1038lab

切り抜いた素材を、ComfyUI内で配置したい人へ。 RMBG v3.3.0のCanvasは、複数素材の移動・拡大縮小・回転に対応。 切り抜きだけなら導入を急がず、合成までまとめたい時の候補に。 ※公式更新の整理。未実測／添付は説明図。 github.com/1038lab/ComfyUI-RMBG/blob/main/update.… t.co

[Update] ComfyUI-RMBG v3.3.0: Interactive Canvas studio node, VRMBG temporal video matting, FeyNoBg hair matting, YOLO26-seg, Florence-2 local offline rewrite, and native FP16 Hi everyone, We just tagged v3.3.0 for ComfyUI-RMBG. Over the past few weeks, between investigating compatibility crashes on newer Transformers releases and trying to streamline multi-node matting and compositing setups, we ended up overhauling large parts of the repository. 1038lab/ComfyUI-RMBG Rather than staggering these into separate small patches, we put everything into this release. It covers fixes for broken upstream dependencies, new dedicated video and fine-detail matting backbones, direct detection-to-matting plumbing, and an interactive compositing node on the graph. Here is a breakdown of what was added, what changed, and why it is useful: 1. Canvas (RMBG): Interactive Multi-Layer Compositing Studio Chaining multiple blend nodes, transform nodes, and mask combiners just to place cutouts over a background gets tedious quickly. We built an interactive staging node directly on the ComfyUI canvas: - Interactive Stage Handles: Move, rotate, scale (aspect-locked via 4 corners), flip horizontally or vertically, and tune opacity directly on the visual stage. - Dynamic Layer Expansion: Connect layer_1 through layer_5 and the node dynamically expands its inputs, with live thumbnail cards and hover-to-remove actions. - True Alpha Transparency: When set to Alpha background, it outputs a clean 4-channel RGBA image without downstream nodes clipping transparent regions to solid black. - Merged Foreground Mask: Automatically merges all active layer silhouettes into a single combined MASK output, ready for immediate downstream inpainting, relighting, or shadow generation. Workflow： github.com/1038lab/ComfyUI-RMBG/blob/main/example… 2. VRMBG: Flicker-Free Video Background Removal (AILab_VRMBG) Standard image matting nodes applied frame-by-frame on video sequences often produce noticeable edge jitter and strobing hair boundaries. - We integrated BRIA AI's VRMBG-3.0 architecture, which uses autoregressive temporal state tracking conditioned on prior frame foreground features to keep edges coherent across motion. - Added a chunked batch reset_state toggle, letting you process long video sequences in smaller batches without memory fragmentation or GPU driver timeouts (TDR). - Fully offline: Place models manually in ComfyUI/models/RMBG/VRMBG-3.0/. Workflow： github.com/1038lab/ComfyUI-RMBG/blob/main/example… 3. FeyNoBg: Ultra-Fine Matting for Hair, Glass, and Fur (AILab_FeyNoBg) For challenging subjects where general-purpose segmenters struggle, we added FeyNoBg (based on a deepened 263M Swin-L backbone): - Tailored specifically for complex dichotomous boundaries: wispy hair, animal fur, transparent glassware, lace, and motion-blurred edges without requiring manual trimaps. - Features a built-in sub-pixel refinement toggle (refine_foreground) to eliminate residual halos and color fringing on fine borders. 4. Finegrain Box Segmenter: 1024px High-Definition Bounding Box Matting (AILab_FinegrainBoxSegment) - Integrates the Swin-B MVANet architecture to perform high-resolution 1024px object cutouts driven by bounding box coordinates. - Plugs directly into the new BBOX outputs from Florence-2 and YOLO for automated detection-to-matting workflows. - Automatically falls back to full-image subject matting if no bounding box input is provided. Workflow：github.com/1038lab/ComfyUI-RMBG/blob/main/example… 5. Florence-2: Native Local Rewrite & BBOX Output (AILab_Florence2) - Native Local PyTorch Architecture: The previous implementation relied on Hugging Face remote code (trust_remote_code=True), which broke on recent transformers updates with AttributeError: forced_bos_token_id. We completely rewrote Florence-2 into a clean local PyTorch module. It runs 100 percent offline, eliminates remote script vulnerabilities, and avoids version locking. - Complete Polygon Masks: Fixed the issue where complex polygon boundaries (such as car wheels and facial contours) were abruptly truncated mid-generation. - Matplotlib Removed: Replaced external plotting dependencies with native tensor and image rendering, reducing installation footprint and preventing memory leaks. - New BBOXES Output: Added a dedicated 4th output port (BBOXES) providing parsed coordinates for direct connection to Finegrain Box Segmenter. - UI Polish: Placed prompt at the top for easier workflow chaining, simplified widget names, and added an unload_model toggle. 6. YOLO Segmentation Node Overhaul (AILab_Yolo) - YOLO26-seg and YOLO11-seg Families: Upgraded to state-of-the-art YOLO26 and YOLO11 instance segmentation models (yolo26n-seg through yolo26x-seg, plus yolo11*-seg), with auto-downloading from 1038lab/YOLO. - Native RGBA Cutouts: The primary IMAGE output now directly emits transparent cutouts (Alpha RGBA) or solid color backdrops (Color RGB), eliminating downstream Mask Applier helper nodes. - Configurable Annotation Display: The ANNOTATED_IMAGE port supports flexible display modes (Mask & Box, Box Only, Mask Only, or None). - Built-in Mask Refinement: Added edge feathering (mask_blur), boundary expansion/contraction (mask_offset), and mask inversion (invert_output). - Collapsible Advanced Drawer: Foldable parameter drawer keeps the default node footprint minimal while keeping advanced thresholds accessible. Workflow: github.com/1038lab/ComfyUI-RMBG/blob/main/example… 7. SAM3 Segmentation Enhancement (AILab_SAM3Segment) - Added a dedicated 4th output BBOXES (BBOX format) that outputs detected bounding box coordinates ([[x1, y1, x2, y2], ...]), allowing SAM3 detections to feed directly into the Finegrain Box Segmenter or downstream cropping tools. 8. Native FP16 Inference & VRAM Optimization - RMBG-2.0 Native FP16 on CUDA: Over 50 percent reduction in weights and activation VRAM, providing roughly 2x faster throughput using NVIDIA Tensor Cores. - Batch Processing Stability: Implemented automatic intermediate VRAM clearing per batch chunk during video and long sequence background removal to prevent GPU driver timeouts (TDR). - Numerical stability improvements across half-precision RMBG-2.0 and BiRefNet outputs to eliminate transparency artifacts. 9. General Bug Fixes and Stability - Fixed Image Resize scaling factor calculation where scale_by was accidentally applied twice (#211). - Hardened node loader exception handling so a single missing optional library will not crash the entire node pack (#217, #218, #220). - Updated timm.models.layers imports to timm.layers across backbones to silence upstream deprecation warnings (#210). How to Update & Documentation You can update directly via ComfyUI Manager, or run git pull inside your custom_nodes/ComfyUI-RMBG directory. - GitHub Repository: 1038lab/ComfyUI-RMBG - Complete Changelog & Details: github.com/1038lab/ComfyUI-RMBG/blob/main/update.… - Example workflows for the Canvas node, VRMBG video matting, and Florence-2 to Finegrain box workflows are included in the repository's example_workflows/ folder. If you encounter any edge cases with specific video formats or layer setups, feel free to open an issue on GitHub. Hope this saves you some time in your daily workflows.

