Transparent PNGs Straight From an AI Image Model: The MOSS-Image Alpha Add-on in ComfyUI, and Why It Beats Background Removal on Glass
Background removal is a guess made after the fact. The tool looks at a finished picture of a wine glass on a grey table and has to decide, pixel by pixel, which grey belongs to the table and which grey is the table seen through the glass. It usually picks one. The glass comes out solid, or it comes out with a hole where the stem should be.
Every game artist, sticker maker and motion designer knows this tax. You generate the item, you cut it out, you fix the edges, you fake the transparency back in by hand. For a sword it costs a minute. For a fairy wing, a soap bubble or a crystal it can cost the afternoon.
There is another way to do it. Instead of removing a background from a finished image, you have the model paint the object with its transparency already decided. It knows the wing is half see-through because it made the wing.
On October 24 a small add-on called MOSS-Image Alpha showed that working on an ordinary home graphics card, inside ComfyUI, with an open poster model. It is rough, and its licence keeps it out of paid work for now. It is still the clearest look yet at a workflow that will replace a lot of cut-out time.
What the add-on does
MOSS-Image is an open image model from the OpenMOSS team, released in mid-October and built for posters and layouts. The add-on, made by a community creator who goes by paralaif, is a small extra file you load on top of it. With the add-on loaded, the model produces a PNG with a real alpha channel, the layer that stores how transparent each pixel is.
The author describes it simply: "Text-to-image with a real alpha channel for MOSS-Image." You end your prompt with a fixed phrase, the image comes out of the model already cut out, and partial transparency works. The example gallery on the card includes pixie wings, gems, a water creature, the lattice of the Eiffel Tower, and close-ups shown over a gradient so you can see the soft edges.
The scale of the project is part of the story. It was trained on 39 images, in about 40 minutes, on a single RTX 5060 Ti with 16 GB of memory. That is a mid-range gaming card, not a data-centre rack. The author calls it a proof of concept, and that is the right label.
Why painting transparency beats cutting it out
The interesting detail is buried in the card's research notes.
To get an image with transparency out of a model, you need an encoder-decoder (the part that turns the model's internal working into actual pixels, and back) that knows about a fourth channel. The add-on uses one made by researcher Xavier Jara, a version of the FLUX.2 encoder-decoder widened from three colour channels to four and retrained to carry transparency.
You might think that decoder is enough on its own. The author tried that. Plugged straight onto plain MOSS-Image, the card says it "behaves roughly like a chroma key." In other words, it mostly knocks out a background colour, the way a green screen does. Same problem as before: a guess based on colour.
The add-on is what changes the result. It teaches MOSS-Image to produce an image that carries the transparency information from the start. That is why the glass and the wings work at all. The model is not separating subject from background. It is drawing a subject that has no background.
For anyone who composites, this is the difference between a key and a matte rendered out of a 3D package. One is recovered. The other was never mixed in.
Put this into practice: cut-out game items and stickers
The card includes a ready-made ComfyUI workflow. Here is the route, in plain steps.
1. Install the custom node. The add-on folder includes a node called ComfyUI-RGBA-VAE, which adds a "Load RGBA VAE (Flux2)" loader. ComfyUI's standard decoder loader will fail on this file, so the node is not optional.
2. Download four things.
- The add-on itself,
moss_alpha_v1.safetensors, intomodels/loras/. - The compressed MOSS-Image model and its Qwen3.5 9B text encoder from Comfy-Org's MOSS-Image page.
- The transparency decoder from trmz/flux2-klein-alpha, renamed to
flux2_klein_alpha_rgba_vae.safetensors. - Optionally, an 8-pass speed add-on for MOSS, and the standard FLUX.2 decoder if you want an opaque preview.
3. Load the workflow. Open moss-alpha-lora-t2i.json from the add-on's comfyui folder.
4. Keep the strength at 1.0. The author says 0.8 still works. At 0.6, subjects start turning see-through, which is the add-on's way of telling you it has lost the plot.
