selector: Resolved a caching issue where stale selections are passed downstream because the fingerprint signature does not update when upstream images change (despite identical size and count) by dynamically retrieving the upstream node's cached image output from outputs_cache during stack inspection and including its statistical pixel fingerprint in the md5 signature.
- Load Batch From Folder: Added a stuck-title check to prevent node titles from getting permanently locked into progress states (such as "Scanning...", "Loading 45%", or "Resizing 80%") if previous runs were interrupted or errored, dynamically falling back to clean default titles derived from node schemas.
- Vision System Prompts: Added explicit camera angle and vantage point instructions to all vision-related system prompts (Simple Description, Detailed Description, Ultra Detailed Description, Image Analysis, Detailed Analysis, Tags) so the LLM describes the actual camera perspective direction (e.g. "from a low angle looking up", "from eye level at a three-quarter angle") instead of vague shot-size labels like "medium shot".

Version: 4.1.6

Fix
- Image Selector: Replaced unstable memory pointer-based tensor signature generation with a stable, shape-and-sampled-stats-based fingerprint to prevent input-images signature mismatch loops when loaders reload on queue runs.
- Load Batch From Folder: Added python-level memory caches to the standard, Step Advanced, and legacy Advanced loaders (py/RvImage_LoadBatchFromFolder.py, py/RvImage_LoadBatchFromFolderStepAdvanced.py, py/_legacy/legacy_LoadBatchFromFolderAdvanced.py) linked to FileListCache invalidation to completely bypass repeated folder reading/decoding when queues are re-triggered after interruptions.

Version: 4.1.5

Feat (New)
- Added Load Batch From Folder (Step Advanced) featuring advanced resizing, sorting, a batch_index widget (positioned below frame_end) with control_after_generate support for auto-increment paging on queues, and a files string list output containing the source absolute file paths matching loaded frames.
- Extended the IO Selector node (py/RvPipe_IO_Selector.py) to support slicing a filenames input list, preserving base_path hierarchies, and updating active file details (filepath, filename, source_name, path, base_path) in its output pipe context for loop-based prompt saving compatibility.
- Converted Save Prompt (py/RvText_SavePrompt.py) to is_input_list=True to support internal iteration for batch/list execution, with thread-safe list-unwrapping safety on context pipes.
- Added support to Save Prompt to allow saving outside of the ComfyUI output directory when paths are anchored via source folder placeholders (e.g. %source_base_folder/../prompts), while keeping un-anchored relative paths sanitized and confined.
- Added dual-layer CPU fallback protection with persistent state tracking for the WD14 ONNX tagger model (core/sml/backend_wd14.py) to automatically handle CUDA/cuDNN mismatches/errors gracefully and prevent repeated fallback attempts in loops.
Chore
- Deprecated Load Batch From Folder (Advanced) (py/_legacy/legacy_LoadBatchFromFolderAdvanced.py) in favor of the new Load Batch From Folder (Step Advanced) node.

Deprecated nodes:

Load Batch From Folder (Advanced)

Version: 4.1.4

Feat
- Redesigned the Image Selector toolbar layout by stacking status hints above the buttons, and added a new "All" button to select all images.
- Optimized memory footprint in both standard and advanced Load Batch From Folder nodes by parsing image dimensions from headers and resizing in PIL via high-quality Lanczos downsampling before converting to PyTorch tensors.
Fix
- Image Selector: Resolved a bug preventing the "1 Image per Row" preview mode from applying correct layout styles.

Version: 4.1.3

Feat
- Added a progress bar and dynamic percentage-based node title updates to the Smart Model Loader (LM) node during batch execution and WD14 tagging, resolving display names or custom user-assigned titles rather than internal node IDs.
Chore
- Corrected factual details in 4.1.2 release notes regarding To Batch, Image Selector, and legacy deprecations.

2026-07-09

Version: 4.1.2

Feat (New)
- Added new To List and To Batch conversion nodes (py/RvConversion_ToList.py, py/RvConversion_ToBatch.py) using dynamic MatchType to convert/consolidate between lists and batches.
- Added Pipe Selector node (py/RvPipe_IO_Selector.py) to handle lockstep slicing, broadcasting, and standard context alignments. Outputs are configured as list-outputs except for image (batch tensor) and pipe (context dict).
- Added optional pipe output to Image Selector (py/RvImage_Selector.py) to output confirmed selection indices inside the pipe context.
Feat
- Refactored all switches, routers, conditionals, and passer nodes (such as Any Dual-Switch, IF A Else B, Stop, Show Any, and 14 typed passer nodes) to use ComfyUI V3 Generic Type Matching (io.MatchType) instead of legacy wildcards, enabling strict type safety and dynamic list/batch propagation.
- Added in-memory preview image caching (_stored_ui_images) to Image Selector to bypass disk writes and prevent browser reloading/flickering.
- Optimized Image Selector signature calculation using fast O(N)O(N) data pointers (data_ptr()) and shapes instead of CPU-heavy mean/std tensor operations.
- Added batch image splitting support to Tile Split (py/RvImage_TileSplit.py) to split batch frames and output tile pipes.
- Marked legacy Convert To List and Convert To Batch nodes as Deprecated, moving them to py/_legacy/ under names legacy_ConvertToList and legacy_ConvertToBatch.
Fix
- Image Selector: Removed legacy list output (routing variables through pipe), implemented strict workflow interrupt/stop, added stack-frame cache-busting when no selection is confirmed, updated toolbar warning when count is 0, and resolved nested 4D image batch extraction bugs.
- Pipe Selector: Implemented factor-division indexing mapping and singleton pass-through fallbacks to slice conditioning, text, images, and latents matching selected indices, and fixed slicing checks for both list-based and tuple-based conditioning chunks.
- Smart Model Loader (LM): Aligned node to fully support list-inputs and list-outputs natively, running the generation loop aligned to input list dimensions.
- Conditioning & Image Lists: Implemented input conditioning normalization inside KSampler and Passer nodes to wrap raw tuples, and added recursive nested list/tuple flattening to _normalize_to_list in all 9 image processing nodes to prevent direct type errors on list structures.
- Configured Show Any and Show Any Stop nodes to print raw PyTorch tensor values directly instead of calculating stats.
Chore
- Reformatted the entire codebase (py/, core/, extern/) using black for PEP 8 compliance.
- Moved legacy Convert To List and Convert To Batch nodes to py/_legacy/ folder.