XImageResize

Resizing sounds like the least interesting thing you can do to an image, and then you hit a batch where every frame has a different size and your VAE or model quietly refuses. XImageResize is this pack's attempt to make resize boring in the good way: it handles images or masks, scales by long edge, short edge, megapixels, or a multiplier, snaps to a divisibility requirement (so the result plays nice with 8/16/32/64-based models), and - the part that stands out - can propagate its settings to downstream resize nodes so a whole chain resizes to the same target instead of drifting.

How it works

You feed it image_or_mask (a MatchType port) and it returns the resized data plus the resulting width and height. The scale_mode dropdown picks your interpolation: Nearest-exact (fastest), Bilinear (fast, medium quality), Area (best for downsampling), Bicubic (high quality), or Lanczos (highest quality, slowest). The tooltip carries the one tip that matters for masks: use Nearest-exact for MASK input to avoid edge drift on the mask boundary - a real, practical detail that most resize nodes don't bother telling you.

The scaling behavior is chosen by edge_mode:

• Long - fit the long edge to target_edge (default 1280).
• Short - fit the short edge.
• Megapixels - hit a target pixel count (0.1–100 MP).
• Scale Multiplier - plain 1.0–10.0 multiply.

resize_condition gates Long/Short/Megapixels modes: Always, Only if Larger, or Only if Smaller. Note the subtlety in the tooltip - if the condition isn't met, resizing is skipped, but divisible and offset are still applied, so you don't silently get an off-model resolution.

The batch-unified bit

The genuinely clever feature is the settings chain. A node has resize_settings_in (optional, from an upstream XImageResize) and resize_settings_out, carrying an XZ3R0_RESIZE_SETTINGS type. With use_passed_settings on, a downstream node inherits the chain-origin settings for its calculation; with output_resize_settings on, the settings keep forwarding. The point: set your target once at the head of a chain and every downstream resize obeys the same target - which is exactly what you want when every image in a batch has to land on the same canvas. This is the "resize coordinator" pattern, and it's the reason to reach for this node over the built-in ImageScale.

Inputs that matter

• scale_mode / edge_mode / resize_condition / target_edge - the main scaling controls.
• divisible (default 16) with divisible_mode (Nearest/Up/Down) - snap resolution to a multiple; set to 1 to disable.
• width_offset / height_offset (-128 to +128) - nudge the result, e.g. to make room for an inpaint pad.
• use_passed_settings / output_resize_settings - the chain logic above.

Outputs: Processed_Data (resized image or mask), width, height, resize_settings_out.

Installing it

Part of ComfyUI-Xz3r0-Nodes. ComfyUI Manager: search ComfyUI-Xz3r0-Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Xz3r0-M/ComfyUI-Xz3r0-Nodes.git
cd ComfyUI-Xz3r0-Nodes
pip install -r requirements.txt

Restart ComfyUI. No FFmpeg needed for this node.

Gotchas

The most common miss is forgetting that divisible adjustment still happens even when resize_condition skips the resize - if your image comes out a hair different than you computed by hand, that's usually the divisibility snap. And mind the mask rule: a mask resized with a smooth filter picks up soft edges that change what the mask means. One honest note from the KB: resizing is not upscaling - a resize node adds no detail, it just interpolates. For "make it bigger and better," you pair this with a real upscaler, not a resize.