ComfyUI-EdgeTTS v2.0.0: fixed the sudden 403 error, and Whisper STT now exports synced .srt subtitles 1038lab/ComfyUI-EdgeTTS A few days ago, like many of you, I hit Queue Prompt on my video workflow, sat back waiting for the voiceover to generate, and was immediately greeted by that lovely red error box: 403 Invalid response status. A Turns out Microsoft decided to roll out an unannounced security handshake overhaul across their speech endpoints, effectively bricking Edge TTS worldwide. If you have ever had a complex 50-node video pipeline grind to a screeching halt because the audio generator suddenly refused to talk to the server, you know that very specific feeling of developer despair. So we bumped the project to v2.0.0 and completely overhauled the pipeline. Here is what is in this update, why we made it a major version, and what it actually means for your daily workflows: The Main Event: Your TTS Is Back From the Dead First and foremost: the speech synthesis works again. Instead of asking everyone to dive into their terminal, fight with conflicting pip packages, or break their embedded Python environments trying to update external libraries, we rebuilt the connection engine directly under the hood. We handled the new handshake requirements internally, added an automatic retry backoff so random network hiccups won't nuke your batch queue, and built multi-backend stream fallbacks so audio saving never throws buffer errors on newer PyTorch versions. You do not need to perform virtual environment surgery—just update the node, and your voices will work exactly as they did before. The New Superpower: Native SRT Subtitle Export on Whisper While we had the hood open fixing the connection, we decided to tackle another persistent pain point for creators: subtitles. Let us be real: if you are generating speech inside ComfyUI, nine times out of ten you are pairing it with video. And if you are making videos, raw text from speech recognition is only half the battle. You still had to drag the audio into another application or write helper scripts just to figure out the timestamps. In v2.0.0, Whisper STT now has a dedicated SRT output socket alongside TEXT. When you feed your audio into Whisper, it does not just transcribe what was said; it hands you a ready-to-use, millisecond-accurate .srt subtitle block (00:00:01,000 --> 00:00:03,500) directly from the model segments. You can immediately pass these timecoded subtitles into video overlay nodes, lip-sync tools, or save them right alongside your render. No external dependencies, no extra steps. Demo Workflow: github.com/1038lab/ComfyUI-EdgeTTS/blob/main/exam… What You Need to Do If your workflows were stalled by the 403 error, just update the node: - Via ComfyUI Manager: hit Update on ComfyUI-EdgeTTS. - Via git: run git pull inside your custom_nodes/ComfyUI-EdgeTTS folder. 1038lab/ComfyUI-EdgeTTS There is also a complete end-to-end example workflow sitting in example_workflows/ that connects Edge TTS directly into Whisper STT so you can verify both audio synthesis and subtitle generation in one click. Let me know if your voices are speaking again, and if there are any other quality-of-life additions you want to see in the audio stack next!