Netflix VOID, part 2: a Mac Studio expedition

2026-05-24 · Source: Machine Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Content Creation & Production · Depth: Advanced, short

Summary

An analysis of the Netflix VOID video inpainting model's performance on Apple Silicon, specifically an M4 Max with 128GB unified memory, demonstrates its viability for professional VFX pipelines. The standalone VOID implementation, after specific patches for device detection and output, achieved production-comparable quality in float32, processing 75-frame shots in approximately 18 minutes. This outperformed a community ComfyUI node, which had architectural shortcuts and slower processing. While VOID Pass 1 proved effective, Pass 2 was impractical on Mac due to OOM errors, hardcoded A100 timeouts, and no significant quality improvement without Netflix's proprietary Star upscaler. A surprising finding was that direct Log3G10 input yielded cleaner generations than Rec.709, suggesting better dynamic range preservation. However, VOID is limited to small to medium masks, failing on large masks (e.g., 60% of frame) due to architectural reliance on spatial context.

Key takeaway

For VFX artists or AI engineers evaluating video inpainting solutions on Apple Silicon, Netflix VOID Pass 1 offers production-comparable quality after specific patches. You should prioritize the standalone implementation over community ComfyUI nodes, which are less capable. Be aware that VOID is a precision tool best suited for small to medium masks and requires manual integration into your color pipeline. Consider experimenting with direct Log3G10 input for potentially superior results, challenging the conventional Rec.709 conversion advice.

Key insights

Netflix VOID Pass 1 achieves production-comparable quality on Apple Silicon with specific patches, but has architectural limitations.

Principles

• Community reimplementations may lack core features.
• Log footage can preserve dynamic range for AI models.
• Diffusion inpainting needs sufficient spatial context.

Method

The standalone Netflix VOID workflow on Apple Silicon requires patching device autodetect from CUDA, switching output to 16-bit PNG via OpenCV, and adding un-padding for the temporal VAE.

In practice

• Patch VOID for MPS float32 and 16-bit PNG output.
• Test Log3G10 input for better dynamic range.
• Use VOID for small to medium inpainting masks.

Topics

• Netflix VOID
• Video Inpainting
• Apple Silicon
• VFX Pipeline
• Color Management
• Log3G10

Best for: AI Engineer, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• 🍎Video Object Deletion🍎 👉Void by Netflix is a novel video object removal framework... — Void by Netflix offers physically-plausible video object removal for complex scenarios.
• Built an open-source one-prompt-to-cinematic-reel pipeline on a single GPU — FLUX.2 [klein] for character keyframes, Wan2.2-I2V for animation, vision critic with auto-retry, music + 9-language narration in the same pipeline — A single-GPU, open-source pipeline generates cinematic video reels from one prompt, integrating multiple AI models.
• Netflix Acquires Ben Affleck’s AI Startup — Is This the End of Hollywood, or a New Beginning? — AI in film production is enabling efficiency and creative freedom, not eliminating jobs.
• LiveEdit: Towards Real-Time Diffusion-Based Streaming Video Editing — LiveEdit enables real-time, stable streaming video editing via a three-stage distillation pipeline and AR-oriented mask cache.
• TextureFlow part II: full ComfyUI walkthrough - powerful AI animation tool — TextureFlow is a ComfyUI workflow for abstract animations, integrating IP adapters, ControlNets, and AnimeDiff for detailed control.
• Toward More Controllable AI Video Editing: An Early Research Exploration at Netflix — Netflix's Vera and VOID models advance controllable AI video editing by isolating changes and ensuring physical plausibility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning on Medium.