5. Write prompts about the object, never the place. End every prompt with the trigger phrase: ", isolated on a transparent background". Then describe only the subject. "A bronze key with a ruby set in the bow, isolated on a transparent background" will work. Anything that implies a setting will fight you. The card warns that prompts suggesting a location, a famous landmark for instance, can produce a full scene and break the transparency. For characters, add "full body" to keep them inside the frame.
6. Pick your speed. Steps are how many refinement passes the model makes on each image. With the 8-pass speed add-on, run 8 steps. Without it, run 28. The author's settings for each are in the workflow, and the tested sizes are 1024 square, plus 832 by 1216 and 1216 by 832.
7. Save with the "RGBA PNG" node. RGBA means red, green, blue plus alpha, the transparency layer; this is the node that writes it into the file. A normal save node will flatten it.
8. Check it over three backgrounds. Drop the PNG over black, white and a mid grey in your editor before you trust it. Halos show up on dark backgrounds. Faint leftover patches show up on light ones. Grey tells you whether the soft edges are soft in the right places.
That last step costs a minute and saves you from discovering a white fringe after the sticker is printed.
Who this is for
If you make game assets, this is a direct fit: swords, potions, chests, gems, creature sprites, user-interface icons. The card's gallery is heavy on exactly that kind of thing.
Sticker and merch designers get clean die-cut shapes without tracing. Motion designers get overlay elements, a glowing rune or a floating crystal, that drop straight onto a timeline. Logo and badge experiments work too; the gallery includes a game logo and a café logo.
It pairs naturally with the transparent video work that showed up this week. Wan-Alpha++, released on October 19, animates a single transparent asset into a transparent clip. A painted-transparency still is exactly the kind of input it wants.
Where it breaks
This is a 39-image experiment, and it behaves like one.
Halos and soft patches. Subjects far from the training examples can come out with halos, faint patches or soft edges. The three-background check catches them.
Dark clothing goes see-through. The card's example is a witch's cloak that turns translucent. If your character wears black, expect trouble.
Fire and smoke are the weakest category. Which is a shame, because those are precisely the effects compositors most want with real transparency.
Glass is inconsistent. It is the category that shows off the idea best, and the one least likely to work on the first try. Run several seeds.
It changes the picture. Compared with plain MOSS-Image, the add-on shifts composition and costume details, and it favours centred, full-body subjects. Off-centre layouts are not its strength.
You need a real graphics card. The author trained it on a 16 GB NVIDIA card. The card does not state a minimum for running it, and you are loading a large 9B text encoder (the part that reads your prompt) alongside the image model, so a smaller card will struggle.
And the big one: you cannot sell what it makes. Eight of the 39 training images were generated with Qwen Image 2.1, whose licence restricts commercial use, so the author releases the add-on for non-commercial research and evaluation only. Separately, the transparency decoder it depends on is listed on its own card under the FLUX Non-Commercial License. MOSS-Image itself is Apache 2.0, but the chain is only as open as its least open link.
So this is a tool for testing the workflow, building personal projects and prototyping a game's look, not for shipping a client's asset pack.
The bigger point
What makes this add-on worth your evening is not its quality. It is how little it took. Thirty-nine images and forty minutes on a gaming card were enough to teach an open model to paint transparency. That tells you the hard part is solved in principle, and that a properly trained, commercially licensed version is a matter of someone doing the same thing with clean data.
When that arrives, the cut-out stage of a lot of creative work shrinks to almost nothing. The skill that survives is the one this workflow already asks of you: describing an object without a world around it, and checking edges like someone who has been burned by a white fringe before.
So try it now, while the stakes are low. Pick five assets from a project you know well, one of them glass or a wing. Generate each one with the add-on and cut out the same subject the old way. Put both over black. You will see exactly where painted transparency wins, where it still loses, and how much of your week it might give back once the licence catches up.