Day to night exterior photo: GPT Image 2.5 or Ideogram 4.5?
Turn a daytime building photo into twilight: Ideogram 4.5 keeps the source shape at $0.075; GPT Image 2.5 needs aspect_ratio auto, $0.0835 at high.

To turn a daytime exterior photo into a twilight or night shot with an API, send the photo as the only reference to either ideogram/ideogram-v4.5 (omit aspect_ratio to keep the source shape) or openai/gpt-image-2.5 (set aspect_ratio: "auto"). On Sume an Ideogram edit at medium costs $0.075 and a GPT Image 2.5 edit at high costs $0.0835 for a 1024x1024 image with one reference.
Both can do it. The differences are the parameters that keep the frame stable, the price tiers you can pick between, and what you can add, like a mask for the sky or windows.
Keeping the frame the same
The geometry rule differs between the two. For Ideogram 4.5 on Sume, an edit without aspect_ratio keeps the source geometry, and a resolution of 2K on its own does not change that. For ChatGPT Image 2.5 the docs say to use aspect_ratio: "auto" on edits to match the reference, and that omitting the field is not the same as auto.
For real-estate and architecture the frame is not negotiable: a building with a moved roofline is a different building. Whichever model you use, compare the result with the source for edges, window positions and roof angles. Overlay the two images at 50 percent opacity in any image tool: if the outline of the building doubles, the edit moved something, and that image goes back for a rerun at a higher tier. Check the output pixel size too, since a result that is a different size from the source is a sign the frame was not held.
The prompt
A good twilight prompt names the light you want and the things that must stay: deep blue sky with a faint orange horizon, interior lights on in every window, warm light on the front door, same camera angle, same building, same landscaping. Ask for lit windows explicitly, because a dark night photo of a house with black windows does not sell anything.
curl -X POST https://api.sume.com/v1/images \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "ideogram/ideogram-v4.5",
"prompt": "Convert this daytime photo to twilight: deep blue sky with a faint orange horizon, warm interior light in every window. Same building, camera angle and landscaping.",
"input_references": [
{"type": "image_url", "image_url": {"url": "https://example.com/house-day.jpg"}}
],
"quality": "medium"
}'Cost for 12 listings
GPT Image 2.5 amounts are the billed prices in Sume's pricing fixtures for one 1024x1024 edit with one input image. Ideogram amounts are Sume's per-image price.
| Model and quality | Per image | 12 images |
|---|---|---|
| Ideogram 4.5, low | $0.0375 | $0.45 |
| Ideogram 4.5, medium | $0.075 | $0.90 |
| Ideogram 4.5, high | $0.275 | $3.30 |
| GPT Image 2.5, medium | $0.021 | $0.252 |
| GPT Image 2.5, high | $0.0835 | $1.002 |
Details that sell a twilight photo
The best twilight shots are not dark. They keep enough ambient light that the facade still reads, with a sky that is blue rather than black and lit windows that are warm rather than white. Ask for dusk, not night, unless you really want a dark scene. Mention the ground: wet driveways reflect lights and look premium, but the model may add puddles you did not ask for, so say whether the ground is dry.
Watch the sky and the glass. Skies are where models invent stars, moons and clouds, and window glass is where they invent reflections. If you see an invented moon, rerun with a prompt that says no moon, no stars, and a plain gradient sky. Compare every output with the daytime source so lamp posts, power lines and parked cars do not appear or disappear.
When GPT Image 2.5 is the better pick
Pick GPT Image 2.5 when you want a mask. You can mask only the sky and the window glass and leave the brick, roof and landscaping out of the model's reach, which protects the parts a buyer inspects. Its medium tier is also cheaper than Ideogram's medium. Pick Ideogram 4.5 when the frame must not move and you want a no-mask edit with a fixed price per image.
Sources
Related posts
More in Comparisons
- Decagon says Chord is 90% indistinguishable: run your own blind test
Decagon cites about 90% indistinguishability across three voice samples. How to run a fair blind listening test on any TTS output, with a cost estimate.
- Decagon's Oct 1 launches: which of the four matter to a video team?
Voice 3, Personal Agent Gateway, Agent Modules and Duet Apprentice launched together. Only one touches voice; here is what each does and who needs it.
- Decagon Voice 3 handles 70+ languages mid-call. A TTS job takes one
Voice 3 detects language and switches mid-sentence. A Sume TTS job speaks one language per request. How to plan a multilingual script around that.
- Decagon Voice 3 duplex agent or a voiceover job: which do you need?
Voice 3 listens while it speaks. A voiceover job does not. How to choose between a live voice agent and a file-based TTS job for your project.
Written by Sume