OpenLayer
Alpha v0.37.0 · free & open source
OpenLayer connects Photoshop to your own ComfyUI server. Generate with Stable Diffusion, SDXL, and Flux — then import results as editable layers. Your art, your GPU, your machine. No cloud. No credits. No subscription.
v0.10 setup that explains itself
OpenLayer starts with dependable local workflows — text to image, image to image, sketch to image, experimental inpaint and outpaint, pixel upscale, lossless PNG source capture, live previews, diagnostics, and clean import into the active document. v0.10 turns to the part nobody enjoys: setup. Workflow Health now tells you when a "missing" model is really just one folder over, and names the repository behind a missing custom node. The plugin zip is also built to spec for the first time, which is what makes it unpack correctly on macOS.
Tools
Prompt, negative prompt, size, steps, CFG, and seed into your checkpoint of choice — SD 1.x, SDXL, FLUX.2 Klein, Z-Image-Turbo, Krea-2, and Qwen-Image 2.1, which can hand back a transparent cut-out layer and letters spelled exactly.
Capture the active layer or visible canvas as lossless PNG, reinterpret it with a prompt, and import the result — optionally automatically.
Say what should change — “make the jacket red”, “remove the parked car” — and the rest of the layer stays where it is. FLUX.2 Klein by default; Qwen-Image 2.1 keeps a cut-out a cut-out.
Guide generation with your linework through an SD 1.x LineArt ControlNet workflow, with adjustable control strength.
Make a Photoshop selection, capture source and mask, repaint just that region with SD inpainting or Flux Fill, and import aligned to your selection.
Extend content beyond the canvas edges with Flux Fill and per-edge padding and feathering control.
Enhance layers with local upscale models such as 4x-UltraSharp — fast pixel/model upscaling through ComfyUI.
One button, and the subject comes back on its own layer with real alpha — about two seconds, no prompt and no checkpoint, because there is no diffusion in the graph at all. The pixels are your own at full resolution with only an alpha channel added, so nothing is re-rendered or softened.
Read a depth, line-art or normal pass off any layer, returned at that layer’s exact pixel size so it sits straight over it. Photoshop has no normal-map or line-art export, and its Depth Blur filter needs an internet connection — this runs on your machine. A few seconds per map, no prompt and no checkpoint — there is no diffusion in the graph. Line art comes back black on white, ready to ink or colour over.
Describe any layer or canvas into editable prompt text with Florence-2 PromptGen, then send it straight to Text to Image.
Keep painting and OpenLayer keeps repainting — a fast local preview tier that follows your strokes, plus a slower Krea-2 Turbo refine pass on demand.
Capture several layers into an ordered list and compose them into one picture with FLUX.2 Klein — clothing, props, setting, and lighting carry across from each reference. A specific person's face does not; treat it as scene composition, not as a way to place someone recognisable into a picture.
Hand it a flat layer and get the picture back as separate layers with real transparency, imported in stacking order. A photograph of someone on a bench comes back as ground, bench, and person. It needs a subject standing in front of a background — a close-up that fills the frame has no front and back to find.
Send pixels the other way. Export the active layer, the current selection, or the selection mask — to a file you pick, or straight into ComfyUI's input folder where a workflow can use it.
One click tells you which presets are ready on your ComfyUI — missing models, missing nodes, and beginner-friendly next steps.
Results · v0.20
Every shot below is the panel and the Layers panel in one frame, because the layer that arrives is the entire point. Nothing here is a mockup or a marketing render.
Interface shots · v0.6
Each view below is a real local workflow, compressed for the web and loaded only when it approaches the viewport.
Live demos
Full local sessions, not clips — a real ComfyUI server and a real Photoshop window, start to finish.
Oldest to newest
The visual language stayed dark, local, and Photoshop-aware. Each release removed friction while preserving the original goal: make AI output behave like editable layer work.
Workflow
Setup
Get OpenLayer from its Adobe Exchange listing with one click. Want the latest alpha instead? Grab openlayer-v0.37.0-alpha.ccx from the latest release and double-click it. Creative Cloud installs the panel — no developer tools, nothing to build.
The Exchange install is reviewed by Adobe; the direct .ccx is unsigned, so Creative Cloud will note it is not verified. Accept, then find it under Plugins > OpenLayer. Prefer to build from source? The README has the UXP Developer Tool route.
python main.py --listen 127.0.0.1 --port 8188 --preview-method autoCheck ComfyUI, choose a checkpoint, enter a prompt, generate, and import the result into the active document.
Missing a model? The Setup screen lists every one the presets need and downloads it for you — one at a time, resumable, and only after a confirmation naming the file size, the folder it lands in, and the host it comes from. Licence-gated weights and custom node packages are the exceptions and stay a manual job; the screen says which is which.
OpenLayer defaults to http://127.0.0.1:8188, which is ComfyUI's own default port, so an untouched install already matches — and if yours runs somewhere else, you do not have to move it. If ComfyUI is already running anywhere, open Settings and click Find ComfyUI Active Port; OpenLayer finds it, connects, and reports Ready. --preview-method auto enables live step previews.
