COMFYUI-SAVE-IMAGE-XMP

Stack

Python + ComfyUI node API

Two custom nodes — Save Layered TIFF (XMP) and Save Image (XMP) — drop in as a regular ComfyUI extension.

libtiff via Pillow

Multi-page TIFF write with per-page (per-layer) IFD tags. Each page carries PageName (TIFF tag 285) as its layer name.

Deflate / ZIP + delta predictor

Lossless compression for every layer; ~30–50% smaller than uncompressed with full bit-depth preservation.

XMP packet

Custom cfl:* namespace embedded in the file's XMP block: workflow JSON, prompt JSON, model hashes, author, layer names, arbitrary user JSON.

PNG tEXt chunks

PNG outputs also get native workflow / prompt tEXt entries so ComfyUI re-loads them like any first-party save.

Per-session model-hash cache

Model SHA256 collection is automatic from the run graph and cached so a 20-image batch doesn't re-hash a 7 GB checkpoint twenty times.

Process

The problem. ComfyUI's default save behavior is one image, one file, with the workflow embedded as a PNG tEXt chunk. That's fine until your graph produces a meaningful intermediate (a mask, an outpainted region, a depth pass, a heatmap, a preview) — then you're flooding the output folder with files whose relationship lives only in the filename convention you happened to use that day. Six months later the connection is gone.

Why TIFF. TIFF is the only widely-supported lossless format that natively handles multiple named layers in a single file — the same reason it has been the standard archival format in print, broadcast, and medical imaging for decades. Each page sits in its own Image File Directory (IFD) with its own compression tag, so a preview layer can be JPEG at quality 90 while mask layers stay lossless Deflate and a float32 heatmap goes uncompressed — all in one file. The current nodes use Deflate + delta predictor uniformly for simplicity.

The metadata block. A custom XMP namespace cfl: carries the full ComfyUI node graph JSON, the execution prompt graph, SHA256 hashes for every model used in the run, an arbitrary user-supplied JSON blob, an author string, and (for TIFF) a comma-separated layer name list. Model hashes are collected automatically from the run graph and cached per-session so batched runs don't pay the hashing cost more than once.

Why this is worth the work. macOS Finder, QuickLook, and Preview.app render multi-page TIFFs natively — spacebar a file in Finder and you see the preview layer; arrow through to see masks and depth passes. Drop the TIFF into Photoshop or Affinity Photo and every page appears as a named layer in the Layers panel without a plugin. The XMP block survives all of these. Six months later the file is still self-describing.