AI Panorama Guide: How to Make 360° Images and Turn Them Into Video

A venue manager asked me for a "360 video tour" of a converted warehouse space. Budget: a few hundred dollars. Timeline: four days. No camera crew, no gimbal, no stitching software licence.

What I actually delivered was two things that get sold under the same label: a set of 2:1 equirectangular panoramas you can drag around in a viewer, and a short flat video that pans across one of them. She was happy. But the gap between "a 360 image" and "360 video" is where most AI panorama projects go wrong, and it is worth being precise about it before you spend a weekend generating the wrong asset.

This guide covers all three things "AI panorama" can mean, how to generate a clean equirectangular image, how to fix the seam and the poles, and the three routes for turning a still panorama into motion — including the one that actually produces something you can publish.

TL;DR

• "AI panorama" means three different assets: a 2:1 equirectangular still for a 360 viewer, a cubemap or skybox for a 3D scene, and true 360 video. Most AI tools can do the first, some can do the second, and none of the mainstream video models do the third properly.
• The format is 2:1, not 16:9. Equirectangular maps a full sphere onto a flat rectangle. A 16:9 image will not wrap correctly in any 360 viewer, no matter how wide it looks.
• Seams and poles are the whole quality problem. The left and right edges have to match, and the top and bottom get stretched into pinwheels. Fixing both is a five-minute compositing job, not a regeneration.
• Prompt for a scene, not a subject. A panorama needs coverage in every direction. Describe the room, the light and what is behind the viewer, and put the horizon at the vertical centre.
• You cannot turn a still panorama into real 360 video with a video model. You can produce a convincing flat pan, a re-framed set of angles, or a 2.5D parallax move. Those are three different deliverables and clients need to be told which one they are buying.
• The cheapest credible workflow: generate the still free, fix the seam, animate a pan, publish in a viewer that accepts 2:1. Total cost is minutes, not a licence.

First, Decide Which "Panorama" You Need

This is the step people skip, and it costs them the whole project. The three formats are not interchangeable, and each one has a different downstream tool.

Format	Shape	What it is for	Can AI make it?
Equirectangular still	2:1 rectangle (e.g. 4096×2048)	360° viewers, virtual tours, Google Street View-style embeds	Yes — the default output of most image models when asked
Cubemap / skybox	6 square faces, or 1:4:1 cross	Game engines, 3D scenes, HDRI lighting	Yes, with conversion from equirectangular
True 360 video	Equirectangular video, 2:1 per frame	VR headsets, head-tracked playback, walkable tours	Not from mainstream video models

The distinction that matters commercially: an equirectangular still plus a viewer feels like a 360 experience to 90% of web visitors. They drag, it moves, they feel inside the room. True 360 video adds head tracking and parallax as you move — which is a different product with a different budget, and it is not something a text-to-video model will hand you.

If the client said "360 tour" and meant "something I can embed on my site", you need the still. If they said "VR", you need to have a harder conversation.

Why the Format Is 2:1 and Why That Breaks Things

An equirectangular image is a sphere unwrapped onto a rectangle. Longitude maps to the horizontal axis, latitude to the vertical. Because a full sphere is 360° horizontally and 180° vertically, the correct aspect ratio is 2:1.

Three consequences follow, and all three cause visible defects:

1. The left and right edges are the same place. They are the seam where longitude wraps from 360° back to 0°. If they do not match pixel-for-pixel, you get a vertical discontinuity — the single most obvious tell of an AI panorama.
2. The top and bottom are single points. The zenith and nadir get stretched into a full row of pixels, producing a pinwheel or smear. Models handle this badly because it is unlike anything in their training distribution.
3. Straight lines bend. A doorway near the horizon is fine; the same doorway near the top or bottom curves dramatically. This is correct projection behaviour, not a model error — but it looks like one to a client.

Once you know these three, the fix list writes itself, and it is much shorter than "regenerate until it looks right".

