EA Trim Window

When you're curating training data, "roughly this part of the video" isn't good enough. You need frame 120 through frame 175, exactly - a fixed-size window at an exact start, because your dataloader expects consistent clip lengths and your captions reference specific moments. EA Trim Window is that surgical tool: give it a start frame and a frame count, and it extracts precisely that window, reporting back the exact bounds it used.

It's the pack's "precise" trim, aimed squarely at dataset curation - the source docstring says as much - and it sits next to the simpler EA Trim Frames (head/tail removal) and EA Auto Trim (automatic detection) as the manual, deterministic option.

What you set

• start_frame - where the window begins (0-indexed).
• frame_count - how many frames the window spans. Default is 56, which is no accident: that's a common short-video training clip length, and it's the number you'll keep for many Wan 2.2 training setups.
• clamp_to_bounds (default on) - the behavior switch. On, the node quietly clamps your request to what's available (start within bounds, frame count limited by what remains). Off, it raises a ValueError instead - strict mode, for when you'd rather fail loudly than silently get a shorter clip than your dataloader expects.

The outputs

The window itself as images, plus the metadata to trust it:

• first_frame / last_frame - previews of the window's endpoints, so you can confirm the cut lands where you meant.
• frame_count - the actual number of frames extracted (may differ from your request if clamping kicked in).
• start_frame, end_frame (inclusive), and total_frames - the exact bounds used and the input's full length. Wire these to a text display and you've got a self-documenting data-prep step: no guessing what the node actually did.

Why the exact-bounds outputs matter

This is the node where the reporting is the feature. Training runs are only reproducible if every clip is accounted for, and having start_frame/end_frame/total_frames as data lets you log exactly what went into the dataset - including the cases where clamping trimmed your window shorter than requested, which is precisely the silent failure that wrecks a training run's clip-length expectations.

Install

Part of comfyui-ea-nodes: ComfyUI Manager → search comfyui-ea-nodes, or git clone https://github.com/ExoticArts/comfyui-ea-nodes into custom_nodes/, restart, hard-refresh. No extra Python dependencies. Feed it frames from EA Video Load, and save the window with EA Video Save (Idempotent) for a tidy data-prep loop.