How to Turn ChatGPT-Generated Architecture Images Into a 30-Second Video

You asked ChatGPT for a house on a hillside, liked what came back, and now want it to move. ChatGPT gives you still images and nothing else: there is no video step after the image, so the film has to be made somewhere else. The hard part isn't the video tool. It's getting six views of the same building rather than six different buildings, exporting the files at full size, and saying plainly that the result is a concept. This guide covers each of those steps, the three ways to get from ChatGPT images to a 30-second vertical video, and which AI markers survive the trip.

Key takeaways

• ChatGPT Images 2.5 creates and edits still images; OpenAI closed the Sora app in April 2026 and its API in September 2026, so the video step happens in a separate tool.
• Generate all six views in one conversation as edits of your first image, with materials, light, camera height and one aspect ratio written down before you ask for a second view.
• Download each image with Save, never a screenshot: a screenshot drops the Content Credential and can degrade the invisible watermark OpenAI embeds.
• A video model animates pixels, not a 3D model, so it invents what lies past the frame edge, adds its own people, and can't give you a measured walkthrough.
• The image's own credential does not travel into a video made by another tool, so write the caption as if the file carried no label at all.

Why ChatGPT gives you images but not a video

ChatGPT can't turn your images into a video. As of October 2026, ChatGPT Images 2.5 creates and edits still images, and OpenAI's provenance help page lists video with a single note: "Sora is discontinued" (OpenAI Help Center). To get a film, download the images and bring them to a separate image-to-video tool, which is also where the decisions about length, format and labeling get made.

OpenAI announced ChatGPT Images 2.5 on September 8, 2026, available to all ChatGPT tiers, and says it "is better at preserving the subjects in your reference photos, and follows editing instructions more reliably across multiple turns" (Introducing ChatGPT Images 2.5). That second claim is what makes six matching views of one building possible, and the rest of this guide leans on it.

The video side closed this year: the Sora app on April 26, 2026, the API on September 24, 2026 (what the Sora shutdown means for studio archives). Older advice to "ask ChatGPT to animate it" no longer works.

Six views of one building, not six buildings

Every fresh prompt is a fresh building. Ask ChatGPT for "a timber house on a slope" six times and you'll get six timber houses with different roofs, window counts and cladding, and no video tool can stitch those into one project. The fix is to treat the first image as the reference and get every other view as an edit of it inside the same conversation. OpenAI's help page confirms you can "upload an existing image and describe the changes you want ChatGPT to make," and "generate images in any aspect ratio" through the aspect ratio picker, as of October 2026 (Images in ChatGPT).

A prompt scaffold that holds the building still

1. Generate the hero view first and refine it until the design is right. Don't move on until the roof, openings and materials are what you want; this image is the reference for everything else.
2. Write a building sheet in the next message. List what must stay fixed: materials by name (board-formed concrete base, charred cedar cladding, standing-seam roof), storeys and bays, window pattern, site, light, camera height (eye level) and one aspect ratio. Ask ChatGPT to hold all of it for every image that follows.
3. Request each new view as an edit of the first image. "Using the image above as the reference, show the same house from the garden side, same materials, same light, same camera height, same aspect ratio." Naming the reference every time matters more than any other phrase.
4. Make one change per message. Moving the camera and changing the time of day in one request doubles what the model has to reinterpret. OpenAI's help page also warns that "edits may extend beyond the area you selected," so look at the whole image after each edit.
5. Cover six roles, not six angles of the facade. Context, approach, the main space wide, a quieter secondary space, a material detail and a closing view (the six roles a project video needs). Ask for interiors last, once the exterior is locked, and name the window pattern you expect to see from inside.

A consistency checklist before you download

• Same number of storeys and bays in every exterior.
• Same roof form and overhang depth from every side.
• Same cladding, the same color temperature of light, shadows falling the same way.
• Interior windows match the openings visible from outside.
• One aspect ratio across all six images.
• No readable signage or house numbers; video models smear small lettering.

Fix problems here, where another edit is cheap. A video tool extends whatever it's given, including mistakes.

Export the right files

Download every image with the Save control, which OpenAI's help page describes as the way "to download the image to your device"; don't screenshot it. Images generated in ChatGPT carry two provenance signals, Content Credentials (C2PA metadata) and a SynthID watermark, and OpenAI says the credential "can be lost through everyday actions, such as taking a screenshot or converting a file" (OpenAI Help Center). A screenshot also captures the screen's resolution, not the file's.

Check the pixel size of what you downloaded. The ChatGPT help page doesn't publish output dimensions, but OpenAI's API image guide lists 1024×1024, 1536×1024 and 1024×1536 as recommended sizes for the same model family, custom sizes up to a 3840-pixel long edge at aspect ratios between 1:3 and 3:1, and PNG, JPEG and WebP as output formats, as of October 2026 (OpenAI image generation guide). A 1024×1536 portrait is workable source material for a vertical film; a copy that went through a messaging app is not.

For a 9:16 film, choose a portrait or square ratio in the picker and hold it across all six; a wide image still works, but it gets recomposed into a vertical frame and anything pinned near an edge can drop out. Keep the six originals in a project folder with a link to the conversation: OpenAI states that "provenance signals do not include details about the user, organization, or prompt," so the file will never tell you what you typed.

