Repository navigation
v0.35.0-alpha — Qwen-Image 2.1: transparent layers, lettering, editing
Pre-releaseQwen-Image 2.1 comes to three tools you already have. Transparent layers straight from the model, lettering that comes out spelled right, instruction editing, and composition from up to six layers. It also fixes four bugs in the shipped panel.
Licence first. Qwen-Image 2.1's weights are under the Qwen Research License — research and evaluation only, not commercial work. Every 2.1 preset ends in "(research licence)", none of them is a default, and the panel never downloads those files for you. Everything else OpenLayer ships is unaffected.
New: Transparent Background
In Text to Image, choose txt2img-qwen-image-21 and a Transparent Background switch appears under Model. Turn it on, and the result lands as a cut-out layer with a real alpha channel. It's generated that way, not cut out afterwards.
- A studio teapot came back with a clean hole through its handle and a soft contact shadow, and a desk lamp kept its thin pull chain.
- Your prompt stays exactly as you typed it.
- About 15 seconds at 1024² on a 12 GB card.
New: Qwen-Image 2.1 in three tools
- Text to Image: strong lettering. Put the exact words in quotes. Native up to 2048 × 2048 (about 90 s; 1024² takes about 20 s).
- Image to Image, edit: write what should change, e.g. "make it a rainy evening", "replace the cart with a bicycle". Objects stay where they were. Edit a cut-out layer and you get a cut-out back.
- Multi-Reference: name your layers in the prompt as
<image1>,<image2>… in list order. Transparent layers keep their edges. Up to six references, at about 30 s each.
Setup: ComfyUI 0.37.0 or newer, three files, 14.2 GB, and no custom nodes.
Fixed
- Remove Background and Layer Maps results land on their layer again. They were landing in the middle of the canvas.
- The Preview panel's Import button works for Multi-Reference and Unflatten. It used to do nothing.
- Generate with an empty prompt tells you so, instead of silently doing nothing.
- Setup's note on licence-gated models is now accurate.
- The leftover "Spike: Model Download" button is gone from Settings.
Known limitations
- Research licence for everything Qwen-Image 2.1, as above.
- Qwen-Image 2.1 edits repaint the whole frame. Untouched areas can come back slightly darker, so compare before you keep one. Editing only a selection was built and cut: in a smooth sky the edge showed.
- Qwen-Image 2.1 doesn't replace Remove Background. It failed on real photographs in testing; Remove Background stays on BiRefNet.
- Lettering is strong, not guaranteed. Clean type came back exact; stylised hand-painted lettering didn't.
The rest is in known limitations. Verified on Windows 11 with Photoshop 2025; macOS is still untested.
Installing
Download openlayer-v0.35.0-alpha.ccx and double-click it. Creative Cloud installs the panel, and Photoshop lists it under Plugins › OpenLayer. The "not verified by Adobe" warning is expected, because the plugin isn't from Adobe Exchange.
If double-clicking does nothing, load openlayer-v0.35.0-alpha.zip through the UXP Developer Tool: Add Plugin, select the unzipped manifest.json, then Load.
You also need a local ComfyUI server and a GPU with 8 GB of VRAM or more (12 GB targeted). openlayer-comfyui-setup-0.35.0.zip lists every model, and the panel's Setup screen checks the same list against your machine. Checksums are in checksums.txt.
Feedback
Does any of this work on macOS? That's still the one big unknown. Discussions · Issues
What's new, in full
Qwen-Image 2.1 arrives as three opt-in presets in tools you already have: Text to Image, Image to
Image and Multi-Reference. It brings transparent layers straight from the model, exact lettering,
native 2K and composition from up to six layers. Its weights are under the Qwen Research
License — research and evaluation only, not commercial work — so every 2.1 preset says so in its
name, and none of them is a default. The release also fixes four bugs an architecture audit
found in the shipped panel.
