Atlas Attach Source Plate 🎞

Most of the pack works on ordinary IMAGE tensors, but the pack's final plate - the one that gets handed to Nuke, Maya or an EXR exporter - is a different object: an ATLAS_PLATE_REF, registered with colourspace and bit-depth metadata so the DCC handoff is float-safe and colour-managed. Atlas Attach Source Plate 🎞 is the bridge between the two. It takes a registered plate reference and attaches it to a solve, so the solve's projection now references the real file instead of a tensor that exists only in graph memory.

What it's for

Think of it as "tell the solve which file is the official plate." You'll reach for it in a few situations:

• You converted or re-processed a plate (RAW → ACEScg EXR via AtlasLoadRAW, or a debanded/regraded file) and you want the solve to point at the converted artifact rather than the original.
• You want the browser viewport to show the new plate's preview.
• You're wiring a multi-view rig and need each source to carry its own registered plate.

The mechanism is simple: it clones the solve, attaches the plate_ref to the chosen projection source, and - the subtle bit - refreshes that source's embedded image_b64 preview so the viewport doesn't keep showing stale pre-conversion pixels while the exporters already reference the new file.

The input that bites people

• solve - required, the solve you're attaching to.
• plate_ref - required, the ATLAS_PLATE_REF from AtlasRegisterPlate or the RAW loader.
• source_name - leave blank and you attach to the PRIMARY source plate (the original behavior, right for single-photo work). Set it to a ProjectionSource name (e.g. photo_2 from a multi-view solve) to retarget that source's plate instead.

That source_name field is where the gotcha lives. In a two-photo rig, attaching only to the primary leaves the second source still tagged Linear Rec.709 while the first is ACEScg - and only the anchor plate renders correctly in Nuke. Retargeting the non-anchor source by name is the only way the converted plate reaches it. If the name doesn't match a known source, the node raises with a list of the known names, so a typo is loud rather than silent.

Output is a single updated ATLAS_SOLVE - wire it onward into the viewport, the layer stack, or the exporters. It's a small node, deliberately so: its whole job is to make sure the solve's plate reference is the right one before the heavy lifting downstream. Install is the pack's clone-and-go, no extras needed.