ComfyUI-RMBG 3.3: Canvas Studio, Video Background Removal and New Matting Models

ComfyUI-RMBG by 1038lab is one of the most popular background removal packs for ComfyUI, and version 3.3.0, released on October 20, 2026, is its biggest update in months. It adds a Canvas node for layering cut-outs on new backgrounds right inside ComfyUI, a video background removal node built on Bria’s VRMBG-3.0 that keeps edges steady from frame to frame, two new matting models for hair, fur and glass, and box-based segmentation that works with Florence-2, SAM3 and YOLO. RMBG-2.0 also now runs in native FP16, which the changelog says halves its VRAM use and doubles its speed. This guide covers how to install or update the pack, every model you need with its exact folder path, four practical workflows (photo, hair and fur, single object, video), the Canvas node, licences and fixes for the common errors.

What Is New in ComfyUI-RMBG 3.3

The release also fixes the Image Resize node squaring its scale_by factor (512 px at 1.2x now gives 614 px, as it should), a Florence-2 crash on Transformers 5.x, and a bug where one failed optional import could stop the whole pack from loading.

Install or Update the Pack

Pick one of these three methods.

1. ComfyUI Manager: open Manager, search for ComfyUI-RMBG, install (or click Update if you have an older version) and restart.

2. Comfy CLI:

comfy node install ComfyUI-RMBG

3. Git clone:

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

Whichever method you use, install the requirements with the Python that runs ComfyUI. For a normal install or venv, from ComfyUI/custom_nodes:

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

For the Windows portable build, run this from the ComfyUI_windows_portable folder:

python_embeded/python.exe -m pip install --no-user --no-cache-dir -r ComfyUI/custom_nodes/ComfyUI-RMBG/requirements.txt

If you already have the pack, update with git pull inside ComfyUI/custom_nodes/ComfyUI-RMBG and run the pip command again, because 3.3 changed some dependencies (matplotlib is gone, and the YOLO nodes are new). The pack’s example workflows are in example_workflows on GitHub.

Models and Exact Folder Paths

Most models download automatically to ComfyUI/models/RMBG/ the first time you pick them in a node. If your PC is offline, or a download fails, put the files in the folders below yourself. Folder names are case sensitive on Linux, so copy them exactly.

To download a model by hand with the Hugging Face CLI, run it from the folder that contains ComfyUI, for example:

hf download 1038lab/RMBG-2.0 --local-dir ComfyUI/models/RMBG/RMBG-2.0
hf download 1038lab/FeyNobg --local-dir ComfyUI/models/RMBG/FeyNoBg
hf download 1038lab/finegrain-box-segmenter --local-dir ComfyUI/models/RMBG/finegrain-box-segmenter

Note that the Hugging Face repo is spelled FeyNobg but the folder ComfyUI-RMBG expects is FeyNoBg.

Which Model Should You Use?

Workflow 1: Remove a Background From a Photo

1. Add Load Image and pick your photo.
2. Add the RMBG node and set the model to RMBG-2.0 (or BiRefNet for commercial work). Connect the image.
3. Leave sensitivity at 0.5 and process_res at 1024 for a first run. Set background to Alpha for a transparent PNG, or pick a colour.
4. Connect the image output to Save Image, and the mask output to a Preview Image through a Mask to Image node if you want to check the mask.
5. Queue. The first run downloads the model to ComfyUI/models/RMBG/RMBG-2.0/.

If edges look hard, add 1 to 3 of mask_blur. If a thin halo of the old background remains, set mask_offset to -1 or -2 to shrink the mask slightly. Turn on Refine Foreground to clean colour spill on edges.

Workflow 2: Hair, Fur and Glass With FeyNoBg

Hair and fur are where most background removers fail, leaving either chunky clumps or a grey fringe. FeyNoBg is a dedicated matting model (263M parameters, Swin-L backbone) that predicts soft alpha values, so single strands and semi-transparent edges survive.

1. Load your portrait or pet photo with Load Image.
2. Add FeyNoBg (AILab_FeyNoBg). Keep refine_foreground on, since it removes the colour of the old background from semi-transparent pixels.
3. Output to Save Image with an alpha background, or straight into the Canvas node (below) to place the subject on a new background.

