Geeky RemB v4.0 - Ultimate AI Background Removal & Video Layering

If your only goal is to cut a person out of a photo, you don't need this node - ComfyUI ships BiRefNet natively these days, and a native Image Remove Background plus a composite node does the same job with fewer surprises. What GeekyRemB is actually for is the all-in-one move: feed it a foreground and a background, and it cuts the subject out, scales, rotates, repositions and blends it onto the new scene, even animates it across a frame sequence, all in a single pass. That's a workflow that normally takes four or five nodes. It's a kitchen sink, and like most kitchen sinks it does the core job well and lets the accessories get ahead of the marketing.

The honest read first

The README is a masterpiece of overstatement. "Multimodal processing", "physics-aware generation", "self-optimizing workflows" - none of that is in the shipped code. What's actually there is one node (GeekyRemB, ~950 lines) that wraps real models: BiRefNet for cutting, rembg's RMBG-1.4 as an alternative, and an OpenCV chroma keyer. Same story with the 15+ blend modes: in the source, only multiply, screen and overlay do their own math; the rest fall through to a plain alpha blend. Call it a comfortable wrapper, not a revolution, and you'll set your expectations right. The node itself is MIT, and the models underneath (BiRefNet, RMBG, rembg) are all open with no commercial gotchas.

How it works

Pick removal_method. The default rmbg_v2_birefnet (and birefnet_zhengpeng7) loads BiRefNet through HuggingFace transformers from the cocktailpeanut/rm repo, resizes your image to 1024 square internally, and upscales the mask back - which means the model caps how fine those hair edges can get no matter what processing_resolution you set. rmbg_v1_4_briaai uses rembg's ONNX session. chroma_key_advanced is pure HSV keying with spill suppression, and hybrid_ai_chroma blends a BiRefNet mask 70/30 with a chroma mask.

Then the compositor takes over: 12 position_mode choices, scale_x/scale_y, rotation, flips, opacity, and the animation_type/easing_function pair for motion over a frame batch. "Video" here means it loops over the batch dimension of your IMAGE tensor - pipe in frames from VHS Load Video and you get a batch of composited frames back. It's frame-consistent, not fancy.

The inputs that actually matter

• foreground and background - both required IMAGE inputs. Nothing happens without both.
• removal_method - the quality switch. BiRefNet for hard edges, RMBG-1.4 when you want speed, chroma key when you shot on a green screen.
• skip_background_removal - set this when your foreground already has an alpha channel (e.g. from LayerDiffusion) and you only want compositing.
• mask_blur / edge_feathering - your edge-quality levers; feather 2–5 for hair, leave the rest alone.
• blend_mode / opacity - for compositing into the scene rather than just pasting.

Outputs are composite_result (the merged image), foreground_mask, and foreground_removed_bg (the cutout with alpha) - wire composite_result to Save Image and you're done.

Installing it

ComfyUI Manager, search "GeekyRemB", install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/GeekyGhost/ComfyUI-GeekyRemB.git
cd ComfyUI-GeekyRemB
pip install -r requirements.txt

Two install gotchas. First, the requirements file has onnxruntime-gpu>=1.16.0 listed with a comment saying "uncomment if using GPU" - but it's not commented, so a CPU-only machine will try to install the CUDA build and can fail at import. Swap it for plain onnxruntime if you have no GPU. Second, the first run downloads models: BiRefNet into your HuggingFace cache and the RMBG-1.4 ONNX into ~/.u2net. Give it a minute and let it finish - don't cancel.

Where people get burned

The big one is silent failure. If anything inside throws - model download fails, CUDA issue, weird input - the node catches it and returns your unprocessed background as the composite, so you get a result that looks fine and is completely wrong. Keep the ComfyUI console open on first run; the errors land there. Also note that because the mask output is built as an RGB image tensor, treating foreground_mask as a true single-channel MASK can misbehave downstream - prefer the foreground_removed_bg RGBA output for further compositing. And remember: 1024 is the BiRefNet ceiling here, so for 4K source you're still getting 1024-model edges. For a genuinely overengineered single node, it works surprisingly well for the cut-and-place workflow - just don't buy the "ultimate" part.