Figma Weave caps GPT Image 2 at 1K. Replicate caps it at 1K. WaveSpeed caps it at 1K.
fal.ai, Flora, Freepik, and Higgsfield all support 2K and 4K on the same model, pulling from the same OpenAI API.
Same model. Different infrastructure.
| Tool | Resolutions | Access |
|---|---|---|
| fal.ai | 1K / 2K / 4K | API |
| Flora | 1K / 2K / 4K | UI |
| Freepik | 1K / 2K / 4K | UI |
| Higgsfield | 1K / 2K / 4K | UI |
| Adobe Firefly | 1K | UI |
| Figma Weave | 1K | UI |
| Replicate | 1K | API |
| WaveSpeed AI | 1K | API |
Eight platforms added GPT Image 2. Four offer the full resolution range. Four cap at 1K. The model is identical in all of them. So what's different?
The aggregator problem
Replicate and WaveSpeed are pass-through aggregators. You send a request, they route it to the upstream API, you get a result. The value is access to hundreds of models in one place. The constraint is you get whatever parameters the wrapper was built to expose, and nothing more.
Replicate's gpt-image-2 implementation tops out at 1K right now. It's not a permanent ceiling. The integration will update. But today the wrapper doesn't expose the full flexible sizing the model supports.
Figma Weave is a visual canvas that aggregates from multiple backends, including Replicate, fal.ai, and others. For its OpenAI models, Weave almost certainly routes through Replicate. If that's correct, you're getting Replicate's 1K ceiling, not OpenAI's.
fal.ai built the other way. They host inference directly, which means they control the full API surface they expose. They offer gpt-image-2 with the complete flexible sizing, from 1K up to around 4000px per edge. Flora, Freepik, and Higgsfield each built direct integrations and got the same range.
Adobe Firefly fits the same pattern. Firefly is a platform that hosts both Adobe's own model and third-party models including gpt-image-2. Their gpt-image-2 integration caps at 1K today. Whether that's a backend routing decision or a deliberate tier choice, the result is the same.
The pattern is straightforward. If the tool sits between you and the model, you get what the middleman built. If the tool talks to OpenAI directly, you get the whole thing.
Before you start using 4K outputs
A detail worth knowing: OpenAI's own documentation labels gpt-image-2 outputs above 2K as experimental, with "mixed results at those sizes." The practical working range is 1K to 2K. That's already an improvement over gpt-image-1.5, which topped out at 1536px natively.
More important: gpt-image-2 at 4K and NanoBanana Pro at 4K are not the same thing.
NanoBanana Pro (Gemini 3 Pro Image from Google DeepMind) has been generating natively at 4096x4096 since launch. Over a billion images. A year of actual production use. It's available at full resolution through Freepik, the Vertex AI API, and others. When you get a NanoBanana Pro 4K image, you're getting 16 megapixels of native output from a model that's been battle-tested at that size.
When you get a gpt-image-2 4K image right now, you're getting an experimental output from a model that's been live for less than 48 hours at that resolution, from a provider OpenAI themselves flags as producing mixed results.
Flora ran gpt-image-2 against NanoBanana Pro across 13 dimensions before adding it to their platform. Their finding on image-to-image editing: NanoBanana Pro still wins. It makes surgical changes to what you ask for, gpt-image-2 tends to shift the overall scene instead of just the target. That matters for carousels, product shots, anything where you're editing one zone and need the rest to hold.
gpt-image-2 made genuine improvements in other areas. Text rendering is at 99% accuracy, up from the mid-80s. The edit fidelity on targeted zone changes is strong. The ComfyUI team specifically noted you can colorize a photo or shift lighting at up to 2K without the rest of the frame changing. It earned the top spot on LM Arena before it launched publicly.
But "4K available on gpt-image-2" and "4K on NanoBanana Pro" are not interchangeable claims. One is new and experimental. The other has a year of real production evidence behind it.
What to actually do with this
If you're on fal.ai, Freepik, or Flora: test gpt-image-2 at 2K on image edit tasks where text accuracy matters. That's where the model made the clearest jumps. The targeted edit story is worth testing against your own use case.
If you're on Replicate or Weave: you have gpt-image-2, just at 1K. I tested it on image expansion, asking the model to extend the frame on a product shot, and the detail loss is real. Textures go soft in a way you notice immediately. That's not a model problem. It's a resolution ceiling, and it will move when the integrations update.
And if someone shows you 4K results from gpt-image-2 and compares them against NanoBanana Pro: ask which tool, whether that's direct generation or an upscaled 1K, and how long the model has been running at that size. The number on the label is the starting point, not the answer.