ComfyUI-RMBG v3.3.0: Canvas Studio and Video Matting

The Canvas studio node, from the repository's example workflows.

A compositing stage inside the node graph

Canvas (RMBG) is the headline addition. It is a multi-layer compositing stage rendered on the node canvas itself, so cutouts can be arranged without leaving ComfyUI.

• Layers accept up to five inputs, layer_1 through layer_5, with the slots expanding as connections are made.
• Each layer can be moved, rotated, flipped horizontally or vertically, scaled with four-corner aspect-locked gizmos, and given its own opacity.
• Canvas size follows the background input's aspect ratio, or can be set freely between 320px and 4096px in multiples of 8.
• Selecting Background: Alpha keeps a true four-channel RGBA output, so transparent pixels survive into preview and save nodes instead of collapsing to black.
• The node also emits a merged single-channel MASK covering every active foreground layer, which feeds straight into inpainting or further matting.

Video matting arrives with VRMBG

The VRMBG node and its example workflow, from the v3.3.0 changelog.

AILab_VRMBG is the pack's first node built for sequences rather than single frames. It runs Bria's VRMBG-3.0 and conditions each frame on foreground features from the previous one, which keeps boundaries stable instead of flickering frame to frame. A reset_state toggle lets long videos be fed in consecutive batches without losing that continuity.

Automatic downloading is off for this model. Weights go into ComfyUI/models/RMBG/VRMBG-3.0/ by hand, and the briaai/VRMBG-3.0 repository is gated, so access has to be granted on Hugging Face before the files can be pulled at all.

Two finer matting backbones

The Finegrain Box Segmenter fed by Florence-2 boxes, from the changelog.

FeyNoBg targets the hard cases: wispy hair, fur, glass, lace, and motion-blurred edges. It uses a 263M parameter Swin-L matting backbone with a built-in refine_foreground pass, and the changelog notes it needs no manual trimap. The weights ship as a single 1.05 GB safetensors in 1038lab/FeyNobg.

Finegrain Box Segmenter takes a bounding box and returns a 1024px cutout using a Swin-B MVANet architecture, so boxes coming from Florence-2 or YOLO become high-resolution mattes without a second detection model. With no box connected it falls back to segmenting the whole image.

To feed it, both the Florence-2 and SAM3 segmentation nodes now expose a fourth BBOXES output in BBOX format, which plugs directly into the box segmenter.

Florence-2 moves off remote code

The Florence-2 node previously loaded Hugging Face's dynamic remote script with trust_remote_code=True, which broke whenever ComfyUI or transformers moved. v3.3.0 rebuilds it as a native local PyTorch implementation: no remote script execution, no matplotlib dependency, and no version pinning. Two reported defects go with it, polygon masks that were cut off mid-generation and the 'Florence2LanguageConfig' object has no attribute 'forced_bos_token_id' crash under transformers 5.x.

YOLO segmentation overhaul

The rebuilt YOLO segmentation node, from the changelog.

The YOLO segmentation node moves to YOLO26-seg and YOLO11-seg, with the yolo26n-seg through yolo26x-seg and yolo11*-seg checkpoints hosted in 1038lab/YOLO and downloaded on demand. Its IMAGE output can now return a transparent RGBA cutout or a solid background colour directly, removing the Mask Applier step that used to follow it. Annotation overlays are switchable between mask and box, box only, mask only, or none, and mask_blur, mask_offset, and invert_output handle edge refinement inline.

Performance and fixes

• RMBG-2.0 runs in native FP16 on CUDA. The changelog claims more than 50% less weight and activation VRAM and roughly 2x throughput from Tensor Cores.
• Video batches clear VRAM per chunk, which removes the memory fragmentation behind GPU timeout crashes on long sequences.
• Image Resize no longer squares the scale_by factor, so 512px at 1.2x gives 614px instead of 736px (#211).
• Node discovery survives a failed optional import, fixing a NameError that could take down the entire pack (#218).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-RMBG

Install dependencies with the ComfyUI Python environment, then restart:

python -m pip install -r ComfyUI-RMBG/requirements.txt

Model placement for the three new backbones:

Repository: 1038lab/ComfyUI-RMBG
Changelog: update.md v3.3.0
Example workflows: example_workflows