Adobe currently documents the Firefly model at 1024×1024 output when returned to Photoshop desktop and web. That limit explains why a large Firefly fill can look soft after it is fitted into a much larger selection. It does not establish the output of every partner model in Photoshop. For a deliverable that needs a specified 2K or 4K result, use a model and route that expose supported output dimensions, then composite and inspect the result.
The source is Adobe’s current Photoshop generative AI FAQ, reviewed July 14, 2026. The distinction between Firefly, partner models, generation output, fill output, and placed canvas size matters throughout this guide.
Why do Firefly fills look soft on a large selection?
A 1024×1024 generation contains about one million pixels. If Photoshop fits that result into a 3000×2000 selection, the placed area needs six million pixels. Resampling can distribute the generated pixels across the larger area, but it cannot recover missing generated detail that was absent from the result.
The visible mismatch is usually texture rather than composition. A generated wall, fabric, or product surface can read correctly from a distance while resolving less grain and edge detail than the captured pixels beside it. Adobe’s resizing and resampling guide makes the PPI distinction explicit: changing PPI without resampling does not change pixel dimensions or pixel count. With resampling enabled, changing PPI changes pixel dimensions because Photoshop adds or removes pixels, but that operation cannot recover missing generated detail. Selection dimensions and the actual generated output still govern what detail was supplied.
This is an imaging explanation, not a claim that every soft region has the same cause. Feathering, compression, depth of field, and a low-detail prompt can also affect the result. Compare the fill layer with nearby source detail before choosing a remedy.
What is the documented output limit?
Adobe’s FAQ says the Firefly model currently returns 1024×1024 output to Photoshop on desktop and web. The exact scope is important: the statement is evidence for the documented Firefly path. It does not establish one 1024 rule for every partner model exposed by the model selector.
Adobe separately describes Firefly Image 5 as producing native 4 MP generation without upscaling. That is a generation capability; it does not establish that every Photoshop fill uses Image 5 output or returns 4 MP.
The current model-selector workflow shows a selection, model, prompt, optional reference, and Generate action. Adobe’s documented Generative Fill workflow does not expose a manual output resolution control.
| Adobe route | Documented output | Credit behavior reviewed July 14, 2026 | Evidence boundary |
|---|---|---|---|
| Firefly model in Photoshop desktop and web | 1024×1024 | 1 credit per standard Adobe Firefly Generative Fill | Firefly path only |
| Firefly Image 5 | Native 4 MP generation without upscaling | 10 credits per generation | Generation capability, not a promise for every fill |
| Gemini 3.1 with Nano Banana 2 in Adobe | 1K or 2K on Adobe’s rate card | 20 credits at 1K; 40 credits at 2K; limited-time rates | Photoshop Generative Fill uses 10-40 credits by output size |
The Firefly credit entries come from Adobe’s generative credits FAQ. The partner values and limited-time markers come from Adobe’s partner-model rate card. Rates can change, so the date belongs beside the table.
Which models are currently selectable in Generative Fill?
Adobe’s June 18, 2026 model-picker guide separates Adobe models from partner models.
- Adobe models: Firefly Image 5, Firefly Fill & Expand, and Firefly Image 1.
- Partner models: Gemini 3.1 with Nano Banana 2, Gemini 3 with Nano Banana Pro, Gemini 2.5 with Nano Banana, FLUX.2 pro, and FLUX.1 Kontext [pro].
Partner models use a premium-credit path. Adobe’s partner-model page lists model-specific rates and says Photoshop Generative Fill consumes 10 to 40 credits depending on output size. It also labels several values as limited-time rates. A partner model selected inside Photoshop is therefore not the same billing route as calling that provider’s API directly.
Can Generative Upscale repair a soft fill?
It can produce a larger result, but it does not recover ground-truth camera detail that the source fill lacked. A generative upscaler can synthesize plausible new detail. That can be useful, yet it creates a new inspection surface for labels, texture, faces, repeated patterns, and edges that were already approved.
Adobe’s Generative Upscale guide documents 2x or 4x output. It lists Firefly Upscaler up to 6,144×6,144, Topaz Gigapixel up to 56 MP, and Topaz Bloom up to 9 MP. Adobe marks the Topaz choices as requiring premium access.
