CHANGELOG.md

Changelog

All notable changes to ComfyUI-CustomNodePacks are documented here.

[1.14.1] – 2026-05-04

Fixed

• Hidden multiline-string widgets no longer leak as giant textareas on top of editor nodes. On modern ComfyUI, multiline STRING widgets are backed by a real <textarea> DOM element parented to a div.dom-widget wrapper. Setting widget.type = "hidden" was no longer enough to hide them — the DOM textarea kept rendering on top of the canvas and made PointsMaskEditor, SplineMaskEditorMEC, MECAdvancedPaintCanvas, and InpaintCompositeMEC look like a wall of empty text fields. The hide path now also sets element.style.display = "none" and collapses the dom-widget wrapper so the visible UI matches what those widgets are supposed to be: invisible internal state. Affected widgets:
  • PointsMaskEditor.editor_data
  • SplineMaskEditorMEC.spline_data / mask_color / mask_opacity
  • MECAdvancedPaintCanvas.canvas_data
  • InpaintCompositeMEC mode-conditional widgets (blend_mode_override, color_match, upscale_method, feather_edges, feather_radius)
• Verified end‐to‐end via Playwright: created PointsMaskEditor and SplineMaskEditorMEC, captured pre/post screenshots, and exercised the interactive editors. Points editor accepts left/right click for positive/negative points and Ctrl-drag for bboxes — editor_data payload contains real coordinates with distinct label:1/label:0 values (no hardcoded zeros). Spline editor accepts 4 control points and renders a smooth catmull-rom curve closed loop, not a polygon.

[1.14.0] – 2026-05-04

Fixed

• SAM Multi-Mask Picker — render real per-mask thumbnails: the JS widget previously called an empty buildMaskThumbnails() stub and then drew a flat red fillRect over all three thumbnails identically, so every candidate looked the same regardless of what SAM produced. The Python node now serializes the three candidate masks as grayscale PNGs to ComfyUI's temp directory (stable input-hash filenames) and ships them on the UI payload as mask_thumbs. The JS now loads each PNG via /view?...&type=temp into this.maskImages[i] and composites the real mask onto each thumbnail using an offscreen canvas with globalCompositeOperation="source-in" plus a per-thumb tint.
• SAM Multi-Mask Picker — widget no longer renders ~600 px below the node body: the addCustomWidget registration was forwarding node.pos[0] and node.pos[1] + widgetY into the inner draw/mouse handlers, but LiteGraph already translates the canvas context to the node origin before invoking widget callbacks — the result was a doubled offset that pushed the picker off the bottom of the node. Registration now passes widget-local coordinates (0, widgetY for draw; pos[0], pos[1] for mouse), so the picker lives inside the node body where it belongs.
• SAM Multi-Mask Picker — stop force-rerunning SAM on every queue: removed IS_CHANGED → NaN, which was unconditionally invalidating the cache and re-running SAM on identical inputs. Standard cache semantics now apply.

[1.13.1] – 2026-05-04

Fixed

• Latest-ComfyUI compatibility: removed deprecated disable_noise=False kwarg from two comfy.sample.sample_custom(…) call sites in MECBuilderSampler (mec_paint_suite.py). The argument was dropped from comfy.sample.sample_custom's signature in recent ComfyUI builds and was raising TypeError: unexpected keyword argument 'disable_noise' at sample time.
• Audited all 41 nodes against the current ComfyUI runtime: full pack imports cleanly, no other deprecated APIs in use (prepare_noise, sampler_object, calculate_sigmas, OUTPUT_NODE, IS_CHANGED all match current signatures).

Notes

• ComfyUI's V3 node API (io.ComfyNode / define_schema / MatchType / Autogrow / DynamicCombo) is opt-in and additive — the V1 API used throughout this pack is fully supported and not scheduled for removal. No mass migration required.

