v0.40.0-alpha — remove, select, re-shoot

Four new things to do with a layer: remove an object (LaMa for small things, LaMa + FLUX.2 Klein for a whole person), select by naming it (SAM 3.1), see it from another camera position (Camera Angle, experimental), and restore it at 2x with SeedVR2. Results also land at the right size in documents that are not 72 ppi.

Install

1. Start ComfyUI (the panel looks for it on 127.0.0.1:8188; Settings › Find ComfyUI Active Port finds it elsewhere).
2. Double-click openlayer-v0.40.0-alpha.ccx. Creative Cloud installs it and Photoshop lists it under Plugins › OpenLayer.
3. If double-clicking does nothing, unzip openlayer-v0.40.0-alpha.zip and load its manifest.json in UXP Developer Tool (Add Plugin › Load).
4. New to ComfyUI models? Open Setup and press Download starter. Each new tool lists its own files in Setup, with sizes and download buttons.

The Adobe Exchange listing can trail this release until Adobe approves an update.

Tested versions

Manifest minimum	Photoshop 25.0 (2024)
Verified	Windows 11, Photoshop 2025, RTX 4070 Ti (12 GB), 32 GB RAM, ComfyUI 0.37. Each tool was checked in Photoshop in its own pull request.
Photoshop 2024	Not run for this release; reports welcome in issue 127
macOS	Untested

checksums.txt lists the SHA-256 of every asset.

What's new

Four new things to do with a layer: remove an object, select by naming it, see it from another
camera position, and restore it with a generative upscaler. Plus a fix for every import into a
document that is not 72 ppi.

Added

• Remove object (LaMa), a new Inpaint preset. It fills the selection from its surroundings
  in about 2 to 4 seconds, with no prompt and no diffusion model. It is for small things: a sign,
  a cable, a bird in the sky. On large areas and on people it smears. It needs the
  comfyui-art-venture node pack and big-lama.pt (206 MB, Apache-2.0).
• Remove object (LaMa + Klein), experimental, for large things such as a standing person.
  LaMa fills the selection first, then FLUX.2 Klein repaints the fill with real detail, in about
  16 seconds once loaded. The prompt describes what should be there instead ("dry hillside with
  mountains behind"); with an empty prompt the fill stays a smear. The panel grows the selection
  by 24 px for this preset, because a selection that hugs the person leaves their outline behind
  and Klein paints them back from it. Denoise opens at 1. No new download if Remove object (LaMa)
  and Klein inpaint already run.
• Select by Text. Type what you want, for example backpack or chair, table, and it
  becomes a live selection in your document, using Meta's SAM 3.1 in about 2 seconds. Nothing
  else in the document changes. If nothing matches, it says so and leaves your selection alone.
  The selection has hard edges; feather it or use Select and Mask for a soft one. One 1.7 GB file,
  commercial use allowed under Meta's SAM License.
• Camera Angle, experimental. Capture a layer, choose where the camera stands (8 directions,
  4 heights, 3 distances) and get the picture re-shot from there as a new layer. It uses
  Qwen-Image-Edit 2511 with fal's multiple-angles LoRA, all Apache-2.0, about 31 GB of files.
  Measured on a 12 GB card: 34 seconds for the first run, 20 to 22 seconds after it. The whole
  picture is redrawn from the new position, background included, and a face will not keep an
  exact likeness.
• SeedVR2 3B upscale, a second Upscale preset. It restores a layer at 2x and can invent
  detail that was not in the source, unlike 4x-UltraSharp, which only enlarges pixels. Core
  ComfyUI nodes only, Apache-2.0. It refuses a source whose result would not fit your card's
  VRAM, instead of running for minutes and failing; the limit scales with the card.

Fixed

• Imports into documents that are not 72 ppi. Every placed result came in at the wrong size
  in, for example, a 240 ppi document: 3.33 times too large. Inpaint then reported "No pixels are
  selected" and fell back to an unmasked layer. Results now carry the document's resolution and
  land at 1:1. Checked in Photoshop at 240 ppi and 72 ppi.
• Inpaint no longer generates from a selection you have since changed. If the live selection's
  bounds differ from the captured one, Generate stops before anything is sent and asks you to
  capture again. Deselecting keeps the captured selection.
• A failed capture no longer leaves the previous one in place. In Inpaint, Image to Image,
  Sketch, Outpaint, Upscale, Style Reference, Unflatten and Prompt from Layer, a failed capture
  kept showing the old source and left Generate enabled, so the next press ran on a capture you
  had already replaced.
• The Preview panel can import Remove Background and Layer Maps results. It showed the
  picture but never offered the Import button for these two tools.
• The Preview panel says why Import is unavailable. It blamed a running generation whenever
  Import was disabled, including when nothing was running and the picture was simply an older
  result.
• The Upscale picker lists only SeedVR2's own file for the SeedVR2 preset, instead of every
  diffusion model on the server.

Known limitations

• Camera Angle and memory: on the 32 GB test machine, system RAM dropped to about 3 GB free
  during the first run. With 16 GB of RAM the first run may fail or swap.
• The Inpaint selection check compares only the selection's bounding box. A change that stays
  inside the same box is not caught.
• Select by Text, Camera Angle and Remove object are not on the MCP agent bridge yet.
• One unexplained refusal: in one test, the first Inpaint Generate after loading the panel was
  refused with "The captured selection sits outside its own context". Capturing the same
  selection again worked.
• Tested on Windows 11 with Photoshop 2025 and an RTX 4070 Ti (12 GB). Not run on Photoshop 2024
  or macOS. The Adobe Exchange listing can trail this release.