A ComfyUI custom node toolbox focused on multi-image references, SAM3/SAM3.1 person cutouts, SCAIL-2 mask workflows, image batch utilities, boolean routing, and group-based workflow switching.

Current version: v2.9.0. The latest version of ManualRefCollage is v2.2.2.

Installation

ComfyUI Manager

Search for ComfyUI-FeiHou-Toolbox in ComfyUI Manager, install it, then restart ComfyUI.

Manual Install

Open your ComfyUI custom_nodes directory and run:

git clone https://github.com/FX-FeiHou/ComfyUI-FeiHou-Toolbox.git

After installing or updating, restart ComfyUI and hard-refresh the browser page.

Nodes Overview

Node display titles follow the current ComfyUI language setting. This English README shows the English titles.

| Display Title | Node ID | Version | | --- | --- | --- | | Create SCAIL-2 Colored Mask V2 | SCAIL2ColoredMaskV2 | v2.6 | | Auto Ref Collage | AutoRefCollage | v1.0 | | Manual Ref Collage | ManualRefCollage | v2.2.2 | | Switch V2 | ComfySwitchNodeV2 | v2.0 | | Invert Boolean | InvertBoolean | v2.3 | | Image Batch Multi V2 | ImageBatchMultiV2 | v2.4 | | Fast Groups Bypass Switch | FastGroupsBypassSwitch | v2.0 | | Random Seed Noise | RandomSeedNoise | v2.8 | | Video Combine 🎥🅥🅗🅢 V2 | VideoCombineV2 | v2.9 |

Create SCAIL-2 Colored Mask V2

An enhanced version of ComfyUI's original SCAIL-2 colored mask node, built for mask routing, separation, and batch handling in multi-reference workflows.

prefix_mask_mode supports:

Multi Image Single Color: each output batch image is single-color, matching the current reference_image_mask foreground color.
Multi Image Multi Color: each output batch image uses the node's object color palette rule.
Single Image Multi Color: output batch images use batch-order colors: blue, red, green, magenta, cyan, then loop after 5 images.

render_device supports:

gpu: renders masks on the graphics card for speed, using VRAM.
cpu: offloads mask rendering to system memory to reduce VRAM use, at a slower speed.

Auto Ref Collage

Automatically uses SAM3/SAM3.1 to cut people out from multiple reference images and compose them into a single multi-person reference collage.

Manual Ref Collage (v2.2.2)

A manual collage node that loads SAM3/SAM3.1 cutouts onto an editable canvas, allowing users to adjust position and scale before outputting the final composed image.

Switch V2

A revised switch node for toggling between two workflow paths while reducing issues caused by inactive branches.

Invert Boolean

An input-only boolean node that inverts true to false and false to true. It is compatible with PrimitiveBoolean output chains.

Image Batch Multi V2

A copy-style variant of KJNodes ImageBatchMulti for creating a batch from multiple image inputs. It keeps the dynamic input-count workflow, but skips missing or bypassed image inputs instead of adding black placeholder frames. If no image input is available, it returns no image output.

Fast Groups Bypass Switch (v2.0)

A group bypass and switch node that binds two ComfyUI Groups, switches which group is active, bypasses the inactive group, and outputs the corresponding input.

Changelog

v2.9.0

Added Video Combine 🎥🅥🅗🅢 V2, based on VideoHelperSuite's Video Combine interface, format presets, VAE/LATENT handling, date-prefix handling, and preview behavior.
When audio is connected, Video Combine V2 writes one final video file containing the video stream, audio stream, and ComfyUI prompt/workflow metadata.
Restored the complete VideoHelperSuite format list and per-format dynamic widgets. The preview auto-plays muted and enables audio while hovered.

v2.8

Added RandomSeedNoise, a rgthree-style random seed controller that outputs standard NOISE for SamplerCustomAdvanced.

v2.6

Changed the render device to gpu and cpu options. gpu is the default; if a long video runs out of VRAM, manually switch to cpu.

v2.5

Optimized SAM3/SAM3.1 cutout memory usage in AutoRefCollage and ManualRefCollage.
Wrapped CLIP prompt encoding and SAM3 inference in torch.no_grad() to avoid retaining inference graphs during preview loading and workflow execution.
Reduced the matched-object refinement peak by keeping only the best mask/box before refinement and avoiding a full original-image GPU copy before cropping.

v2.4

Added ImageBatchMultiV2 (Image Batch Multi V2 / 图像组合批次（多重）V2), based on KJNodes ImageBatchMulti.
Kept the original dynamic inputcount and Update inputs workflow.
Changed missing-input handling: unconnected or bypassed image inputs are skipped instead of being replaced by black frames.
If no image inputs are available, the node returns no image output.

v2.3

Added InvertBoolean (Invert Boolean / 反转布尔值), a single-input single-output boolean inverter.
The node is compatible with PrimitiveBoolean output chains and flips true to false, false to true.

v2.2.2

Changed ManualRefCollage size handling so the manual canvas reads and applies current width / height inputs only when 应用尺寸 is clicked.
ManualRefCollage no longer auto-applies linked width/height during node creation, connection changes, or 载入图片, avoiding early reads before KJNodes Set / Get variable chains are ready.
Added compatibility for KJNodes-style Set_variable / Get_variable width and height routing; clicking 应用尺寸 now attempts to resolve the matching Set node and read the value from its input chain.

v2.2.1

Fixed ManualRefCollage not correctly reading externally connected width / height inputs in both the frontend manual canvas and backend execution.
Manual collage now prioritizes linked width/height inputs and falls back to the node's own width/height widgets only when linked values cannot be resolved.

v2.2

Updated Create SCAIL-2 Colored Mask V2 prefix_mask_mode behavior.
Multi Image Single Color renders every output batch image as a single color matching the current reference_image_mask foreground color.
Added Multi Image Multi Color for rendering every output batch image with the node's object color palette rule.
Single Image Multi Color renders output batch images with batch-order colors: blue, red, green, magenta, cyan, then loops after 5 images.

v2.1

Fixed ManualRefCollage preview loading images from ignored/bypassed inputs.
Fixed ignored/bypassed image inputs still being included during manual collage execution.
Fixed externally connected width / height being overridden by stale layout_json dimensions.

v2.0

Added FastGroupsBypassSwitch (多框忽略并切换).
Supports enabling, bypassing, and output switching between two ComfyUI Groups.
UI follows the rgthree Fast Groups Bypasser style for large grouped workflows.

v1.5

Finalized ManualRefCollage (多参图像手动拼接).
Supports SAM3/SAM3.1 cutouts and manual layout for up to 5 reference images.
Supports using the first video frame as a composition reference while keeping the final output on a black or white background.

v1.0

Initial release.
Added Create SCAIL-2 Colored Mask V2.
Added AutoRefCollage.