If you want output today, start here: Launch Wan 2.2 Now →

Step 1: Prompt for Coverage, Not for a Hero Shot

Most people prompt a panorama the way they prompt a normal image: a subject, a style, a lens. That produces a beautiful frame with dead space at the edges, which wraps into a viewer as a wall of blur.

A panorama prompt needs to describe the whole environment and where the viewer stands in it. The pattern I use:

360 degree equirectangular panorama of [space], camera at eye level in the centre of the room,
horizon line at the vertical centre of the image, [light sources and direction],
[what is at each cardinal direction], seamless left and right edges, no foreground subject,
photorealistic, consistent scale

Four elements do the heavy lifting:

• "360 degree equirectangular panorama" — the format instruction. Without it you get a wide-angle photo, which will not wrap.
• "horizon line at the vertical centre" — anchors the projection. Without it the horizon drifts and the floor/ceiling proportions look wrong in a viewer.
• "what is at each cardinal direction" — forces coverage. "Behind the viewer: a steel roller door and a fire exit" is the phrase that stops the model from leaving a blank wall.
• "no foreground subject" — a person standing two metres from the camera becomes a stretched monster at the poles.

Two things to avoid: "wide angle lens" (pushes the model toward 16:9 distortion) and any mention of a specific focal length (the projection is spherical; focal length language confuses it).

You can generate the still with a free image model — the Z-Image generator handles interior scenes well and runs without setup, and if you would rather start from a photo you already have, image-to-image works for restyling an existing equirectangular shot.

Step 2: Fix the Seam Before You Judge the Image

Here is the part that separates a usable panorama from a demo. A raw AI panorama will almost always have a seam mismatch of some kind. You do not need to regenerate — you need ten minutes in any editor with a clone or offset tool.

The offset test. Roll the image horizontally by half its width. Now the old left and right edges sit in the middle of your screen as a vertical line. Everything wrong with your seam is now visible at once. Fix it, roll back, save.

The pole patch. At the top and bottom, replace the smeared 5% with a flat colour sampled from the surrounding area, or a soft radial blur. Nobody looks at the nadir in a viewer, and a clean gradient reads as "floor" to the eye. This is far better than leaving a pinwheel.

The horizon check. Draw a horizontal line at exactly 50% height and look at whether real-world horizontals (table edges, window sills, wall tops) sit near it. If they fan upward or downward, the projection is off and the viewer will feel like the room is tilted.

The scale check. A doorway at 0° should be roughly the same height as a doorway at 180°. If one is half the size, the model treated the panorama as two pictures side by side and the space will feel wrong the moment someone drags.

A checklist you can run in five minutes:

Check	Pass condition	Fix if it fails
Seam continuity	No vertical discontinuity after 50% roll	Clone/heal across the join
Poles	No pinwheel or radial smear in top/bottom 5%	Patch with sampled colour or radial blur
Horizon	Straight line at 50%, real horizontals near it	Regenerate with horizon instruction
Scale consistency	Matching object heights across 0°/180°	Regenerate; this one is not fixable
Content coverage	No large empty/blurred region	Add "behind the viewer" detail to prompt

Step 3: Turn the Still Into Motion (Three Routes)

Now the part that gets oversold. You have a 2:1 equirectangular still. What can you actually do with it?

Route A: Image-to-Video Pan (fastest, flattest)

Feed the panorama into an image-to-video model and prompt a camera move: slow horizontal pan, slight push-in, gentle tilt. The model does not know it is looking at a sphere — it treats the image as a flat frame and animates within it.

What you get is a flat video that reads as a pan, not a 360 video. In a standard player, that is often exactly what the client wanted: a 10-second movement across a room for a listing page or a social post.