Upgrading from v0.20 or earlier? The default port changed from 8190 to 8188. OpenLayer used 8190 for twenty releases and now uses the port ComfyUI itself starts on. If you had ever connected successfully your address was saved, so nothing changes for you. If the panel does come up on 8188 while your server is on 8190, it scans for the running server by itself on first launch, and Settings › Find ComfyUI Active Port puts it back in one click at any time — 8190 is still in the scan list.
Current alpha status
A second dockable panel shows the current generation at any size, mirroring every tool including Live Painting. It can be pinned to one tool, and it imports the result into your document without going back to the dashboard. It has passed a real Photoshop UXP smoke test.
Claude, Codex, or anything else that speaks the Model Context Protocol can call thirteen of OpenLayer's tools directly, running the panel's own buttons in your open document. It is local-only, off until you turn it on, and not included in the download — it lives in the repository. Verified end to end generating real images in Photoshop with Claude, and separately confirmed connecting and reading panel state with Codex.
OpenLayer v0.37.0-alpha is for local testing and feedback. It is not production-ready yet.
The Qwen-Image 2.1 presets — transparent layers, lettering, editing and composition — use weights Qwen licenses for research and evaluation only. Each one says so in its name, none is a default, and the panel never downloads those files for you. Everything else OpenLayer ships is unaffected.
With FLUX.2 Klein, 48 test generations found clothing, props, setting, and lighting carry across from a reference layer — a specific person's face does not, even from a clean studio portrait. It is honest as a composition tool and should not be used expecting a recognisable result.
Testing found that what decides whether a picture separates is its composition, not where it came from: a generated close-up fails exactly as a photographed one does, and a subject standing clear of its background succeeds either way. The number of layers you ask for is a ceiling — empty ones are left out rather than imported. As of v0.21 the panel detects a run that separated nothing and reports it, instead of presenting the unchanged picture as a result.
v0.37.0 retains the upload and alignment fixes while output quality across checkpoints continues to be validated by testers.
Flux1-dev fp8 Text to Image, Flux Fill Inpaint, and Outpaint run locally but still need careful validation and tuning.
OpenLayer keeps API and source workflow folders, but custom workflow import and automatic node mapping are not built yet.
txt2img-basic, img2img-basic, and sketch2img-linecn-basic are starter workflows, not final production presets.epicrealism_naturalSinRC1VAE.safetensors and offers LineArt, Scribble and Depth ControlNet presets, each needing its own ControlNet model.--preview-method auto..ccx/.zip download — a plugin package inside Photoshop cannot start a Node program, so it lives in the repository. The panel asking the assistant for a prompt additionally depends on the AI client supporting MCP sampling, which is optional in the protocol.Roadmap
Text to Image, preview, import, checkpoint selection, Image to Image, Sketch LINECN, source capture, and setup documentation.
PNG capture, diagnostics, Prompt from Layer, Upscale, queue-aware Cancel, History metadata, live previews, and experimental Inpaint and Outpaint.
A denser dashboard, sticky headers, determinate progress, Advanced settings, clearer availability, and prompt fields hardened in the real Photoshop renderer.
A separated dockable preview panel mirroring every tool, a generated ComfyUI setup pack with verified model download locations, and one fewer custom-node dependency.
Persistent preview pinning, the missing editable workflow sources plus a source/API drift check, four user-visible fixes to status and progress, and a smaller App.ts.
Export the active layer, the current selection, or the selection mask to a file or straight into ComfyUI's input folder — and every tool's messages finally stay on its own screen.
Workflow Health finds models sitting in the wrong folder and names the repository behind a missing node — and the plugin zip is finally built to spec, so it unpacks correctly on macOS.
A Setup screen listing every model and node package with its folder, size and status — usable before ComfyUI is even running — plus a ranking of which presets your card will actually run well.
A Text to Image preset for the GGUF-quantised FLUX.2-dev, built on ComfyUI's own advanced sampler chain — and every preset the panel lists is now one you can actually run, because the two that never could have been are gone.
An optional LoRA on eleven presets, two new Sketch to Image ControlNet presets, and a Setup screen that downloads missing models itself from the project's own pinned URLs instead of only telling you where to find them.
Compose one image from an ordered list of captured layers with FLUX.2 Klein, gated on 48 live test generations before any screen was built: reference order affects output more than reference count, and a specific person's likeness does not carry the way clothing and setting do.
A flat layer comes back as separate layers with real transparency, inside a group, in stacking order — the layers in front rebuilt from your own full-resolution pixels wearing the model's matte. Gated on eight questions answered with live generations first, two of which overturned assumptions the plan was built on.
One button, and the subject comes back on its own layer with real alpha — about two seconds, no prompt and no checkpoint, because there is no diffusion in the graph at all. The pixels are your own at full resolution with only an alpha channel added. The rest of the release is honesty work: the panel now says when Unflatten failed to separate anything, and seven warnings it had been hiding are back on screen.
Finish Inpaint and Outpaint alignment, strengthen workflow validation, and turn the metadata foundation into repeatable layer-level iteration.
Guided custom workflow import, ControlNet panels, LoRA browsing, batch variants, persistent AI layer metadata, and faster live-preview generation.
OpenLayer is early, practical, and open for contributors who care about artists, local workflows, and clean Photoshop integration.