Using this LORA, you can generate images for VR with a 360-degree panorama. When actually loaded into VR, the effects are acceptable except for the low resolution. However, there is a serious issue that remains unresolved: the sense of unreality at the seams. Future updates will need to address this problem.
このLORAを使用すると、VR用の360度パノラマ画像を生成することができます。実際にVRに装載すると、解像度が低い以外は、効果はまずまず受け入れられます。しかし、解決できない深刻な問題があり、それは継ぎ目で生じる非現実感です。後続のアップデートでこの問題に対処する予定です。
Look like this
Description
I increased the training material to 80, maybe the effect of natural landscapes will be better, but the seam problem is still difficult to solve.
FAQ
Comments (16)
Great Job! Works really well.
Seams are actually pretty good compare to other 360 Loras I've tried.
I'd be very interested in what solution you use for seams. I crop and swap left and right side of the image and have tried Fooocus inpainting and SEGS detailers to try to generate a blended image to cover the seam (which is now in center of image). I still have never been happy with the result. Also, when I upscale the image I get a new seam. :-(
Next challenge -- stereoscopic 180 (or 360) for depth. Is that even possible?
I've done stereo images by animating the viewpoint moving sideways and using two images side by side. Maybe you could get two frames out of animating with this lora as well? I imagine the subject in the center is going to be the only area where the image holds up though.
UPSCALE EXPERIMENTS IN COMFYUI
Started with low rez (1024 x 512 & 512 x 256) for the composition. Upscaled it directly & it was a mess. So I added Controlnets to reinterpret the original generation as depth or normalized & color so it would keep it's main features.
However, the upscaler just greebled the thing into oblivion. Maybe a better upscale process? Any ideas?
maybe try
extras
Postprocess upscale by: 8, Postprocess upscaler: R-ESRGAN 4x+
目前效果非常差，希望博主能有更好的作品，辛苦！
stable diffusion VR bj's when? lol
Cool!!! Really looking forward
Hey, I'd like to train a stable cascade 360-degree LoRa. Do you have any tips for the dataset? Did you train the SDXL LoRa simply with 360-degree images, or is there something else to consider?
希望使用不少于100万张全景图进行训练谢谢
@8888 I found a page with resources, just need to manually change each image from HDR to PNG. https://polyhaven.com/
@denrakeiw
You can find resources on certain specific websites. For example, some industrial sharing or stargazing websites offer free image resources. Ideally, you should take the photos yourself or use graphic software to modify and create them. This can help avoid mismatches between the head and tail, which can lead to training failures
I don't see the point in offering a Lora that isn't perfect. I imagine that it must be quite complicated, but it can be used because the hardest part is generating images that loop seamlessly... Good luck anyway in trying to improve it!
It wont ever be "perfect" and ai in general will keep improving.
CONGRATULAIONS!!!: This is truly groundbreaking. Skybox Ai does basically the same thing, but what you are doing will give access to unlimited models and environments. KEEP GOING! MAKE IT A REALITY and YOU get a lot of BUZZ from me! Literally! VR-VR-VR mates with AI is going to be great if done right.
thank you for your support