What image-to-video keeps and changes in a ChatGPT image

An image-to-video model has no geometry behind the picture. It reads pixels, infers depth and extends what it sees; it hasn't read a plan. A ChatGPT image is the same kind of input as a rendering or photograph, except that some of its details were already invented once, and the video model will extend those too.

• Motion past the frame edge is guesswork. A slow push-in stays inside what the image shows; a pan, pull-back or orbit asks the model to invent the rest of the street or the next room. Which camera moves keep lines straight ranks them if you're directing the camera yourself.
• People are added, not animated. Arch2Video films always include a small cast of original, fictional people in at least one scene, and people already in your images aren't animated; most clip tools render their own figures too. Leave figures out of the ChatGPT images rather than fighting over them.
• There's no measured walkthrough. A video from six images is a sequence of views, not a route through the plan. A floor plan can be one of the six for Arch2Video, but it's treated as a document and given a slow pan, and the AI director decides where it sits.

Review the result before anyone else sees it, and present it as concept work: a building that exists as six ChatGPT images is not a building.

Three routes from six images to a video

The clip route gives you the most control and costs the most time: a prompt per image, re-renders for the clips that bend, then an edit with music you've licensed. The finished-film route trades that control away: Arch2Video takes the six images (JPG, PNG or WebP, full resolution is fine), plans the order itself, which may differ from your upload order, and delivers the MP4 with a choice of soundtrack (Still Air, Quiet Piano, Concrete, or none) that you can switch afterward and re-export without using another video credit. A commissioned animation is the right call when a client needs a route through the building or a fixed shot list; for a concept that exists only as ChatGPT images, it's usually more than the project warrants.

Metadata and labels: what survives, and what to write

Two signals ride inside a ChatGPT image. Content Credentials are metadata and carry detail but are easy to lose; SynthID is embedded in the pixels and, in OpenAI's words, is "designed to withstand common changes" such as "light edits, copying and pasting, cropping, or screenshots," while "heavy image cropping, file compression" and similar changes "can make a watermark undetectable" (OpenAI Help Center). Neither one is carried into a video, because the video is a new file with new pixels.

Check any downloaded image at openai.com/verify, which accepts PNG, JPG and WebP and reports whether it finds C2PA metadata, a SynthID watermark, or neither, as of October 2026. Check the finished MP4 once per video tool in the Content Credentials Verify tool; what a content credential is and who reads it explains the result. Know what your export carries before you rely on the file to do any disclosing for you.

Platform rules don't wait for a credential. YouTube requires disclosure when AI "generates a realistic scene that didn't actually occur," and says it "may automatically apply an AI label" to content carrying C2PA metadata (YouTube Help). Meta will "require people to use this disclosure and label tool when they post organic content with a photorealistic video or realistic-sounding audio that was digitally created or altered" (Meta). LinkedIn shows a C2PA icon on signed image and video content and notes that "it's not yet possible to identify and label all AI-generated and modified content" (LinkedIn Help). All three as of October 2026.

Write the caption as if nothing in the file will speak for you: "AI-generated concept study. Images created in ChatGPT and animated with an AI video tool. Not a built project." OpenAI itself notes that watermarks and C2PA metadata "do not replace visible labels, banners, or other notices that may be required." Disclosure wording you can copy has versions for built and unbuilt work, and if you plan to boost the post, Meta's AI info label on ads covers what a paid placement adds.

Frequently asked questions

Can ChatGPT make a video from my images?

No. As of October 2026, ChatGPT Images 2.5 generates and edits still images, and OpenAI's help page lists video only with the note "Sora is discontinued"; the Sora app closed in April 2026 and the API in September 2026. Download the images and use a separate image-to-video tool: a clip tool, a six-image service that returns a finished film, or an animator.

Why does the building change between my ChatGPT images?

Because each new prompt is a new generation unless you anchor it. Keep all six views in one conversation, name the first image as the reference in every request, write the materials, storey count, light, camera height and aspect ratio into a building sheet the model must keep fixed, and change one thing per message. OpenAI says Images 2.5 follows edits "more reliably across multiple turns," but it only holds what you've told it to hold.

Can I mix ChatGPT images with real site photos in one video?

Yes, if they belong to the same project, and it's a sensible way to show an unbuilt addition in its real context. Match the light and camera height between the photo and the generated views, or the cut will announce itself. Arch2Video accepts any mix of renderings, AI concept images and photos from one project, and anyone in the photos is replaced by the fictional cast rather than animated. The caption then has to say which parts are concept: "existing building photographed, proposed addition AI-generated" is the kind of line to write.

Does the finished video keep the AI label?

Not from the image. The Content Credential and SynthID watermark in a ChatGPT image stay with that image; the video is a new file, and whether it carries its own credential depends on the tool that made it, so check the exported MP4 in a verification tool before you post. YouTube may add a label automatically when it detects AI content or C2PA data, Meta requires its disclosure tool for photorealistic AI video, and LinkedIn only shows an icon when the file is signed. Write the disclosure yourself every time, and any platform label will agree with you.