DIGIT Drift Gate (Paths)

DIGIT Drift Gate (Paths) is the file-based twin of the Drift Gate node. Where the original takes reference and generated as IMAGE tensors from your graph, this one takes two absolute file paths - reference_path and generated_path - and loads the images itself. Same comparison math, same outputs, zero tensor plumbing.

Why does that variant exist? Because in a production pipeline the reference plates live on disk, not in a graph. Think render farms, automated review loops, or a script that drops a fresh render into a folder and expects the gate to judge it against a locked plate. If your workflow is already file-based - render lands in shots/dev_0010/, reference sits next to it - this node skips the "load it, wire it, convert it" dance entirely.

How it works

Under the hood it's the exact same engine as the tensor version: resize-align, pixel SSIM + edge SSIM scoring with detail_weight, grid-based hotspot detection, and a pass/fail decision against confidence_threshold. All the same tuning knobs are here - resize_mode, fit_method, compare_width/compare_height, grid_size, diff_threshold, reject_mode, and the save_qc/qc_output_dir/qc_filename trio for writing the burn-in QC sheet and JSON sidecar.

The one difference in setup: the two paths are required inputs, and qc_filename (when you save QC) auto-derives a base name from generated_path if you leave it empty - output is always <name>_qc.png/.json. That's a nice touch for keeping QC artifacts tied to the file they judged.

Outputs are identical to the tensor gate: passed, confidence, drift_report, qc_json, qc_filepath, gated_image, qc_sheet, diff_heatmap, edge_diff, reference_aligned, and generated_aligned. So you can build the same rejection logic and the same interactive review (wired into DIGIT Drift QC Preview) whether the image arrived as a tensor or a path.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/thedepartmentofexternalservices/comfyui-digit.git
cd comfyui-digit
pip install -r requirements.txt

Or ComfyUI Manager → search comfyui-digit → install → restart. Local-only, no API keys.

Which version to pick

Honestly, it comes down to where your images are. If they're already outputs in a ComfyUI graph (a freshly generated render right off a node), use the tensor gate. If they're files - a reference plate on disk, or renders being dropped into a watch folder - the paths version removes a whole step. The path-based gate is also the one you'd script against: you can drive it purely by changing which files it points at, no graph rewiring. And because it produces the same QC artifacts, your downstream record-keeping doesn't care which one produced them.