From 0a40784129fdecdc56651d4c5b19e62bb0c39c3c Mon Sep 17 00:00:00 2001 From: AI Lab <129358391+1038lab@users.noreply.github.com> Date: Tue, 20 Oct 2026 00:39:55 -0700 Subject: [PATCH] Update fmt.Println message from 'Hello' to 'Goodbye' --- update.md | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 61 insertions(+) diff --git a/update.md b/update.md index ce2d357..c54f212 100644 --- a/update.md +++ b/update.md @@ -1,5 +1,66 @@ # ComfyUI-RMBG Update Log +## V3.3.0 (2026/10/20) +### New Nodes +- **Canvas (RMBG)**: An interactive multi-layer image compositing studio built directly into the ComfyUI node canvas: + ![Canvas_nodes](https://github.com/user-attachments/assets/3edd69e0-ed9f-4c74-aa85-5aa84bff0ec1) + + - **Interactive Stage**: Move, rotate, flip (horizontal/vertical), adjust opacity, and scale layers with 4-corner aspect-locked transform gizmos directly on the visual stage. + - **Dynamic Layer Management**: Clean slot architecture that dynamically expands layer inputs (`layer_1` through `layer_5`) as connections are made, featuring enlarged live preview cards and hover-to-remove actions. + - **Adaptive Canvas Sizing**: Automatically syncs aspect ratio with background inputs or supports free-form width and height scaling (snapped to 8px multiples, 320px–4096px bounds). + - **Alpha Transparency & Full RGBA Pipeline**: True 4-channel RGBA compositing and output when `Background: Alpha` is selected, preserving transparent pixels for downstream preview and save nodes without black background clipping. + - **Merged Foreground Mask**: Automatically composites and outputs an exact single-channel `MASK` covering all active foreground layers for downstream inpainting and matting workflows. +- **VRMBG Video Background Removal (`AILab_VRMBG`)**: + ![VRMBG_nodes](https://github.com/user-attachments/assets/c1377e46-3533-423d-9289-0f1647314c48) + - **Flicker-Free Temporal Matting**: First dedicated video matting node designed specifically for sequential video batches and animation frames. + - **Autoregressive State Tracking**: Conditioned on prior frame foreground features to maintain temporal continuity and eliminate boundary flickering across frames. + - **Chunked Video Processing**: Features a `reset_state` toggle allowing creators to feed long videos in consecutive batches without losing temporal continuity. + - **Manual Model Loading**: Automatic download removed in favor of manual placement in `ComfyUI/models/RMBG/VRMBG-3.0/` for full offline control. +- **FeyNoBg Ultra-Fine Matting (`AILab_FeyNoBg`)**: + - **Deepened Swin-L Matting Backbone**: Introduces the 263M parameter FeyNoBg model specializing in boundary reconstruction and complex dichotomous segmentation. + - **Micro-Hair & Semi-Transparent Subjects**: Drastically improves extraction quality on wispy hair, fur, transparent glassware, lace, and motion-blurred edges with zero manual trimap required. + - **Sub-Pixel Foreground Refinement**: Built-in `refine_foreground` ensures clean, halo-free subject boundaries. +- **Finegrain Box Segmenter (`AILab_FinegrainBoxSegment`)**: + ![Finegrain Box Segmenter and Floorence-2_nodes](https://github.com/user-attachments/assets/556fb73c-1521-4b4c-bfb6-739bed5c8d90) + + - **1024px High-Definition Bounding Box Matting**: Integrates the Swin-B MVANet architecture to perform high-resolution object cutouts using bounding boxes. + - **Florence-2 & YOLO Workflow Integration**: Directly connects with `BBOX` outputs from Florence-2 and YOLO nodes for an end-to-end detection-to-matting pipeline. + - **Fallback Full-Image Mode**: Automatically falls back to high-res subject cutout when no prompt box is connected. +- **Florence-2 Segmentation Enhancement (`AILab_Florence2`)**: + - Added dedicated 4th output `BBOXES` (`BBOX` format) returning parsed bounding box data for direct plug-and-play connection to `Finegrain Box Segmenter`. +- **SAM3 Segmentation Enhancement (`AILab_SAM3Segment`)**: + - Added dedicated 4th output `BBOXES` (`BBOX` format) returning detected bounding box coordinates (`[[x1, y1, x2, y2], ...]