Two practical rules:

1. Crop before you animate. A full 2:1 panorama animated as a flat frame gives you a very wide, very small subject. Crop to 16:9 from the region you want to move through, then animate.
2. Keep the move slow. Fast pans on a still panorama expose the resolution drop and the seam. A 5-second, 15-degree move looks like camera work; a 90-degree whip looks like a bug.

Free tool for this: image to video.

Route B: Re-Frame With Camera Angle Control (best for sets of stills)

If you need multiple angles of the same space rather than a moving shot, angle control is a better fit than video. Give the model the panorama plus a direction instruction and you get a coherent re-framed still — useful for a gallery, a floor-plan overlay, or a listing that wants "living room / kitchen / mezzanine" from one source image.

The limit is honest: this produces plausible alternate views, not views that stay geometrically consistent with a real space. For architectural accuracy, use a real camera. For mood and layout suggestion, it is fast and good enough.

The Qwen camera control tool is the entry point here, and there is a full walkthrough in the camera angle guide.

Route C: 2.5D Parallax (most convincing, most work)

The trick real-estate video people have used for years: take the panorama, cut it into a foreground layer and a background layer, and move them at different speeds. Add a subtle zoom and you get depth without any model needing to understand the space.

With an AI panorama this is easier than with a photo, because you likely generated it with a clean depth separation — flat floor, mid-ground furniture, distant wall. Cut along those lines, offset the layers over 5–8 seconds, and the result reads as a camera move through a room rather than a photograph sliding.

It takes an editor and maybe twenty minutes. It is also the only one of the three routes that a client will describe as "cinematic".

What You Cannot Do (And Should Say So)

Three claims are common in AI panorama marketing and none of them survive contact with a project:

• "Turn any image into 360 video." No mainstream video model generates equirectangular output. You get flat video from a flat crop of a spherical image. It looks like a pan; it is not navigable.
• "Seamless 360 out of the box." Seam quality is a compositing task. Models produce a seam; you fix it. Any workflow that skips this step produces the vertical-line artifact.
• "Walk through the space." Walking requires multiple viewpoints and geometric consistency between them. A single panorama — however clean — contains exactly one viewpoint. A multi-node tour needs several panoramas, each generated or shot separately, linked by hotspots.

Being clear about these three up front is the difference between a happy client and a revision cycle. For a closer look at how it stacks up against other models, see Kling 2.6 Motion Control vs Wan 2.2 Animate.

Failure Map

What you see	Cause	Fix
Vertical line in the viewer	Seam mismatch at 0°/360°	Roll 50%, clone across the join
Swirl at top or bottom	Pole stretch	Patch with sampled colour
Everything looks tilted	Horizon not at centre	Regenerate with horizon instruction
Room feels "too big" or "too small"	Scale inconsistency between hemispheres	Regenerate; check object heights at 0°/180°
Big blurry region	Prompt described only one direction	Add cardinal-direction detail
Video pan looks like a slide	Move too fast, no foreground motion	Slow to 15° over 5s, or use parallax
Text and signage scrambled	Model limitation on small text	Replace signage in post

The Cost Side, Honestly

The reason this workflow is worth learning is that the traditional route is expensive in a specific way: a 360 camera plus stitching software plus a person who knows how to use both. The AI route trades that for time and a handful of generations.

Rough shape of it:

• Stills: free tier generations cover a handful of attempts; expect to burn 3–5 per usable panorama because of the seam lottery.
• Re-framing: a few more generations per alternative angle.
• Video pan: one 5-second generation per attempt at the platform's per-second rate; budget 2–3 attempts per usable clip.
• Post: the seam fix and any parallax work is editor time, not licence cost.

The trap is treating generations as free and time as free. Three attempts at a panorama plus a seam fix is twenty minutes. If you are doing this for a client, price the twenty minutes.

If you want the motion side on a model with stronger camera control, Wan 2.7 image to video is the option I use for pans that need to hold an interior steady without warping the furniture.

