360 Panorama Maker for Qwen-Image 2.1 Edit 🌐
Got a nice photo of a place and wish you could look around? This LoRA turns one to three regular photos into a full 360° equirectangular panorama you can drop into a 360 viewer, a VR headset, or use as a skybox / environment map.
Your photo ends up right in the middle of the panorama, and the LoRA dreams up everything else around it: what's behind you, the sky above, the ground at your feet, all matching the place and the light of your photo.
How to use it
Load Qwen-Image 2.1 with this LoRA at strength 1.0.
Feed your photo(s) into TextEncodeQwenImage21 as
image_1,image_2,image_3. Setresolutionto 1088.Use this prompt, swapping in your own scene:
Transform this set of images into an equirectangular 360 panorama. Scene: A calm alpine lake below a snow-capped peak, evergreen forest, bright midday sun under a clear blue sky.⚠️ Important: sample from an Empty Latent Image at 1536×768 (or any 2:1 size, like 2048×1024). Don't use the latent that comes out of TextEncodeQwenImage21, because that one copies your photo's shape and you won't get a proper 2:1 panorama.
25 steps, CFG 1.0, euler / simple.
That's it! Open the result in any 360 viewer and look around.
Tips for great results
📸 Your best photo goes first.
image_1becomes the center of the panorama. Extra photos get placed around it automatically.🔄 Shoot from one spot. Stand still and turn. Photos taken from different places will confuse it.
📐 Normal photos work best. Regular phone or camera shots, held roughly level. Landscape, portrait or square are all fine. Skip fisheye shots, straight-up sky shots and already-stitched panoramas.
🏞️ More photos = more faithful. Anything your photos don't show gets invented. One photo works great, but 2-3 photos pointing in different directions get you closer to the real place.
✍️ Describe the whole place, not just your photo: where it is, the lighting (time of day, weather), and the mood. One sentence is plenty.
🎨 Turn off style or realism LoRAs. It was trained without them, and they can bend the panorama out of shape.
Good to know
The left and right edges aren't forced to line up perfectly, so you might spot a seam right behind you.
Big empty areas (behind you, straight up, straight down) are the model's best guess, not a reconstruction.
How it was trained
I took 232 free (CC0) 360° panoramas from Poly Haven covering streets, interiors, forests, beaches, deserts, mountains, ruins and more. From each one I cut out 1-3 random "photos" at different angles, zooms and shapes, just like you'd snap them with a phone. Then I taught the model to rebuild the whole panorama from those photos, with the first photo always in the center. That's 464 examples, 3000 steps, rank 32, trained with ai-toolkit.
About the showcase
Every panorama in the gallery was made from a single photo (shown right after it) with seed 7 and the settings above. The first five photos are free CC0 photos found on Openverse (from rawpixel and StockSnap). The grassy hilltop is a crop from one of the Poly Haven panoramas used in training.
Also on Hugging Face: Gogodr/qwen-image-2.1-edit-pano360-lora
Have fun exploring! 🧭
Description
First release. Trained on 232 CC0 Poly Haven panoramas, 1-3 views per sample. Use an empty 2:1 latent (1536x768) and resolution 1088 on TextEncodeQwenImage21.
FAQ
Comments (3)
wouldn't this be more like HDRI?
Usually HDRIs are in the same equirectangular format, but HDRIs main feature is the additional light, saturation, color information.
In traditional RGB you are constrained to a value from 0 to 255 for each main color (RGB)
In HDRI, all values are relative to each other and infinitely scalable, meaning one can be 1 and the other can be 50,000,000.
Inference can be done on a traditional image / png to convert it into an HDRI.
@Gogodr Ah! gotcha. Much appreciated for the explanation.