`), enabling direct connection to `Finegrain Box Segmenter`, Florence-2 tools, and detection-to-matting workflows. + +### Model Downloads & Offline Placement +- **VRMBG-3.0**: Manually download from official [briaai/VRMBG-3.0](https://huggingface.co/briaai/VRMBG-3.0) to `ComfyUI/models/RMBG/VRMBG-3.0/`. +- **FeyNoBg**: Download from [1038lab/FeyNobg](https://huggingface.co/1038lab/FeyNobg) to `ComfyUI/models/RMBG/FeyNoBg/`. +- **Finegrain Box Segmenter**: Download from [1038lab/finegrain-box-segmenter](https://huggingface.co/1038lab/finegrain-box-segmenter) to `ComfyUI/models/RMBG/finegrain-box-segmenter/`. + +### Performance & VRAM Optimization +- **RMBG-2.0 Native FP16 Acceleration**: Enabled native FP16 (Half Precision) inference for RMBG-2.0 on CUDA devices, reducing model weights and activation VRAM consumption by >50% while leveraging NVIDIA Tensor Cores for ~2x faster throughput. +- **Video & Batch Processing Stability**: + - Implemented automatic intermediate VRAM clearing per batch chunk during video and long-sequence background removal, eliminating memory fragmentation and preventing GPU timeout (TDR) crashes. + - Enhanced numerical stability on half-precision outputs across RMBG-2.0 and BiRefNet to ensure clean edge matting without transparency artifacts. + - Hardened cleanup lifecycle across RMBG and SAM3 Multiplex nodes. + +### Compatibility & Model Architecture +- **Florence-2 Architecture Rewrite**: + - **Why We Changed It**: The original Florence-2 implementation relied on Hugging Face's dynamic remote script execution (`trust_remote_code=True`), which constantly broke or crashed whenever ComfyUI or `transformers` was updated (especially on newer 4.40+ and 5.x releases). + - **Rock-Solid Local Stability**: Rebuilt entirely on a native local PyTorch architecture. It runs 100% offline without remote scripts, eliminating dependency collisions and version-locking issues across different ComfyUI setups. + - **Complete & Accurate Masks**: Fixed the issue where complex polygon masks (e.g. car wheels, full faces) were abruptly cut off mid-generation, now producing complete and precise boundaries. + - **Removed Matplotlib Dependency**: Replaced the heavy external plotting library with native image rendering, cutting unnecessary dependencies, reducing install footprint, and preventing memory leaks. + - **Workflow Quality-of-Life**: Moved `prompt` to the top for intuitive workflow chaining, simplified parameter names, and added an `unload_model` toggle to release GPU memory on demand. +- **YOLO Segmentation (RMBG) Node Overhaul**: + - **Upgraded to YOLO26-seg & YOLO11-seg**: Integrated the state-of-the-art YOLO26 and YOLO11 instance segmentation model families (`yolo26n-seg` through `yolo26x-seg`, plus `yolo11*-seg`), hosted on our Hugging Face repository (`1038lab/YOLO`) with seamless auto-downloading. + - **Native Cutout Output (RGBA)**: Outputs direct transparent cutouts (`Alpha` RGBA) or custom background colors (`Color` RGB) right from the primary `IMAGE` output port, eliminating the need to chain downstream `Mask Applier` nodes. + - **Configurable Annotation Display**: Added `annotation_mode` (`Mask & Box`, `Box Only`, `Mask Only`, `None`) on `ANNOTATED_IMAGE` to flexibly control detection overlays without obstructing workflows. + - **Mask Refinement Suite**: Added built-in edge feathering (`mask_blur`), expansion/contraction (`mask_offset`), and mask inversion (`invert_output`). + - **Foldable Advanced Drawer**: Integrated with `Advanced_nodes.js` to present a clean, minimalist default interface while keeping fine-grained controls easily accessible. + - **Standardized Parameter Naming**: Standardized model selection widget to `model` for full consistency across all 1038lab segmentation nodes. + +### Bug Fixes & Stability +- **Image Resize Scaling Factor**: Fixed an issue in `Image Resize` where scaling via `scale_by` squared the factor instead of applying it once (e.g. 512px with 1.2x became 736px instead of 614px). ([#211](https://github.com/1038lab/ComfyUI-RMBG/issues/211)) + ## V3.2.0 (2026/09/30) ### New Features & Model Upgrades sm3