Do I need to paint a mask, or can I work from text only?

You can do either. The VOID node supports text-guided segmentation via T5-XXL + SAM3 for quick selection, a hand-painted binary mask for precision, or both together (use text to localize and a mask to refine edges).

What if the result leaves faint shadows or reflections?

Expand the mask slightly (dilation) and add a small feather to cover halos. If using text-only selection, make the cue more specific (color, size, location) or add a manual mask where reflections persist.

What hardware and performance considerations should I expect?

The model runs in fp16 but remains GPU-intensive. Higher resolutions, longer clips, and larger temporal windows increase VRAM and render time. For faster runs, downscale the input, shorten the clip, or reduce the temporal window. Consider chunking by hard cuts; process each shot separately.

Does SaveVideo keep the original audio?

No. SaveVideo outputs a silent render. To keep audio, remux the original track in your editor or with a tool like ffmpeg after export.

VOID: Video Inpainting - ComfyUI Workflow

Q: How does this keep results stable across frames?

VOID uses a two-pass, temporally-aware inpainting process. The first pass removes the target per frame; the second propagates structure over time to reduce flicker and preserve motion cues. If you still see jitter, increase temporal consistency/window and soften mask edges.

This ComfyUI tutorial shows how to remove objects from a video using the VOID: Video Inpainting workflow. The graph is simple and fast to learn: LoadVideo reads your clip, a single VOID inpainting node (c3157b75-484a-459e-b8de-57823bef5130) performs segmentation and temporal inpainting, and SaveVideo writes the cleaned result. A MarkdownNote in the canvas provides inline guidance. Under the hood, segmentation is powered by SAM3 (using the sam3.1_multiplex_fp16.safetensors checkpoint). You can guide what to erase with a text cue encoded by a T5-XXL text encoder, a hand-drawn mask, or both for maximum precision.

Technically, the core VOID node runs a two-pass, temporally-aware inpainting model from the Netflix/void-model project. Pass one removes the masked content per frame and estimates plausible backgrounds; pass two propagates structure over time to reduce flicker and preserve motion parallax. This makes it well-suited for real object removal rather than single-frame patching: reflections, shadows, and brief occlusions are handled more coherently across frames. Because it operates in fp16, it balances quality and speed on modern GPUs while keeping memory use manageable.

FAQ

Frequently Asked Questions