VRGDG LTX Preview XYZ Plot

LoRA training spits out preview videos at checkpoints, and if you're training in chunks - which the VRGameDevGirl trainer nodes make you do - you end up with a folder full of small preview clips named by step. Squinting at them one at a time is the wrong way to judge a run. This node waits until training actually finishes and then stitches every saved preview into one side-by-side grid video, the way people eyeball checkpoint progress at a glance. It's the "did it converge" node.

You can think of it as the training loop's final evaluation step. It doesn't sample anything itself - it's a video-assembly node that turns already-written preview files into a single comparison reel. The "XYZ" in the name is aspirational: it's a flat grid of all your previews, not a true XYZ sweep.

How it works

The node is triggered by a vhs_filenames input from VideoHelperSuite's Combine Video node, so it runs right after the trainer's preview videos are written. It filters the preview_folder for videos matching output_name (your LoRA/output name), and only builds the grid once completed_steps reaches total_target_steps. Before that, it does nothing - which is by design: you don't want an interleaved half-finished grid.

cell_width and cell_height size each tile; 0 auto-detects from the first preview video. If you're comparing a lot of checkpoints, 320px tiles keep the grid from exploding; 512 is a nicer balance for a handful. label_tiles adds a label band above each tile using the filename (which should include the LoRA name and step - that's how you read the grid). output_fps defaults to 24.

For the export, render_backend picks between CPU (libx264) and NVIDIA GPU (h264_nvenc). The honest tradeoff, per the tooltip: the grid compose filters still run in ffmpeg either way, so NVENC only speeds up the encode stage - it's not a miracle on a CPU-bound compose. Nice to have if your ffmpeg has NVENC.

Outputs and wiring

Three outputs, and you mostly care about one: xyz_video_path is the STRING path to the finished comparison video. created is a BOOLEAN that flips true when the grid was actually built, and status carries a human-readable message. Wire created into a gate or just watch the console. Feed the resulting video to a Preview/Save node and you've got a record of the whole run in one file.

Install and gotchas

Same pack install as everything else - ComfyUI Manager, search "vrgamedev", or git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl into ComfyUI/custom_nodes, restart, and install kornia, librosa, imageio from the README's requirements. The one dependency to plan around is VideoHelperSuite for the trigger, plus ffmpeg for the encoding (the pack falls back to its bundled imageio-ffmpeg binary if you don't have system ffmpeg).

The main gotcha is the trigger discipline: vhs_filenames must come from the same combine step that writes your previews, or the node never fires. And remember - it only builds the video when completed_steps hits the target, so if you abort a run early there's no grid, just silence. If you want an early look, run with a lower total_target_steps to force a build.