SG Load

If your show tracks work in Flow Production Tracking - the thing everyone still calls ShotGrid - the irritating part of using ComfyUI at work isn't the model. It's the handoff. Someone publishes a depth pass, a matte, a plate, and you go find it, download it, drop it in a folder, and hope you grabbed the right revision. SG Load deletes that step: it queries your site for the Version you describe, decodes its media, and hands you the frames as a normal IMAGE.

Set expectations: this is half of a two-node pack, it's in alpha, and it's aimed at studio pipelines rather than r/comfyui. The entire Reddit corpus holds five threads that mention "ShotGrid". No crowd wisdom to lean on - but unusually thorough docs.

The thing it's actually good at

The rule-based picking is the point. You don't wire up a version number, you wire up a question: the newest Version on this Shot, in these statuses, whose name contains "depth". Step 5 of a pipeline publishes, step 6 consumes it, and no id gets copied between graphs - which is exactly the failure mode that hand-typed ids create.

Where this sits next to the ecosystem's usual habit: ComfyUI's convention is that metadata rides inside the file, since a PNG carries the graph that made it (image-io-metadata.md). A studio is the opposite shape - the record on the site, the pixels on shared storage.

How it decides what to load

The node resolves against the site at execution time using the sg-groundtruth client. Each output takes the best source it can find on its own: image prefers a sequence on storage, then a decoded clip, then a still, then a thumbnail; video prefers a movie PublishedFile, then the movie on storage, then the untouched uploaded mp4. source in the advanced fold is the override for when two candidates both look plausible - pick one file and both outputs read from it.

Decoding goes through ComfyUI's own decoder, the same call core Load Image makes, not Pillow - which is why a 16-bit PNG keeps its levels and a 32-bit float EXR keeps values above 1. Feeding a comp rather than a sampler, that's the difference between usable and not (post-processing.md).

The inputs you'll touch

project and link are the two required combos - and link being empty means "search the whole project", so blank is a real strategy.

Then, of the rest:

• statuses - comma-separated codes, empty accepts any. There's no "approved" concept in Flow Production Tracking - approved is one code among many, and the tooltip appends the codes your project uses.
• name_contains - words that must all appear in the Version name. This is how you say "depth v0" without a filter.
• frame / frame_count (advanced) - frame is the number in the filename, so 1003 means plate.1003.exr, and 0 means "wherever this sequence starts", which is why a 1001–1048 plate needs nothing typed. frame_count 0 reads to the end, 1 gives you a single image, and a batch too big for memory is refused with the number that fits.

The other advanced ones - newest_by, filters, pin_version_id - are for when the show's naming isn't tidy. pin_version_id beats everything above it.

What comes out

image is the batch. video is a real VIDEO, wiring into anything downstream that eats clips, and only fetched when that slot is wired. mask is the decoder's alpha as 1 - alpha, matching core Load Image's convention, with a zero mask when the source has none. Then the record: version_id, code, and colour_space, which is recorded, never applied - nothing converts, nothing guesses sRGB for you.

The load is recorded as an ancestor too, so a downstream SG Publish names what it came from without anyone typing an id.

Install

Through Manager: Custom Nodes Manager, search Flow Production Tracking, Install, restart ComfyUI. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ksallee/sg-comfyui.git
cd sg-comfyui
<comfy-python> -m pip install -r requirements.txt

<comfy-python> matters: it must be the interpreter ComfyUI runs on, not whatever python resolves to. Needs ComfyUI 0.34.0+, Python 3.11, and a site you can log into. Dependencies are just sg-groundtruth from PyPI; torch, numpy and av are deliberately not pinned. Then Settings → SG: site address, Log in, Test, pick a project.

When it goes wrong

• Settings → SG 404s. ComfyUI was already running when the pack landed, so no routes registered. Restart; if it persists, check the startup log - a pack whose import failed registers nothing.
• Pickers are empty. Press Test. An empty project list with a passing Test means the account can see no projects; an empty link list means the project has no entity of that type. A new Shot or Version not showing up is usually the 600-second lookup cache - press Sync from SG.
• Nothing works and the pack looks fine. Run tools/doctor.py with ComfyUI's interpreter; it names both when they differ.
• A Version holding only a zip comes back as a zip. It isn't unpacked.