[1.12.0] – 2026-05-04

Added

Inpaint Composite (MEC) — unified Stitch + Paste Back

• New InpaintCompositeMEC node merges the previous two-node split (InpaintStitchProMEC + InpaintPasteBackMEC) behind a single mode dropdown:
  • stitch_pro — advanced blend pipeline (edge-aware, Laplacian pyramid, frequency, gaussian) with optional Reinhard color match.
  • paste_back — clean resize + paste with optional Gaussian-feathered edges. Faster, deterministic, no blend pipeline.
• Unified output signature (IMAGE, MASK, STRING). In paste_back mode the MASK is the paste rectangle (feathered if enabled) so downstream nodes always receive a usable mask.
• New js/inpaint_composite.js extension hides parameters that don't apply to the selected mode (clean UI without losing widget state).
• Internally delegates to the existing Stitch and PasteBack classes — no duplicated logic, no behavioural drift.

Changed

• Legacy InpaintStitchProMEC and InpaintPasteBackMEC are still registered for backward compatibility, but their display labels now read "— legacy (MEC)" to steer new workflows toward the unified node.
• README updated with a new "Inpaint Composite" section, mode-vs-mode guidance, and an updated reference table.

[1.11.0] – 2026-05-08

Added

Ronin-gimbal cinema stabilization (Inpaint Crop / Stitch / Paste-Back)

• New video_stab_strength (0–1) widget on InpaintCropProMEC activates a two-stage filter cascade on the per-frame mask centroid trajectory:
  1. EMA low-pass at 4 Hz × strength + 0.05 cutoff to kill HF jitter while preserving subject drift.
  2. Critically-damped spring (k = 8·strength + 0.5, damping = 2·√k) integrated with semi-implicit Euler for inertia and a tiny natural overshoot.
  • Pre-warm pass: the spring is run once on the first 12 frames, the final (pos, vel) state is captured, and the real pass is initialised from it — eliminating the snap-to-first-frame artefact that plagues naive spring filters.
• 99th-percentile center-displacement sizing (not full envelope): crop width = P99(detected widths) + (P99(cx) − P1(cx)) + 2·padding. Matches cinema operator framing — tight on the subject, ignoring outliers.
• New video_stab_fps (FPS for filter time-base) and video_stab_padding widgets.
• stitch_data schema bumped to v3 with new frame_offsets field (per-frame (x, y) paste positions in canvas-space). v2 remains fully supported by InpaintStitchProMEC and InpaintPasteBackMEC.
• InpaintStitchProMEC._stitch_v2 and InpaintPasteBackMEC.paste_back rewritten with a per-frame paste loop that consumes frame_offsets when present (v3) and falls back to the original single-position paste for v2 round-trips.
• New helpers in nodes/stabilization_utils.py: _median_filter_1d, _ema_lowpass, _spring_filter, ronin_filter_1d, compute_ronin_bbox_trajectory.

Changed

• InpaintCropProMEC: crop_for_inpaint signature gains video_stab_strength, video_stab_fps, video_stab_padding (all defaulted; backward compatible).
• Step 8.6 crop_mask now emits a per-frame rectangle stack when Ronin is active.
• Step 10 info string reports the active stabilization mode.

Validated

• Numerical: 11.13 px raw frame-to-frame jitter → 1.34 / 1.64 / 1.76 px at strength 0.3 / 0.7 / 1.0 (≥85% reduction); init-snap ≤ 1.2 px (warmup works).
• Round-trip: identity inpaint + paste-back is bit-perfect (mean abs diff = 0.0000) on 12-frame jittery video.
• Static-subject test: 0 px center spread on stationary subject with ±1 px detector noise.

