Gemini app: The Gemini app has no size setting; Google’s own examples put the shape in the prompt (“…with 2:3 aspect ratio”). For an exact shape, add a blank image the shape you want and ask for your picture on it.
In the API it’s gemini-nano-banana-2.1, generally available since October 6, 2026. Write the size with a capital K, as in 2K: lowercase is rejected.
If you don’t set a shape, the model matches the size of an image you add, or otherwise makes a square.
1080 × 1920--ar 9:16 --hdScale to 1080 × 1920 if needed
1080 × 1350--ar 4:5 --hdScale to 1080 × 1350 if needed
1920 × 1080--ar 16:9 --hdScale to 1927 × 1080 (×0.662), then trim 3 px from the left and 4 px from the right
1280 × 720--ar 16:9Scale to 1285 × 720 (×0.882), then trim 2 px from the left and 3 px from the right
1080 × 1080--ar 1:1 --hdScale to 1080 × 1080 (×0.527)
Whole numbers only: --ar 139:100, not --ar 1.39:1. V8.2 has been the default since July 24, 2026.
Midjourney publishes sizes for these four shapes only, and says HD sizes in some shapes may vary slightly. The HD sizes are also what its upscalers make from a standard image.
BFL says the sizes it delivers are rounded to multiples of 16, so the sizes here are too. With reference images the limit is 4 MP.
Or set aspect_ratio (21:9 to 9:21) with resolution 768sq, 1k, 1.5k, 2k or 4k; a width and height override both. Width and height arrived on October 22, 2026.
Set it withaspectRatio: "SHAPE", resolution: "TIER"
Best setting for common sizes
1080 × 1920aspectRatio: "9:16", resolution: "1080p"Scale to 1080 × 1920 if needed
1080 × 1350aspectRatio: "9:16", resolution: "1080p"Trim 29.7% of the height, from the top and bottom, then scale to 1080 × 1350
1920 × 1080aspectRatio: "16:9", resolution: "1080p"Scale to 1920 × 1080 if needed
1280 × 720aspectRatio: "16:9", resolution: "720p"Scale to 1280 × 720 if needed
1080 × 1080aspectRatio: "16:9", resolution: "1080p"Trim 43.8% of the width, from the sides, then scale to 1080 × 1080
On Google Cloud (formerly Vertex AI) as veo-3.1-generate-001 and veo-3.1-fast-generate-001. Clips are 4, 6 or 8 seconds.
Google’s pages disagree about 4K: the model page lists it for veo-3.1-generate-001, but the API reference says 4K is for the preview models only. Check before you rely on it.
1080 × 1920 · Image to videoratio: "720:1280"Scale up to 1080 × 1920 (×1.5)
1080 × 1920 · Text to videoratio: "720:1280"Scale up to 1080 × 1920 (×1.5)
1080 × 1350 · Image to videoratio: "832:1104"Scale up to 1080 × 1433 (×1.3), then trim 41 px from the top and 42 px from the bottom
1080 × 1350 · Text to videoratio: "720:1280"Scale up to 1080 × 1920 (×1.5), then trim 285 px from the top and bottom
1920 × 1080 · Image to videoratio: "1280:720"Scale up to 1920 × 1080 (×1.5)
1920 × 1080 · Text to videoratio: "1280:720"Scale up to 1920 × 1080 (×1.5)
1280 × 720 · Image to videoratio: "1280:720"Nothing to do
1280 × 720 · Text to videoratio: "1280:720"Nothing to do
1080 × 1080 · Image to videoratio: "960:960"Scale up to 1080 × 1080 (×1.13)
1080 × 1080 · Text to videoratio: "1280:720"Scale up to 1920 × 1080 (×1.5), then trim 420 px from the left and right
The model is gen4.5, and Runway’s ratio is the output size in pixels. Runway announced it on December 1, 2025, and it came to the API on February 10, 2026.
6 shapes, each at 480p, 720p, 1080p and 4k: 16:9, 4:3, 1:1, 3:4, 9:16 and 21:9.
ratio: "SHAPE", resolution: "TIER"
How the best size is chosen
Each model makes only certain sizes, so the size you need usually isn’t one of them. For each model, the calculator picks the size that:
needs the least trimming to reach your shape (anything within 1% of the closest counts as just as close),
doesn’t need scaling up, because scaling up softens detail,
has a published pixel size, rather than an estimate,
and is the smallest of those, so you don’t pay or wait for pixels you’ll throw away.
To finish, scale the picture until it covers your size, then trim what’s left over, half from each side. For a 1080 × 1920 video frame, GPT Image 2.6 makes 1088 × 1920: trim 4 px from the left and right. Nano Banana 2.1 makes 1536 × 2752: scale to 1080 × 1935 (×0.703), then trim 7 px from the top and 8 px from the bottom.
When the size isn’t published
Adobe Firefly (web app), Gemini Omni Flash 1.2, Veo 3.1, Kling 4.0 and Luma Ray 3.2 publish their shapes and a resolution such as 1080p, but not the exact pixel size of each, and the same setting can give different sizes in different shapes. For these the calculator shows the setting and the shape, and gives any trimming as a share of the picture rather than guessing a size.
Apps with no size setting
The Gemini app has no size setting, so the shape goes in your prompt. For an exact shape there’s a better way: Google’s documentation says its image models keep the size of an image you add, so add a blank image in the shape you want to your chat and ask for your picture on it. Download a blank PNG at your size above; it’s made in your browser. ChatGPT has its own aspect-ratio picker.
Retired models
These have shut down, though older guides and AI answers still mention them:
Sora (OpenAI, video). Shut down: App and website: April 26, 2026; API: September 24, 2026. There is no direct replacement. Source.
gpt-image-1 (OpenAI, image). Shut down: API: October 23, 2026. Use GPT Image 2.6 instead. OpenAI names gpt-image-2.5-sunburst or gpt-image-2.5-flare as its replacement; 2.6 is newer. Source.
Veo in the Gemini API (Google, video). Shut down: Veo 3.0 and 2.0: June 30, 2026; Veo 3.1 previews: October 22, 2026. Use Gemini Omni Flash 1.2 instead. Veo 3.1 is still on Google Cloud. Source.
Sources
Every size above comes from the model maker’s own documentation, checked on . Makers change their models often, so this page is checked again regularly.