Because the operation processes the image presented to it, inspect previously sharp regions after upscaling. For a soft local fill beside high-resolution source pixels, compare a local regeneration against a whole-image upscale before changing the rest of the document.
How should you work around softness inside Photoshop?
The first option is a smaller-generation workflow. Bielfor’s field practice is to divide a large background repair into overlapping selections whose long edge stays near the documented Firefly output. That is a workflow judgment, not an Adobe rule. It costs additional generations and requires seam inspection.
A practical pass looks like this:
- Duplicate the document or preserve the untouched source layer.
- Make an overlapping local selection and generate on a separate layer.
- Inspect seams, repeated objects, texture frequency, and color noise at final delivery size.
- Match grain only after the structural edit passes.
- Compare the composite with the original around every selection boundary.
Texture overlays and grain matching can hide a frequency mismatch, but they do not create the missing captured detail. Treat them as finishing tools, not evidence that the underlying generation has more resolution.
Which direct provider routes expose larger output?
Google and OpenAI document model-specific output controls. They do not share one generic meaning of 4K.
| Provider model | Documented selectable output | Model-specific boundary |
|---|---|---|
| Gemini 3.1 Flash Image | 0.5K, 1K, 2K, or 4K | Exact dimensions depend on aspect ratio |
| Gemini 3 Pro Image | 1K, 2K, or 4K | Exact dimensions depend on aspect ratio |
| GPT Image 2 | Flexible; longest edge 3,840 px; 655,360-8,294,400 total pixels | Edges are multiples of 16 and the aspect ratio is no wider than 3:1 |
Google documents the presets and model-specific dimensions in its Gemini image generation guide. Its live pricing page is the appropriate source for current output charges because input, thinking, and optional grounding can add to an image request.
OpenAI documents GPT Image 2’s flexible dimensions in the image generation guide. A 3,840×2,160 result is 4K-class, but it is not a literal 4,096-pixel square. The same guide warns that outputs above 2,560×1,440, or 3,686,400 total pixels, are experimental. OpenAI also prices output by token count, so requested size and aspect ratio can materially change output cost.
How should you choose a route?
Use stock Firefly for local repairs that fit its documented output and pass inspection. Test a Photoshop partner model when its reference or prompt behavior is useful and the plan’s premium credits fit the job. Use a direct provider route when the deliverable requires a selectable supported size and the production team can composite the returned image correctly. If the motivation is to stop consuming Firefly generative credits rather than to raise output size, Nano Banana in Photoshop without Firefly credits covers that route separately.
Bielfor builds Creator MAX, so any preference for a BYO-key panel is owner-authored workflow judgment rather than an Adobe, Google, or OpenAI claim. The current product page is the source for product availability and terms. A manual provider-download-place workflow can reach the same model outputs without that panel.
FAQ
Does increasing document PPI make Firefly Generative Fill sharper?
Not by itself. Changing PPI without resampling changes print dimensions, not pixel dimensions. With resampling enabled, changing PPI changes pixel dimensions as Photoshop adds or removes pixels, but it cannot recover detail missing from the generated output. The relevant quantities are the selection’s pixel dimensions, the model’s generated output, and how that output is placed on the canvas.
Is every current Photoshop fill limited to 1024×1024?
Adobe documents 1024×1024 for the Firefly model returned to Photoshop desktop and web. That source does not establish one output size for every partner model.
Does Firefly Image 5 make every fill 4 MP?
Adobe documents native 4 MP generation without upscaling for Firefly Image 5. That capability does not by itself establish the returned size of each Photoshop fill operation.
Is Generative Upscale useful after a fill?
It can be useful when a larger synthesized result is acceptable. Inspect exact text, repeated patterns, fine edges, and previously sharp source areas because an upscaler creates new detail rather than recovering captured ground truth.
Which route should I use for a specified 4K delivery?
Choose a route whose official model documentation exposes the required dimensions, generate at that supported setting, composite into the layered document, and inspect the full result at its final delivery size. For OpenAI, current 4K-class GPT Image 2 output is supported by the size limits but is experimental above the documented 2,560×1,440 pixel threshold.