[1.9.0] – 2026-04-29

Added

MEC Paint Suite (MEC/Paint)

• MECAdvancedPaintCanvas – interactive paint canvas (DOM widget) with Nuke-style procedural mask math: raw alpha → mask_hardness core threshold → mask_expansion morphological dilate/erode → Gaussian mask_blur_radius lerped against the hard mask by mask_blur_strength. Outputs the alpha-composited painted image plus the processed mask.
• MECContextInpainter – smart blend-back of an inpainted image. Per-channel Reinhard colour match, CIE LAB lightness rescue when the inpainted region is >5% darker, optional differential-diffusion preservation weight from |orig − inpaint|, plus [SEP] / [SKIP] / [ASC] / [DSC] wildcard parsing across connected mask regions.
• MECToneRefiner – percentile-based black/white-point tone curve (smoothstep), gray-world colour balance, optional bicubic upscale and centre-focus fake DOF driven by ai_dof_focus_depth.
• MECBuilderSampler – KSampler with adaptive CFG (Constant / Linear / Ease Down via set_model_sampler_cfg_function keyed on per-step sigma) and an optional 2-step self-correction polish pass at low denoise.
• js/mec_advanced_paint.js – RGBA canvas widget with L-paint / R-erase, linear-interpolated stamping (no dotted strokes on fast drags), brush cursor mirroring size + hardness, mouse-wheel brush sizing, base64-PNG serialisation into a hidden canvas_data widget so the Python node can decode the user's drawing.

[1.8.0] – 2026-04-29

Fixed

• FolderIncrementer (JS) – removed the isInputLoader block that was preventing filenames from LoadImage / LoadVideo / VHS_LoadVideo from being read. Loaders are now exactly the nodes whose filenames feed the versioning folder name.
• FolderIncrementer (JS) – added source_choice routing (auto/image/video) and trigger_image / trigger_video inputs. In auto mode video is preferred over image when both are connected. Re-syncs on disconnect and on source_choice change.
• Points & BBox Mask Editor – fixed canvas auto-zoom-in/out jitter on first image load and on graph execution. The double-pass fitView (rAF + 300 ms) was measuring a stale bounding rect from a pre-reflow layout, then snapping to the correct rect. Now a single deferred fit (two animation frames) runs after LiteGraph reflow has settled.
• Points & BBox Mask Editor – _hasAutoFitted is no longer reset on reference-image disconnect, so brief upstream disconnects keep the user's zoom/pan instead of crazy-refitting on reconnect.
• Spline Mask Editor – wheel-zoom is hard-clamped (80 px ↔ 8 000 px) to prevent the preview bounds from collapsing or exploding after a few scroll events. Auto-fit only re-runs when the underlying image dimensions change, not on every preview update.

Added

VAE & latent diagnostics

• VAEMergeMEC – merge two (or three) VAEs with 8 algorithms (weighted_sum, add_difference, tensor_sum, triple_sum, slerp, dare_ties, block_swap, clamp_interp). Per-block alpha overrides via JSON or comma list, brightness/contrast post-tuning on decoder.conv_out, CPU-side merge in float32, architecture detection (sd1x / sdxl / flux / mochi). Returns the merged VAE plus a JSON info report.
• VAELatentInspectorMEC – per-channel min/max/mean/std/abs_mean statistics, NaN/Inf counts, and a verdict of corrupt / saturated / low_contrast / healthy. Optional fail_on_corrupt to raise hard.
• VAESimilarityAnalyserMEC – cosine similarity between two VAEs globally and per block, with optional per-tensor breakdown.
• VAEBlockInspectorMEC – per-block weight statistics for any VAE (mean / std / abs_mean / count).

Color science (MaskEditControl/Color)

• ColorSpaceConvertMEC – sRGB ↔ Linear ↔ Rec.709 ↔ ACEScg via built-in matrices and OETFs. No PyOpenColorIO required.
• LUTApplyMEC – Adobe .cube LUT loader (1D and 3D) with trilinear interpolation in pure torch and a strength blend.
• ExposureGradeMEC – exposure (stops), white-balance (temp / tint), and contrast around a configurable mid-grey pivot. Operates in linear by default with sRGB encode/decode round-trip.

