Photoshop's Third-Party Model List Hits 33, and That Is the Real Story

Adobe released Photoshop 27.11, and the headline feature is small: Enhance Edge, a new option in the contextual task bar that cleans up masks generated by Remove Background. It uses RGBA diffusion to eliminate color fringing at the edges, which matters most when you are cutting out hair or semi-transparent materials and the old result left a halo.
Enhance Edge costs generative credits. Twenty per use, separate from the Remove tool itself, which means regular users will likely need a Firefly subscription to lean on it.
Twenty credits for a cleaner mask edge is a fair trade for anyone doing volume cutout work, and a mild irritation for everyone else. The part worth noticing is that the refinement is billed separately from the removal, which is a small signal about how Adobe is slicing the editing workflow into metered steps. Removing a background is one charge. Making the removal look right is another. It is the same pattern that shows up in the Relight feature discussed below, where the preview is free but the committed render costs.
That is a nice quality-of-life update. It is not the interesting part of this release.
The model marketplace is the actual product
Photoshop now lets you generate images from text prompts using GPT Image 2 inside the Generate Image feature, and make edits through Prompt to edit, which Adobe introduced in 27.10. Adobe has expanded the list of third-party AI models available across its products to 33, with Photoshop users also reaching models from the Gemini and FLUX families, plus models from Topaz Labs, which Adobe acquired last month.
Read that list again. Google's model. Black Forest Labs' model. OpenAI's model. All sitting inside Adobe's flagship application.
Adobe spent the last several years positioning Firefly as the commercially safe, IP-clean option for brand teams, the model you use when legal is watching. That positioning survives, though it has been demoted. Firefly is one option in a dropdown now, competing on quality and price against models from the companies that Adobe once framed as the risky alternative.
The strategic logic is defensive and clear. If designers are going to use third-party models anyway, and they are, Adobe would rather they do it inside Photoshop than in a separate tab that gradually replaces the need for Photoshop. Owning the workspace matters more than owning the model.
There is a version of this story where Adobe loses. It happens if the third-party models get good enough that the surrounding application matters less than the output, and designers start generating final assets in a browser tab and opening Photoshop only to touch them up. Adobe is betting that the layer stack, the selection tools, the color management, and the print pipeline are worth more than any single model. Given how much of professional work is finishing rather than generating, that bet looks sound for now. It becomes less sound every time a model ships that can output at production quality in one pass.
Adobe also mentions that it is now using real-world usage data, from users who have not opted out of collection, to improve Photoshop performance on Windows. The release notes claim a measurable performance improvement on common workflows without putting a number on it. Telemetry-driven optimization is standard practice in software, though it lands differently in a professional creative tool where the work itself may be sensitive around what gets measured and what does not.
The credits economy underneath
The Enhance Edge pricing tells you how Adobe plans to make this work financially. The model may come from someone else, but the render still runs through Adobe's cloud, and Adobe bills it in credits. A model marketplace with a metered consumption layer is a healthier business than a model marketplace alone, because it captures value from every generation regardless of whose weights did the work.
That structure has a second effect worth watching. When 33 models are available and each render carries a credit cost, model selection becomes a purchasing decision with real money attached. The feature that will matter most to professionals is not any single model but whatever comparison or cost-preview tooling Adobe builds around choosing between them.
Meanwhile, the beta channel is doing something bolder
Photoshop 27.12 in the beta channel ships AI Relight, which places up to three virtual light sources onto a finished photograph and re-renders the scene's illumination on a separate non-destructive layer. Position, intensity, diffusion, color, and a Close/Far control that determines how light wraps through the scene's depth are all adjustable.
For ecommerce and advertising photographers, this addresses a specific and expensive problem: a shoot has wrapped, the light direction is wrong for the creative brief, and the alternative is booking a reshoot. For retouchers building composites, it means matching a placed subject's illumination to a new background without hand-painting.
The Close/Far control is the detail that makes it more than a filter. It determines how the light wraps through the scene's depth, which is what separates illumination that reads as physically plausible from illumination that reads as painted on. Diffusion is the other lever that matters: low values create hard-edged directional shadows, high values produce soft wrapping light with gradual transitions. Getting those two controls right is most of the craft in product photography, and moving them into a panel is a meaningful transfer of expertise from the studio to the software.
The economics are structured to encourage careful setup. On-canvas preview renders continuously at low resolution for free, so you can move lights and adjust intensity all session without spending anything. Adobe charges 30 credits only when you click Done to commit the full-resolution cloud render. Reopen a saved Relight layer and commit again, and it deducts another 30.
Mapping a standard three-point setup, key light for direction, fill light for shadow detail, rim for subject separation, onto the three available slots produces predictable results and minimizes the number of paid renders you need.
What this says about where editing is going
Two releases, one shipping and one in beta, point in the same direction. The value in professional image editing is migrating from the generation model to the workflow around it: the non-destructive layer stack, the credit meter, the ability to swap models without leaving the canvas.
Whoever owns that workflow gets paid regardless of which lab has the current best weights. Adobe is making sure that whoever is Adobe.
The open question is whether that position holds. Model quality is converging faster than workflow quality, and the gap that justifies staying inside one application narrows every quarter. Adobe's answer is to keep adding capabilities that are hard to replicate outside a mature editor, which is why a lighting tool like Relight matters more strategically than it looks. Relighting a finished photograph is a scene-understanding and compositing problem rather than a generation one, and compositing has been Adobe's home turf for three decades.
Related articles
Satellite Photos Are Now Robot Training Data
The bottleneck in physical AI training stopped being compute or model capability. It became the quality of the synthetic world.
Google's Gemini 3.5 Live Translate Removes the Pause
Translation that runs continuously, in the speaker's own voice, on a phone already in your pocket, moves the feature from something you open to something simply on.
China Wrote the First Mandatory Safety Standard for AI Agents
Safety moves from a feature you advertise to a gate you pass. The risk inventory sits at 13 categories and 97 items.
AI-Generated Content Now Has to Declare Itself
This step doesn't solve every problem, but it turns “AI-generated” from an option you could hide into a question you have to answer.