The Bottom Line

AI panorama is a format problem before it is a generation problem. Get the 2:1 projection right, fix the seam and the poles, and a free image model will give you a usable 360 still. Then decide what "motion" actually means for your client: a flat pan, a set of re-framed angles, or a parallax move.

What it will not give you is true 360 video, and knowing that before the kickoff meeting is worth more than any prompt template in this article.

Try the Whole Panorama Workflow Free

You can build the still, fix the seam and animate the pan without paying for a camera, a licence or a subscription. The 360 still is the expensive part in the traditional workflow; here it costs a generation and a clone-stamp pass.

• Free image generation for the equirectangular still — no setup, no card
• Free image-to-video to animate a slow pan out of your crop
• Camera angle control when you need a gallery of re-framed views from one source image
• Wan-family video models on the same account when the pan needs to hold an interior steady
• No watermark, no export gate — download the clip and publish it in your own viewer

Start with the free Z-Image generator for the still, then run it through image to video for the move. Two tools, twenty minutes, one deliverable.

Related guides

• Kling 2.6 Motion Control vs Wan 2.2 Animate: AI Motion Generation Comparison
• GLM-Image vs Z-Image: Next-Gen AI Image Generators Compared
• GPT Image 2 vs Free AI Image Generators: Complete Comparison Guide for 2026

FAQ

What is an AI panorama?

An AI panorama is a wide, wrap-around image generated by an AI model rather than captured with a 360 camera. In practice the useful output is an equirectangular image with a 2:1 aspect ratio, which any 360 viewer can map onto a sphere so visitors can drag to look around.

What size should an AI panorama be?

Keep the 2:1 ratio. Common sizes are 2048×1024 for web weight, 4096×2048 for crisp detail on desktop, and 8192×4096 for headsets. Resolution matters more than usual here because the viewer magnifies small segments of the image — a 2048-wide panorama looks soft the moment someone zooms.

Can I turn an AI panorama into 360 video?

Not true 360 video. Mainstream video models do not output equirectangular frames, so animating a panorama produces a flat clip that reads as a camera pan. That is usually enough for a web embed or social clip; genuine 360 video requires a 360 camera or a multi-viewpoint capture.

How do I fix the seam in an AI panorama?

Roll the image horizontally by half its width so the left and right edges meet in the middle of your canvas, then clone or heal across the join. Roll it back and save. This reveals every seam defect in one view and takes a few minutes in any editor.

Why does the top of my panorama look stretched?

Because equirectangular projection compresses an entire pole into one row of pixels. Patch the top and bottom 5% with a colour sampled from the surrounding area, or apply a soft radial blur. Nobody inspects the nadir in a viewer, and a clean gradient beats a pinwheel every time.

Is AI panorama good enough for real estate listings?

For mood, layout suggestion and social previews, yes. For accuracy — room dimensions, exact finishes, window positions — no. AI panoramas reimagine details, so treat them as marketing imagery rather than a representation of the property, and check whether the listing platform requires real photography.

What is the difference between equirectangular, cubemap and HDRI?

Equirectangular is a 2:1 rectangle wrapping a sphere. A cubemap splits the same sphere into six square faces and is what game engines usually consume. HDRI is a high-dynamic-range equirectangular image used for lighting, not just viewing — useful if you are placing 3D objects into the scene rather than showing a tour.

Related Guides

• How to change the camera angle of a photo with AI
• Camera angle control in AI video: the full workflow
• Best image-to-video AI tools for turning stills into motion

References

• What is an equirectangular image? (Panorama AI) — 2:1 projection, distortion behaviour and viewer mapping
• How to write prompts for AI 360° panoramas — horizon, seam and pole prompting patterns
• Equirectangular skybox vs cubemap vs HDRI — format selection for 3D and lighting use
• ThingLink — crafting prompts for 360-degree images — descriptive coverage in panorama prompts
• VirtualTourEasy — AI 360° panorama generator — 2:1 equirectangular output and tour linking workflow