EXR I/O & metadata (MaskEditControl/IO / Metadata)

• LoadEXRMEC, SaveEXRMEC – EXR read/write with OpenEXR-first, imageio-fallback, and final 16-bit TIFF fallback. Half-float by default.
• EXRMetadataReaderMEC – pure-python EXR header parser when OpenEXR isn't installed; surfaces width/height/channels/compression.
• MetadataWriterMEC – write/merge JSON sidecars next to image batches.
• ShotMetadataNodeMEC – read shot.json descriptor (show / shot / task / frame_in / frame_out / fps).
• FrameRangeRouterMEC – slice IMAGE/MASK batches by [start:end:step] with negative-index support.
• BatchVersionManagerMEC – <root>/<show>/<shot>/<task>/v###/ layout with safe sanitization and atomic version reservation via lockfile.

Render passes & geometry (MaskEditControl/Render / Geometry)

• MergeRenderPassesMEC – composite beauty + diffuse / specular / emission / AO with independent gains.
• DepthOfFieldMaskMEC – depth → CoC mask with focus_distance and aperture controls; emits both CoC and in-focus masks.
• DepthWarpMEC – horizontal parallax warp driven by a depth pass (cheap stereo synthesis).
• NormalToCurvatureMEC – curvature mask from a tangent-space normal pass via central-difference divergence.
• PositionPassSplitterMEC – split a world-position pass into per-axis X / Y / Z masks (auto- or manually-ranged).

Plate tools (MaskEditControl/PlateTools)

• GrainMatchMEC – extract grain (reference − denoise(reference)) and re-apply it to a target plate, with seeded per-frame sampling.
• PlateStabilizerMEC – ORB+RANSAC affine stabilization (cv2) with FFT-translation fallback when cv2 isn't available.
• CleanPlateExtractorMEC – pixelwise median across a batch with optional mask exclusion of foreground samples.
• DifferenceMatteMEC – L1/L2 difference matte with threshold and softness.

Diagnostics

• TemporalConsistencyCheckerMEC – flicker / instability score across a video batch using mask_iou, pixel_diff, or Farneback flow_warp.
• ModelMetadataExtractorMEC – inspect safetensors / checkpoint files without ever unpickling. SHA256 fingerprint on size + first/last 1MB, bounded pickletools.dis for legacy pickles.

Changed

• FolderIncrementerMEC – added folder_name_override STRING and reserve_version BOOLEAN. Names are sanitized against Windows reserved names and illegal characters. Outputs use POSIX-style paths.
• SAMModelLoaderMEC – added original_device tracking and module helpers (move_to_inference_device, restore_device) for safe CPU↔GPU shuffling on offload-enabled wrappers.
• MattingNodeMEC – auto_download switch + four-step fallback chain: vitmatte_base → vitmatte_small → cv2 guided filter → torch Gaussian.
• UnifiedSegmentationNode – added force_mode (auto / image / video).

[1.7.0] – 2026-04-02

Added

• DrawShapeMEC – unified 12-shape drawing node with a single dropdown. All parameters exposed as named inputs with descriptive tooltips — no more raw JSON editing. Replaces the 5 legacy per-shape wrapper nodes.
• SplineMaskEditorMEC (JS rewrite) – complete rewrite of the spline editor following Olm SplineMask patterns:
  • Normalized [0,1] coordinates (resolution-independent)
  • Segment insertion via Ctrl+click near curve
  • Close path by clicking first point (highlighted orange)
  • Right-click context menu (Delete, Open/Close, Smooth/Sharp)
  • Zoom-relative point sizes and fonts
  • Property-based persistence with backward-compatible deserialization
  • Status bar with keyboard hints

Changed

• DrawCircleMEC, DrawRectangleMEC, DrawEllipseMEC, DrawPolygonMEC, DrawLineMEC – deprecated, now thin wrappers around DrawShapeMEC. Kept for backward compatibility with existing workflows.
• Node count updated: 38 → 47 (44 MEC + 3 FolderIncrementer).
• README, docs, and project structure updated for all new nodes.

