OpenLayer v0.19.0-alpha

OpenLayer is a local-first Photoshop panel that drives your own ComfyUI. This is an alpha for testing and feedback — see become a tester.

Multi-Reference Composition: give the panel a list of captured layers instead of one, and it builds a single picture out of all of them — a background, a person, an object, however many you add. Reference order sets the output canvas and affects composition; four questions about how the technique actually behaves were answered with 48 live ComfyUI generations before any of the screen was built. The finding that matters most for what to expect: clothing, props, setting and lighting carry across from a reference — a specific person's face does not.

Install

One-click (Windows): download openlayer-v0.19.0-alpha.ccx and double-click it. Creative Cloud installs the panel; it will warn that the plugin is not from Adobe Exchange, which is expected. Then open Photoshop → Plugins → OpenLayer.

Manual (all platforms, and the fallback if double-click does nothing):

1. Start ComfyUI on http://127.0.0.1:8190.
2. Download and unzip openlayer-v0.19.0-alpha.zip.
3. In the UXP Developer Tool: Add Plugin, select the unzipped manifest.json, then Load.
4. In Photoshop: Plugins → OpenLayer.

openlayer-comfyui-setup-0.19.0.zip is optional — it is the model/node checklist the Setup screen reads, for setting ComfyUI up from scratch. Multi-Reference needs nothing beyond what Text to Image with FLUX.2 Klein already needs.

Tested versions

Two separate claims:

	Version
Required by the plugin manifest	Photoshop 25.0.0 (2024) or newer
Actually verified this release	Photoshop 2025 on Windows 11, ComfyUI on a 4070 Ti (12 GB)

macOS is untested. If you run it there, a report either way is valuable.

Checksums (SHA-256)

7dfd851961070d9ef77805e4fa9191de58c88cfc3f7c8af7a9ad5b853b0eb441  openlayer-v0.19.0-alpha.zip
e02fd14f04ae6d68ff95a9d11737406129ee7de3f4fc13900ff769e5dbee310e  openlayer-v0.19.0-alpha.ccx
869799e5068847bad98c37849fe845e12ec5ce575f541a706cc608cae44082c2  openlayer-comfyui-setup-0.19.0.zip

Verify: shasum -a 256 -c checksums.txt (the checksums.txt asset is attached), or on Windows Get-FileHash -Algorithm SHA256 openlayer-v0.19.0-alpha.zip.

What's new

v0.19.0-alpha - 2026-08-28

This release adds Multi-Reference Composition: give the panel a list of captured layers instead of one, and it builds a single picture out of all of them. Before any of the screen was built, four questions about how the technique actually behaves were answered with 48 live ComfyUI generations, each run as a control/variant pair — full results in docs/multi-reference-gate-findings.md. Two of the four answers shrank the feature that got built; the third is the honest limit stated everywhere the tool is described.

Added

• Multi-Reference Composition (experimental). An ordered list of reference layers — Add Active Layer or Add Canvas to grow it, Up/Down/Remove on each row — composed into one image with multi-reference-flux2-klein, which chains a ReferenceLatent per reference onto both conditioning branches of FLUX.2 Klein. Every node is core ComfyUI, and it shares the Klein 4B stack already needed for the other Klein presets, so it downloads nothing new.

  Reference order is a real control, not a convenience: the first reference sets the output canvas, and gate testing found a wide, thin object that has to sit behind the other subjects can render duplicated or distorted unless it comes earlier in the chain — moving one item fixed a duplication that reproduced on every seed tried. Testing found no reference count, up to six, at which composition quality fell off, so the eight-reference cap is a sanity bound rather than a measured limit. Positional prompt wording ("the second image") turned out not to matter at all — the model matches a reference to the prompt by what it depicts, not by its position in the list — and masking or cutting a reference out from its own background made no measurable difference, because Klein already isolates the subject on its own. Neither slot numbering nor a masking control made it into the screen as a result.

  What did not survive testing is likeness. Clothing, props, setting and lighting carry across from a reference; a specific person's face does not. Four real photographs — two modern, two archival — all came back as a plausible stranger rather than the person photographed, including a clean, frontal, evenly lit studio portrait, which rules out framing or resolution as the cause. Generated sources fare much better, because the model reproduces its own output far more faithfully than it reproduces a photograph — which is exactly the trap the test was designed to catch. The tool's subtitle, its in-panel info note, and its MCP tool description all say this plainly: it composes a scene, it does not place a recognisable person into one.

• multi_reference joins the MCP Agent Bridge, alongside the seven tools already reachable there. Same boundary as every other bridged tool — an agent can set the prompt and sampler values and press Compose, but cannot capture layers, add to the list, or reorder it — and the likeness limit is stated directly in the tool description an agent reads before running it.

Fixed

• Reordering or removing a reference left the progress bar spinning with nothing running behind it. setStatusProgress reads status tone "idle" as work in progress unless the message contains a word from a short completion whitelist (ready, complete, copied, saved, reset, cancel) — right for "Uploading reference 2 of 3...", wrong for "Reference moved to position 2.", a finished action whose wording happened to match none of them. Both list actions now report tone "ready", like every other completed action in the panel already does. Found in a real Photoshop session.

• Composing from two layers captured together produced the wrong background at the wrong size. Captures are named to the minute, so two layers added one after another — the ordinary way to build a reference list — arrived at ComfyUI with identical filenames, and the upload's overwrite=true meant the second silently replaced the first. Every LoadImage in the chain then read one picture: the background became a copy of the last reference, and the output canvas took that reference's size instead of the first one's. Multi-Reference is the first tool to upload more than one image per generation, which is why nothing had hit this before. Each reference now uploads under a name derived from its own list entry rather than its capture filename. Reproduced against live ComfyUI before the fix and confirmed against the original report's screenshot after it.