TiledWan Inpaint Crop Improved

If you've ever tried to inpaint a moving object across a video one frame at a time, you know the specific misery this node exists to fix: the mask slides around, so every frame crops a different region, and the result strobes like a broken neon sign. TiledWan Inpaint Crop Improved is the crop half of a two-node inpainting pair that solves it by thinking in video, not single images.

It's one of six nodes in Baverne's comfyUI-TiledWan pack, which exists to push Wan 2.1 VACE video inpainting past its native 832×480 / 81-frame wall (the TiledWan Video VACE Pipeline article covers that half). The crop node is a rework of lquesada's ComfyUI-Inpaint-CropAndStitch - the pack that made "crop, inpaint, stitch" the standard way to guarantee unmasked pixels stay untouched. Where the original assumed stills, this one was rebuilt for batches: per the pack README, it ignores size variation (which caused inconsistencies) and handles masks that appear and disappear over time.

How it works: you feed it a batch of frames plus a mask. It finds the tight box around the masked area, expands it into a context region, then runs a preprocessing chain on the mask - fill holes, expand by N pixels, optional invert, a feathered blur, and a high-pass threshold that kills weak or noisy mask values. Optionally it extends the canvas for outpainting and preresizes the input to a sane range. Finally it resizes the crop to your inpainting model's preferred resolution, pads it to a multiple you choose, and packages everything into a stitcher record so the matching TiledWan Inpaint Stitch Improved can put it back exactly.

The one setting that makes it video-aware is keep_window_size. Instead of letting each frame pick its own crop box, the node takes the largest window across the batch, forces every frame to use it, and linearly interpolates the box's position frame to frame. That's the anti-flicker trick, and it's why you should have it on whenever you're feeding actual footage.

The few inputs you'll actually touch:

• keep_window_size - on for video, off for stills
• context_from_mask_extend_factor (default 1.2) - how much surrounding context to keep. More helps the model understand the scene; too much just wastes generation budget
• output_target_width / output_target_height (default 512×512) - set these to your model's native resolution; 832×480 if you're feeding Wan VACE
• output_padding - keep it a multiple of 8 for VAE, 32 for many diffusion models

It also takes an optional mask (if you leave it empty, the whole image gets processed) and an optional_context_mask if you want to force areas into the crop without inpainting them.

Outputs: stitcher (STITCHER), cropped_image (IMAGE), and cropped_mask (MASK). Wire the image and mask into your inpainting stage, then feed the result plus stitcher into TiledWan Inpaint Stitch Improved. Depending on the build you have, the node may also expose a row of DEBUG outputs showing each intermediate stage - they're the author's debugging aids, and you can ignore them.

Install is the pack install: ComfyUI Manager (search "comfyUI-TiledWan"), or cd ComfyUI/custom_nodes && git clone https://github.com/Baverne/comfyUI-TiledWan, then restart. The pack declares no extra pip dependencies.

Two gotchas worth knowing. First, there's a guard: if you feed a batch without keep_window_size, output_resize_to_target_size must be on, because every frame in a batch has to come out the same size. Second, an unedited mask straight from Load Image sometimes arrives at the wrong shape; the node detects and fixes that, but a manually painted mask is more trustworthy - give it a clean pass before you commit to a long render.