Video masking tools for ComfyUI: process a masked subject at high resolution inside a moving crop, then paste the result back into the original frames without jitter or visible seams.

Naive per-frame crops around a moving subject jitter in position and size, and video models read that jitter as camera motion. Subject Crop produces batches that are stable by construction: the crop holds still through mask noise, occlusions, and brief excursions, and follows only sustained motion.

example workflow

Mask Cleanup removes segmentation noise, Subject Crop cuts a stable batch around the subject, sampling runs on the crops (optionally masked via Mask To Latent Space into Set Latent Noise Mask), and Subject Uncrop pastes the result back into the full frames.

Examples

Subject Crop / Uncrop

Synthetic stress cases at default settings. Green is the tracked crop, red is zoomed.

| | | |---|---| | moving_approach | recede | | Subject grows while crossing the frame: zoomed follows position and size smoothly, tracked shows the constant-size answer. | Subject shrinks and is still moving at the last frame: the clip ends on a steady, fully padded crop. | | occlusion_dip | oscillate | | The face-mask case, where half the mask vanishes for a stretch: the crops don't react. | Side-to-side swings: absorbed by a slightly roomier crop that stays put instead of chasing every reversal. | | hard_cut | offscreen_entry | | Scene cut: one clean jump to the new framing, no hunting before or after. | Subject enters from off-screen: the clipped mask distorts the apparent size, but the crop settles without lurching. |

Mask To Latent Space

Replacing a car with a German shepherd in an LTX video: feeding the pixel mask straight to Set Latent Noise Mask leaves ComfyUI to trilinear-resize it, which blurs the mask across frames and lets the car bleed through. The max-reduced mask from Mask To Latent Space inpaints cleanly.

mask to latent comparison

Installation

Clone into custom_nodes and restart ComfyUI. Needs ComfyUI v0.15.0 or newer, with no dependencies beyond ComfyUI's own.

cd ComfyUI/custom_nodes
git clone https://github.com/drozbay/MaskVidExperiments

Nodes

Summaries only. Field-level usage lives in each node's tooltips.

MVEx Subject Crop

Crops a region around the masked subject from every frame, sized so the whole batch stacks into one tensor. Position, size, and shape are chosen together over the whole clip, so the crop holds still through mask jitter and follows only sustained motion.

combined: one static crop with padding around the subject's whole travel.
tracked: a constant-size crop that stays still until the subject would leave it, then moves as little as possible. Pixel-exact slices, no resampling.
zoomed: the crop also follows the subject's size, resampled to one fixed output resolution in which the subject keeps a constant share.

crop_scale sets the padding around the subject as a ratio of its size, padding and prefer set how firmly that padding is kept and whether stillness or tightness pays for it, and the debug output summarizes what the planner chose. upscale_megapixels upscales every crop to about that many pixels on the divisible_by grid, keeping the shape the planner chose and never downscaling, so no resize node is needed on either side. Setting crop_scale to 0 skips cropping entirely: the whole frame passes through, still sized to the divisible_by grid and the upscale_megapixels floor, so the rest of the workflow works unchanged with tracking off. MVEx Subject Crop (Advanced) exposes every internal dial, with the standard node's padding and prefer settings as presets over them.

MVEx Subject Uncrop

Pastes processed crops back into the original frames with a feathered border, optionally confined by the cropped masks. The paste is pixel-exact when the crop size was not changed, and crops that were resampled along the way (zoomed mode, upscale_megapixels, or a resize node in between) are scaled back to their box automatically. The input order is bypass-correct: bypassing both Subject Crop and Subject Uncrop passes the processed frames straight through, so the whole crop pipeline can be A/B'd with two clicks.

MVEx Mask Cleanup

Removes specks and brief flickering blobs from a video mask batch while keeping the real subject, including its soft edges. The shrink_grow method erases anything too thin to survive a shrink and restores the survivors' exact shapes, and the components method drops blobs that are both small and short-lived.

MVEx Frame Range Mask

Builds a batch of solid masks at a given resolution, fully masked on the frames you list and empty on the rest. Useful for masking a stretch of a clip for inpainting, or for feeding a temporal mask straight into Mask To Latent Space.

MVEx Mask To Latent Space

Reduces a pixel-space mask batch to latent resolution using the VAE's spatial and temporal compression, aligned to how causal video VAEs group frames instead of the trilinear resize ComfyUI applies on its own (see the example above). Feed the result to Set Latent Noise Mask. auto reads the geometry from the connected VAE, including the exact frame cycle and 2x2 token grid of chunked models like MiniMax H3, and manual enters it directly. The mask is unified over each model token, so what you preview with MVEx Latent Mask To Mask is exactly what the model acts on.

MVEx Audio Mask To Latent

The audio counterpart: regenerates the chosen time ranges of an audio latent and keeps the rest, on a joint AV latent (MiniMax H3, LTX-2) or a bare audio latent. Timing and layout come from the connected audio VAE or from manual widgets, ranges from time_ranges text, a timeline mask, or start/end times. The node attaches the noise mask itself, since Set Latent Noise Mask cannot reach the audio side of a joint latent, and merges with any mask already there so copies chain. MVEx Audio Mask Debug reports which time ranges a latent's audio mask keeps and regenerates.

MVEx Differential Diffusion (Soft)

Drop-in variant of ComfyUI's Differential Diffusion. The stock node turns a grayscale denoise mask into hard binary masks that grow as sampling proceeds, so every intermediate mask has a razor-sharp edge no matter how feathered the input was. This version thresholds with a soft ramp instead, so each per-step mask keeps a soft edge whose width follows the blur already present in the input mask: heavily feathered masks stay feathered at every step, sharp masks stay sharp. Purely per-pixel with no spatial blur, so it is temporally stable on video, and softness 0 exactly reproduces the stock node.

Acknowledgements

The batchcrop nodes in ComfyUI-KJNodes (which trace back to comfy_mtb) are the prior art that motivated this design. Subject Crop and Subject Uncrop are an independent reimplementation with no code carried over, but they exist because those nodes proved the workflow worth doing.
ComfyUI-Inpaint-CropAndStitch is the mature crop-and-stitch solution for still-image inpainting. Use it for single images; this pack covers the video case, where applying that workflow per frame produces exactly the jitter described above.
Mask To Latent Space generalizes WanMaskToLatentSpace from my ComfyUI-WanVaceAdvanced.
Most of the code in this pack was written with Claude (Fable 5).