[1.6.0] – 2026-03-19

Added

• SplineMaskEditorMEC – interactive spline mask drawing with Catmull-Rom, Bezier, and polyline modes. Outputs mask, SAM-compatible coords, and SPLINE_DATA for downstream chaining.
• MotionMaskTrackerMEC – per-frame motion detection with 4 methods (pixel diff, optical flow, background subtraction, histogram diff), camera stabilization (homography/affine/translation), and post-processing.
• BBoxSmooth – smooth bounding-box sequences across video frames using moving-average or exponential smoothing.
• stabilization_utils.py – shared camera stabilization helpers for motion tracker and inpaint suite.

Changed

• InpaintSuite – uses stabilization_utils for camera-stable cropping.
• BBoxNodes – added BBoxSmooth to the 5 existing BBox nodes.

[1.5.0] – 2025-07-17

Added

• InpaintCropProMEC – Canvas Expansion – crops near image edges now extend beyond bounds via F.pad(mode="replicate"), producing perfectly centered crops everywhere. Ported from lquesada/ComfyUI-Inpaint-CropAndStitch.
• InpaintCropProMEC – optional_context_mask – new optional input: union of context mask bbox with crop bbox to include surrounding context.
• InpaintCropProMEC – Iterative Hole-Filling – _fill_mask_holes rewritten with 14-threshold iterative approach using scipy binary_closing + binary_fill_holes (with pure-torch fallback). Preserves gradient values in soft masks.
• Stitch Data v2 Format – new coordinate system: ctc_x/y/w/h (crop-to-canvas) + cto_x/y/w/h (canvas-to-original) for perfect reversal of canvas expansion in stitching.

Changed

• InpaintStitchProMEC – v1/v2 dispatch: v2 uses canvas coordinates for seamless compositing of expanded crops; v1 fallback preserved for backward compatibility with existing workflows.
• InpaintPasteBackMEC – updated for v2 stitch format with canvas coordinate handling; v1 fallback preserved.
• Points Editor (JS) – computeSize override prevents LiteGraph relayout jitter; widget height computed from actual other-widget heights instead of magic number; locked dimensions after image load.
• Image Comparer (JS) – computeSize override prevents resize jitter after image comparison loads.

Fixed

• unified_segmentation.py – marked as DEPRECATED (dead module superseded by unified_segmentation_node.py + model_manager.py). Prevents confusion from divergent MODEL_REGISTRY.

[1.4.0] – 2025-07-16

Added

• InpaintCropProMEC – 3 new mask preprocessing inputs from original InpaintCropAndStitch: mask_invert (flip inpaint/keep regions), mask_fill_holes (flood-fill enclosed gaps), mask_hipass_filter (threshold out near-transparent noise).
• MaskBatchManager – 2 new temporal operations for video workflows: smooth_temporal (Gaussian blur along time axis to reduce flicker) and reduce_flicker (median filter across frames to remove outliers).
• BBoxSmooth (MEC) – new node to smooth bounding-box sequences across video frames using moving-average or exponential smoothing, eliminating jitter in tracked crops.

Changed

• InpaintCropProMEC – crop centering rewrite: bbox now expands symmetrically around the mask center, then clamps to image bounds. Fixes asymmetric off-center crops when the mask is near image edges.
• Points Mask Editor (JS) – complete toolbar UI overhaul: brighter button colors, 11 px font, 36 px toolbar height, textBaseline="middle" vertical centering, 120 ms active-state flash on click, wider separators, higher-opacity pills.
• Points Mask Editor (Python) – bbox rendering rewritten to use soft Gaussian-edged brushes: positive bboxes use torch.max(mask, brush) (additive, preserves soft point edges), negative bboxes use multiplicative erase. Replaces hard mask = 1.0 overwrite that destroyed soft blending.