No trimap is needed, which is what makes FeyNoBg handy compared with older matting models. The model file is 1.05 GB, so it is heavier than RMBG-2.0. Use it for the hard images and RMBG-2.0 for everything else.

Workflow 3: Cut Out One Object With a Box

Sometimes you do not want everything in the foreground, just the shoe, the bottle or one person. 3.3 makes this a two-node job:

1. Load the image and add Florence-2 (AILab_Florence2) with a grounding prompt such as “the white sneaker”, or the YOLO node to detect common objects. Florence-2 now runs fully offline.
2. Connect the new BBOXES output to Finegrain Box Segmenter (AILab_FinegrainBoxSegment) along with the image.
3. The segmenter returns a 1024 px matte of only that object.

SAM3’s AILab_SAM3Segment node has the same BBOXES output, so you can use a SAM3 text prompt as the detector instead. If nothing is connected to the box input, the Finegrain node simply cuts out the whole image, so it also works as a general background remover.

Workflow 4: Video Background Removal With VRMBG

Running an image model on every frame of a video causes flicker, because each frame’s mask is a little different. Bria’s VRMBG-3.0 is built for video: it looks at the previous frame and its matte, which keeps the edges steady over time. It works at 1024×1024 internally and scales the matte back to your video size.

Get the model (it is gated):

1. Open briaai/VRMBG-3.0 while logged in to Hugging Face and request access. Bria grants academic access on request; anything else needs a commercial agreement with Bria.
2. Once approved, log in on your PC with hf auth login and download the files into the folder ComfyUI-RMBG expects:

   hf download briaai/VRMBG-3.0 --local-dir ComfyUI/models/RMBG/VRMBG-3.0

3. Check that ComfyUI/models/RMBG/VRMBG-3.0/ contains model.safetensors (885 MB), config.json, vrmbg3_config.py and model.py. The node never downloads this model for you.

The workflow:

1. Load your clip with the core Load Video and Get Video Components nodes (or Video Helper Suite’s Load Video), which give you the frames as an image batch.
2. Connect the frames to VRMBG (AILab_VRMBG).
3. For a long video, feed it in chunks, and leave reset_state off between chunks of the same shot so the model keeps its memory. Turn it on when a new shot starts.
4. Save the result with Video Combine from ComfyUI-VideoHelperSuite (installed in ComfyUI/custom_nodes/ComfyUI-VideoHelperSuite/) or the core Create Video and Save Video nodes. Use a green background if your format has no alpha channel, or export an image sequence as PNG to keep transparency.

3.3 also clears VRAM after each chunk, which fixes the memory fragmentation and GPU timeout crashes that long videos used to cause. If you need commercial video matting, run BiRefNet on each frame instead and smooth the mask with a little mask_blur.

The Canvas (RMBG) Compositing Node

Canvas is the most visible addition in 3.3. Instead of round-tripping to Photoshop to put a cut-out on a new background, you can do it on the node:

• Connect a background image and up to five layers. The layer_1 to layer_5 inputs appear as you connect them.
• Drag layers on the node’s preview to move them, and use the controls to rotate, flip, scale and set the opacity of each one.
• The canvas size follows the background, or you set it between 320 and 4096 pixels (it snaps to multiples of 8).
• Set Background to Alpha to get a full RGBA result with no background, plus a merged foreground MASK.

A typical product shot workflow is: Load Image, RMBG (alpha), Canvas with a studio backdrop as the background and the cut-out as layer 1, then Save Image. The merged mask output is handy for a final relight or shadow pass in another workflow, for example with an image editing model like the MOSS-Image Edit LoRA.

RMBG Node Settings Explained

Licences: What You Can Use Commercially

The ComfyUI-RMBG code is GPL-3.0, but each model has its own licence, and this matters more than people think:

• VRMBG-3.0 is non-commercial, gated and “not open source at the moment”. Commercial use needs an agreement with Bria.
• RMBG-2.0, also from Bria, is released for non-commercial use as well, with commercial licences sold by Bria.
• BiRefNet and InSPyReNet are MIT licensed upstream, which makes them the usual choice for client and commercial work.
• FeyNoBg and the Finegrain Box Segmenter mirrors on 1038lab’s Hugging Face have no model card yet, so check the original authors’ licences before commercial use.

