[DRAFT] feat(ByteDance): LAYERS output on Seedream Layer Separation

[purzbeats]

Stacked. Base is #15366 (for the tooltip context). Also requires #15317, which supplies the LAYERS io type — this does not import without it.

Renames the existing layers IMAGE output to layers_batch and adds a new layers output of type LAYERS: the base plate at z_index 0 with each element above it, carrying its own placement, name and stacking order. Wires straight into Create Layered Image — no Batch Images, no Add Layer.

base_image    -> IMAGE
base_mask     -> MASK
layers_batch  -> IMAGE          (renamed)
masks         -> MASK
bboxes        -> BOUNDING_BOX
layers        -> LAYERS         (new)

Why

The outputs currently cannot drive the compositor with placement intact. expand_item_frames applies an item’s single x/y/w/h/name/z_index to every frame of a batch, so 16 layers share one placement. That is why full-canvas mode is the only workable path today — placement is baked into the pixels, so x=0, y=0 is correct for all of them. It costs ~768MB of layer tensors plus ~268MB of masks at 2K with 16 layers, and leaves minimal-size mode unusable with the compositor.

One document item per layer sidesteps this with no change to the compositor: document_items already sorts by z_index, expand_item_frames already handles single-frame items correctly.

The bboxes output is currently dead

Its tooltip promises “Placement boxes for Create Layered Image”, but the bbox input it was written against was removed in 1c4953ff. #15386 revives it generally for any bbox producer; this PR is the direct path for this node specifically. They are alternatives — pick one.

Notes

• New output is appended, not placed next to layers_batch, since outputs are wired by index and reordering breaks saved workflows. Happy to move it; the node is unreleased.
• Layers flagged bbox_degenerate / bbox_out_of_canvas are skipped rather than added as invisible items.
• Crop mode keeps each item at its own size; full-canvas mode puts items at the origin since placement is already in the pixels.

Before judging round-trip fidelity

The compositor composites in linear light (srgb_to_linear in, linear_to_srgb out — symmetric and deliberate). Hard edges reconstruct to within 0.50/255; anti-aliased edges diverge by up to ~30/255 if the model decomposed assuming sRGB-space compositing. Worth confirming with ByteDance which space they assume — the claim that elements “directly restore the original image” only holds under one of them.

Draft for @bigcat88 and @jtydhr88. cc @christian-byrne

API Node PR Checklist

Scope

• Is API Node Change

Pricing & Billing

• Need pricing update
• No pricing update

If Need pricing update:

• Metronome rate cards updated
• Auto‑billing tests updated and passing

QA

• QA done
• QA not required

Comms

• Informed Kosinkadink

[github-actions]

🎉 Thank you for your contribution, we really appreciate it! 🎉

Like many open source projects, we require contributors to sign our Contributor License Agreement (CLA). A CLA makes the ownership of contributions explicit, so contributors and the project share a clear understanding of how the code can be used. By signing, you:

• Confirm that you own your contribution.
• Keep the right to reuse your own code.
• Grant us a copyright license to include and share it within our projects.

CLAs are standard practice across major open source projects including those under the Apache Software Foundation and the Linux Foundation. Ours is based on the Apache Software Foundation's CLA. Most importantly, it would enable us to relicense the project under a more permissive license in the future, giving the project and its community greater flexibility.

✍ To sign, please post a new comment on this PR with exactly the following text: ✍

---

I have read and agree to the Contributor License Agreement

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[bigcat88]

this draft was helpful. closing it, we checked everything from our PR against this draft already.