@LatentLabs Hi,If I used the tiling ,I'll get a completely disordered picture fragment T_T

@LatentLabs  Hi, Can you tell me how to activate asymmetric tiling on the X axis in Web UI

@LatentLabs Thank you very much,I know that ,The Activate asymmetric is a plug-in ^_^

A guess: It was the name of this lora during testing. EQP: EQuirectangular Projection perhaps. If I understand lora correctly the prompt has to call the lora. Notice this is the only lora called. Seems like they just changed the name.

u should provably still make one. It seems everyone attempting to get this to work somehow is still missing about 20% of the data in the center of the image so they end up with a huge visible seam.

One of the things i been trying to figure out is rather that attempt to deal with the seam to generate the whole texture as a 1:1 hemispherical skydome. But cant seem to be able to control the camera so it just looks directly up. If this could be achieve there would be no need for tiling since the horizon would be inside a circle within the 1:1 image.

@toontitan Like these?
I feel like for that you'd definitely need your own LoRA model, as the distortions are quite different.
Either that, or you output one of these, pop into blender and put it on a sphere as a texture, look up at the texture in orthographic view, and then render that out. It's the same process as getting those maps from equirectangular HDRI images.

For seams on equirectangular outputs, you can generate seamless images in SD with arbitrary coordinates (tiling on X OR Y, instead of X AND Y). Results work well enough, but get a bit iffy when you start upscaling them to proper resolutions with SD upscale.Once you offset the pixels so the seam is dead center of the image, content-aware fill in Photoshop usually fixes that issue no problem, or if that doesn't work, inpaint the seam and generate what's there.

Training your own LoRA on the hemispherical maps should be easy enough though, and might actually work better than these. Though I think detail around the extreme edges would get pretty squashed, and they won't generate that well since there's not a lot of pixel density there to do it. For cloud maps and stuff though I don't see that being much of an issue.

The top and bottom distortion issues seem to be related to LoRA architecture itself and not based on the amount of or quality of training data. That's why no one has fixed them yet.

I just released my own 360 LoRA here: https://civitai.com/models/26815/360-diffusion-lora-for-sd-15

Even with a better dataset though, the top and bottom issues still seem to remain, and thus I think it might be related to LoRAs in some way rather than the dataset.

@progamergov Yup, I've taken a look at that as soon as it dropped, I'll be giving it a go shortly. I've noticed in a lot of this one (the LatentLabs360 model), the horizon doesn't seem to consistently line up near the middle on the gens I've done. I haven't given yours a try yet, but I'm thinking it might be the case as well.

I'm wondering if the distortion and horizon issues could be fixed by using a ControlNet with the proper distortions to kind of guide it into place a bit more. Might be worth looking into.

realise i've got the wrong lora here but same issue on both

https://civitai.com/models/4384/dreamshaper