Troubleshooting

• All RMBG nodes are missing after the update: run the pip command again with ComfyUI’s own Python, then restart. 3.3 also stops one failed optional import from hiding the whole pack, so check the console for the specific missing package.
• “VRMBG-3.0 model not found”: it is never auto-downloaded. Request access, download it and check the files are directly inside ComfyUI/models/RMBG/VRMBG-3.0/, not in a nested subfolder.
• FeyNoBg cannot find its model: the folder must be ComfyUI/models/RMBG/FeyNoBg/ with that exact capitalisation, even though the Hugging Face repo is called FeyNobg.
• Florence-2 error about forced_bos_token_id: fixed in 3.3 for Transformers 5.x. Update the pack.
• Video flickers: use VRMBG instead of an image model, and do not turn on reset_state in the middle of a shot.
• GPU timeout or out of memory on long videos: update to 3.3 (VRAM is cleared per chunk), feed the video in smaller chunks, and lower process_res if needed.
• Image Resize gives a bigger result than expected: the old version squared scale_by. Update to 3.3.

FAQ

What is ComfyUI-RMBG?

It is a custom node pack by 1038lab for background removal, matting and segmentation in ComfyUI. It wraps many models, including RMBG-2.0, BiRefNet, BEN2, InSPyReNet, SAM, SAM2, SAM3, GroundingDINO, Florence-2 and now VRMBG, FeyNoBg and YOLO.

Where do ComfyUI-RMBG models go?

Most go in ComfyUI/models/RMBG/<model name>/, for example ComfyUI/models/RMBG/RMBG-2.0/. SAM goes in ComfyUI/models/SAM/, SAM2 in ComfyUI/models/sam2/ and GroundingDINO in ComfyUI/models/grounding-dino/.

Can ComfyUI-RMBG remove the background from a video?

Yes. Version 3.3 adds the VRMBG node with Bria’s VRMBG-3.0, which reduces flicker by using the previous frame. The model is gated and non-commercial, and must be placed manually in ComfyUI/models/RMBG/VRMBG-3.0/.

Which model is best for hair?

FeyNoBg, new in 3.3, is designed for hair, fur, glass and motion blur. BiRefNet matting variants are a good commercial-friendly alternative.

Can I use the results commercially?

It depends on the model. RMBG-2.0 and VRMBG-3.0 are non-commercial without a Bria licence. BiRefNet and InSPyReNet are MIT licensed upstream.

How much VRAM does it need?

Background removal is light compared with image generation. RMBG-2.0 now runs in FP16 with less than half the VRAM it used before, and most models run comfortably on 8 GB at the default 1024 processing resolution. Video and very high process_res values need more.

Final Thoughts

ComfyUI-RMBG 3.3 turns a background removal pack into a small compositing studio. The Canvas node saves a trip to an image editor, FeyNoBg handles the hair and fur that used to need manual cleanup, box matting makes single-object cut-outs easy, and VRMBG finally gives ComfyUI a steady video matting option, as long as you respect its licence. Update the pack, put the new models in the right ComfyUI/models/RMBG/ folders, and start with RMBG-2.0 or BiRefNet before reaching for the specialist models.

More local AI image guides:

• Local Background Removal: Best Practices with U2NETP on CPU
• MOSS-Image Edit LoRA in ComfyUI
• Bria Fibo 2.0 in ComfyUI
• FLUX 3 Image Klein in ComfyUI

Sources

• GitHub: 1038lab/ComfyUI-RMBG (README)
• GitHub: ComfyUI-RMBG changelog (v3.3.0, v3.2.0)
• Hugging Face: briaai/VRMBG-3.0
• Hugging Face: 1038lab/FeyNobg
• Hugging Face: 1038lab/finegrain-box-segmenter
• ComfyUI Wiki: ComfyUI-RMBG v3.3.0 news