How to put AI generated art on The Frame

An image generator and a Frame are an obvious pair: you can make exactly the picture you want for the wall, in the style you want, at the size you want, and change it whenever you feel like it. The catch is that almost no generator outputs 3840 by 2160, so there is a preparation step between the model and the TV that decides whether the result looks like art or like a stretched thumbnail.

Ask for 16:9 at generation time

Cropping a square image to 16:9 throws away a third of the picture and usually the third the model composed around. Set the ratio in the request instead.

Midjourney: add --ar 16:9 to the prompt.
Stable Diffusion, SDXL and Flux: set the dimensions to a 16:9 pair the model was trained near, such as 1344 by 768 for SDXL, then upscale. Generating far above the native resolution produces duplicated limbs and repeated horizons.
DALL-E and ChatGPT images: ask for landscape or widescreen in words; the API sizes are fixed, so you will still upscale.
Gemini and Nano Banana: ask for a 16:9 landscape image; state it in the prompt rather than relying on the default.
Ideogram, Leonardo, Firefly: all have an aspect ratio control in the interface. Use it.

The models that let you request more than 16:9, such as 21:9, are worth avoiding here. The Frame will matte anything wider.

Upscale to 3840 by 2160

Generators typically hand you something between 1024 and 2048 pixels wide. The panel wants 3840. Sending the smaller file works, and it looks soft, because the TV scales it up with a general purpose algorithm across a very large screen.

A dedicated upscaler does better, because it reconstructs plausible detail rather than smearing pixels:

Midjourney's own upscale buttons, then a second pass if needed.
Topaz Gigapixel, paid, still the most reliable for a large single image.
Real-ESRGAN, free and open source, runs locally, excellent on illustration and painting.
Automatic1111 or ComfyUI hires fix or img2img at low denoise, which is the best option if you generated the image locally and still have the seed.
Upscayl, a free desktop front end over Real-ESRGAN if you do not want a command line.

Upscale to a size at or above 3840 wide, then crop and resize to exactly 3840 by 2160 as the final step. Upscalers often output 2x or 4x of an odd source size, which lands nowhere near the exact number the TV wants. The reason that number has to be exact is in the size and format guide.

Prompt for a wall, not for a feed

Generated images that impress on a phone often fail at two metres. What survives on a large panel across a room is different.

Ask for a medium, not just a subject. "Oil on canvas, visible brushwork", "risograph print, two inks", "cyanotype", "gouache on cold press paper". Texture is what stops the image reading as a computer render.
Keep it simple. One subject, one idea, plenty of empty space. Busy compositions turn to visual noise on a wall.
Name a palette. A piece has to live with your room. "Muted ochre and slate", "warm greys, single red accent" gives you something that hangs rather than shouts.
Avoid text. Models still garble lettering, and garbled lettering is the one flaw everyone spots from the sofa.
Mind the middle. The centre of the panel is where a face or a horizon should sit. Generators often compose for the whole frame, and a 16:9 crop of a square moves the subject.
Grain hides the tell. A faint film grain or paper texture pass over the final image does more than any prompt keyword to stop it looking synthetic.

Generate variations of a prompt you like rather than a hundred different prompts. Four pieces from one idea make a set, and a set is what makes a rotating wall look intentional.

Getting a batch onto the TV

This is where AI art differs from every other kind: you do not make one image, you make forty and keep six. Uploading six is easy. Getting there means handling forty files, and the phone route is not built for it.

The efficient loop is on a computer:

Generate in 16:9.
Upscale the keepers.
Put them in a folder.
Send the folder to the TV.

TV Art Manager covers steps three and four. Drop a folder of generations into a collection, and each image is prepared to exactly 3840 by 2160 and sent edge to edge with the matte off, originals untouched. The free version sends one image at a time. Pro uploads a whole selection at once, keeps a collection mirrored on the TV so deleting a bad generation locally removes it from the wall too, and rotates the set daily.

That last part is what makes generated art worth the effort. A wall that shows a different piece from your own set each morning is the payoff, and it costs nothing per image once the pipeline exists.

On ownership and honesty

Two practical notes rather than a debate.

Check the terms of the tool you used if the art will hang anywhere commercial. Policies differ on what you own and what you may sell, most are permissive about personal display, and some are not, particularly on free tiers.

And if a guest asks whether you painted it, say a model made it. The Frame is very good at making a screen look like a painting, which means the question comes up more than you would expect.