Added
-
Transparent Background, in Text to Image. With the Qwen-Image 2.1 preset selected, a new
toggle under Model makes the result a cut-out layer with a real alpha channel — generated that
way, not matted afterwards. Measured on a 12 GB 4070 Ti: a studio teapot came back 69% fully clear
with a clean hole through its handle and a soft contact shadow, and a desk lamp kept its thin pull
chain. About 15 seconds at 1024². The toggle only appears for presets that can do it, and your
prompt stays exactly as you typed it; the RGBA phrasing the model expects is added at submit time. -
Qwen-Image 2.1 in Text to Image (
txt2img-qwen-image-21). Strong lettering: a poster reading
"OPENLAYER LIVE / Friday 26 September - Stockholm" and a handwritten note at 2048² both came back
spelled exactly. Native up to 2048 x 2048. About 15–20 s at 1024², 86 s at 2048². -
Qwen-Image 2.1 editing in Image to Image (
edit-qwen-image-21). Write what should change —
"change the sign to read CAFE OPENLAYER 35", "make it a rainy evening", "replace the cart with a
bicycle" — and objects stay within a pixel of where they were. Edit a cut-out layer and you get a
cut-out back: a transparent teapot went in 69% clear and came out 69% clear with no mention of
transparency in the instruction. About 25 s per edit. -
Qwen-Image 2.1 in Multi-Reference (
multi-reference-qwen-image-21). Name your layers in the
prompt as<image1>,<image2>and so on, in list order — "the woman from<image2>pours tea
from the teapot in<image3>". Transparent layers keep their cut-out edges. Six references
composed a scene, a person, a teapot, a character and two framed pictures in one run. Each
reference adds about 30 seconds (1 → 23 s, 3 → 82 s, 6 → 179 s), which is why the list stops at six
for this preset; the model accepts sixteen.Setup: ComfyUI 0.37.0 or newer, three files, 14.2 GB, all from Comfy-Org's public
repackaging and every node core ComfyUI. Because of the licence, Setup gives you the links and
never downloads them itself.
Fixed
- Remove Background and Layer Maps results land on their layer again. Both captured the layer
trimmed to its own bounds but imported to the centre of the canvas, so a cut-out of an off-centre
subject landed in the middle of the document. They now go back where they came from, as Image to
Image always has. - The Preview panel's Import button works for Multi-Reference and Unflatten. It was enabled for
both and did nothing at all. - Generate with an empty prompt says so, instead of doing nothing. In Photoshop an empty text box
reads asnull, not as empty text, and 43 places in the panel assumed otherwise; the error they
threw was swallowed before any message could appear. - Setup's note on licence-gated models was wrong. It said those files need a sign-in. They
don't — they come from public mirrors — and what actually happens is that OpenLayer never
downloads them for you. It now says that. - Switching Multi-Reference presets applies that preset's steps and CFG, and the hint under the
model picker follows the preset instead of always describing FLUX.2 Klein.
Changed
- The "Spike: Model Download" button is gone from Settings. It was a measurement probe from
before in-panel downloads existed, and wrote a test file to a folder on one developer's machine. - Agent Bridge:
text_to_imageacceptstransparentBackground, and themulti_reference
description names both workflows and which one the faces limitation belongs to.
Known limitations
- Research licence. Qwen-Image 2.1 may be used for research and evaluation only; commercial use
needs a separate licence from Qwen. Every other model family OpenLayer ships is unaffected. - Qwen-Image 2.1 background removal was tested and not adopted. It works on the vendor's sample
image and returned an empty frame on a real photograph of a person in a street, four times out of
four. Remove Background stays on BiRefNet. - Qwen-Image 2.1 edits repaint the whole frame. Positions hold, but untouched areas can come back
slightly darker (6–12 levels measured on photographs), so compare before you keep an edit.
Editing only a selection was built and cut: on busy content the join was invisible, but in
smooth sky it left a visible edge, and feathering only softened it into a darker patch. - Lettering is strong, not guaranteed. Clean type came back exact every time; stylised
hand-painted lettering on a wooden sign came back misspelled. - 2K is slightly off the model's own sampling schedule until ComfyUI fixes
#16447. The 2048² results looked right in
testing. - Whether Qwen-Image 2.1 carries a real person's face through Multi-Reference is untested. It kept
a generated person's hair and clothing; do not assume likeness.