New Open-Source Models Now in ComfyUI: VOID, BiRefNet & Gemma 4

Exciting Enhancements for Our Open-Source Ecosystem

Our open-source network has witnessed significant advancements! We're delighted to introduce ComfyUI compatibility with three new open-source iterations this week:

• VOID from Netflix for object elimination in videos

• BiRefNet for intricate background separation

• Gemma 4, Google's multimodal reasoning framework

Discover their functionalities and implementation below.

VOID: Advanced Scene Editing with Physical Interaction Awareness

Netflix recently unveiled VOID (Video Object and Interaction Deletion), enhancing traditional pixel deletion in video restoration. When eliminating subjects, VOID concurrently erases all associated physical traces—including shadows, reflexes, and displaced items triggered by the subject's presence.

Successful restoration hinges on quadmask: replacing binary masks with grayscale equivalents using four values. This guides the model to discern removable zones, overlapping sections, physically influenced regions, and preserved areas—enabling causal inference beyond basic gap completion.
Tip: Leverage segmentation frameworks like SAM3 to initialize quadmask inputs.

VOID features two operational modes:

• Pass 1 handles core restoration for standard clips

• Pass 2 incorporates optical-flow refinement for improved temporal continuity in complex footage

Implementation Guide

1. Obtain the package and position files in designated model directories

2. Acquire workflows from templates

3. Supply input footage, quadmask, and post-removal scene description

4. Execute the workflow
   Download Workflow

BiRefNet: Precision Segmentation for Complex Visuals

BiRefNet (Bilateral Reference Network) developed for CAAI AIR 2024 specializes in high-fidelity mask extraction from intricate imagery. It captures delicate elements like hair, fur, and transparent surfaces across diverse segmentation applications.

This versatile architecture concurrently addresses salient/obscured object detection within a compact framework.

Implementation Guide

1. Retrieve the model and place birefnet.safetensors in /models/background_removal/

2. Obtain workflows via templates

3. Process images instantly
   Download Workflow

Gemma 4: Multimodal Reasoning Architecture

Google DeepMind introduced Gemma 4, the latest evolution of their open-weight framework. This inherently multimodal system processes text, visuals, audio, and video while offering configurable stepwise reasoning before generating responses.

Available variants:

• E2B/E4B: Parameter-efficient options suitable for consumer GPUs

• 26B A4B (MoE) and 31B dense configurations available through Google

Implementation Guide

1. Download the package and transfer files to /models/text_encoders/

2. Acquire workflows from templates

3. Integrate textual instructions with optional media inputs
   Download Workflow

Happy innovating!