💾 Save Batch Images (BC/NM/POS)

What it's actually for

A texture-projection workflow produces three related passes per camera view: base colour, normal map, and position. ComfyUI's built-in SaveImage has no idea they're related - you'd be chaining three save nodes, juggling filename_prefix to keep view1_bc and view1_nm distinguishable, and hoping nothing collides when the batch is six views deep.

This node takes all three as separate optional inputs on one node, writes them with a consistent naming scheme, and hands the whole thing back as one batched image for a preview widget. It's the boring plumbing that keeps a multi-view pipeline sane.

How it works

Feed it bc, nm, and pos batches and it loops each one, converts each tensor to a PIL image, and writes:

{build}/output/view_bc_1.png
{build}/output/view_bc_2.png
{build}/output/view_nm_1.png
...

So filename plus the channel plus the index - deterministic, greppable, and matching what the pack's Load_batch_images and View2DSaveLoad expect to find. Saving uses PNG with compress_level=1, i.e. fast rather than small; that's a deliberate choice for iterative work, and it means these files are chunkier than you'd expect. There's no metadata embedding here - no workflow chunks, no A1111 parameters - so these are working files, not something you'd upload to Civitai and expect to reconstruct from. Keep your graph saved separately.

Two behaviours worth knowing:

• All three inputs are optional. If all are empty, it returns a 16x16 black placeholder tensor instead of erroring.
• The 16x16 gate. Before writing anything, the node inspects the first available channel and refuses to write if the image is 16px or smaller on either side. That's a deliberate guard for the pack's own pipeline, where the batch-listing node emits tiny placeholder images to keep a graph's structure valid. Useful in-context, mildly surprising if you ever legitimately want to save a 16x16 texture - you can't.

The output is preview: every image written, stacked into one batch, so you can drop a Preview Image node on it and see the lot.

The inputs that matter

• path (default L:/AI_3d/output) - a raw filesystem path, not ComfyUI's output folder. It's the author's drive; change it or your files land nowhere useful. The node creates the directory if it doesn't exist.
• filename (default view) - the base name. Keep it short; it prefixes every file in the batch.
• bc / nm / pos - the three image batches. Wire whichever you have; each is written independently, so a graph with only base colour still works.

One output: preview (IMAGE), the concatenated batch for a viewer.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Antonioilev/ComfyUI_Antonioilev_Lightpack.git
# restart ComfyUI

Manager search: "Antonioilev Light Pack". The repo ships no requirements.txt and __init__.py catches import errors with only a printed node count, so a missing package shows as a missing node. This one is torch, numpy and Pillow - all present in any ComfyUI install.

Where people get burned

• The path is absolute by default and won't exist on your machine. Set it to a folder you own. Windows-style L:/... paths work if you're on Windows; on Linux, use a real mount.
• IS_CHANGED returns the current time, so this node re-runs and re-writes every queue, even when nothing upstream changed. That's intentional for a saver, but it means a graph left running overnight fills a folder.
• Files are large. compress_level=1 trades size for speed - if you're archiving, re-compress afterwards or swap in a standard save node for the final pass.
• No alpha. Wire a batch that has an alpha channel and it's flattened into RGB on save. That's what the pack's alpha variant is for.