Fixed

• utils.py – multi_scale_guided_refine and color_aware_refine now handle 2D grayscale input arrays (was IndexError when img_np.ndim == 2).
• model_manager.py – float8 dtype comparison guarded with hasattr(torch, "float8_e4m3fn") to prevent AttributeError on PyTorch < 2.1.
• Points Editor (JS) – fixed 4 memory leaks: requestAnimationFrame now cancelled in onRemoved, resize debounce timer cleared, keyboard event listener properly removed, complete cleanup chain on node deletion.
• InpaintCropProMEC – _resize_lanczos fix for single-channel masks: squeeze channel dim to 2D for PIL, restore after (was crashing with Cannot handle this data type: (1, 1, 1), |u1).

[1.3.1] – 2025-07-15

Added

• InpaintPasteBackMEC – new lightweight node to paste inpainted crop back onto the original image using stitch_data, with optional Gaussian-feathered alpha blending at the crop boundary.
• InpaintCropProMEC – downscale_method and upscale_method inputs with 5 interpolation modes: lanczos (PIL-based, highest quality), bicubic, bilinear, nearest-exact, area. Upscale method is stored in stitch_data and automatically used by InpaintStitchProMEC.

Changed

• InpaintCropProMEC – padding_multiple now enforces step=2, min=2 to guarantee even-valued padding (required by many tiled/diffusion models).
• Image Comparer (JS) – mode labels now use Unicode icons (◧ Compare, ⊕ Overlay, ≠ Diff); labels fade during drag; divider grip uses a dot-grid pattern; overlay scrubber upgraded to rounded bar + circular handle.

Fixed

• InpaintStitchProMEC – now reads upscale_method from stitch_data (defaulting to lanczos) instead of always using bilinear.

[1.3.0] – 2025-07-14

Added

• ImageComparerMEC – new interactive before/after comparison node with three view modes: drag-slider, adjustable-opacity overlay, and amplified difference heatmap. Rendered entirely in-node via a DOM canvas widget.
• MaskDrawFrame – 7 new shapes: triangle, star, diamond, cross, rounded_rectangle, heart, arrow (total: 12 shapes).
• MaskDrawFrame – rotation parameter (−360 ° to +360 °) applies to every shape including the original five.
• SAMMaskGeneratorMEC – negative_text_prompt input: describe what to exclude; GroundingDINO detects the region and injects negative points into SAM inference.
• InpaintCropProMEC – downscale_factor (0.25–1.0) to shrink the crop before sending to the inpainting model and automatically upscale on stitch.
• InpaintCropProMEC – mask_blur / mask_grow pre-processing: blur and morphologically dilate/erode the mask before computing the crop region.
• InpaintCropProMEC – aspect_ratio selector with 12 presets (1:1, 4:3, 16:9, …) plus custom ratio; crop region is adjusted to match the chosen AR.
• InpaintCropProMEC – 6th output crop_mask: binary mask showing the cropped region in original image space.

Fixed

• Points/BBox Editor (JS) – eliminated canvas jitter caused by updateEditorSize() re-entrantly calling node.setSize() in a ResizeObserver loop. Added 50 ms debounce, integer-snap for pixel positions, and a re-entrance guard.
• Points/BBox Editor (JS) – toolbar separators and pill buttons now render at integer pixel positions, removing sub-pixel aliasing artefacts.

Changed

• Model Manager – .safetensors format is now preferred over .pt/.pth/ .bin when a safetensors variant exists locally. SAM3 (sam3.pt) is explicitly excluded from this mapping.
• InpaintCropProMEC – crop dimensions are now snapped up to the nearest padding_multiple to avoid size mismatches with tiled models.

[1.2.6] – 2025-07-13

Fixed

• GitHub Actions publish workflow: added FORCE_JAVASCRIPT_ACTIONS_TO_NODE24 env var and fork guard.
• Added missing MIT-License file required by PyPI metadata.

Changed

• README overview table expanded from 2 to 4 packs.