imagecondition train

[andorxornot]

I tried to train imagecondition exp with the IF model (because I couldn't set it up with SD), so I replaced the model in the config, and got an error in the "depth loss" line. So I disabled the "depth loss" , and the model after 50k step still doesn't look very much like the ref image

[andorxornot]

My prompt:" a marble sculpture of a young, handsome poet with long hair kneeling before God" cond image ( generated through IF and edited in photoshop): result of 52000:

[zouzx]

Hi, our image-conditioned generation is still an experimental version, it indeed can’t make it like ref image very much for now. We are working on better image-conditioned results, please stay tuned.

[andorxornot]

thanks, I also pre-processed the image with the stable-dreamfusion script, so the depth loss works now, but the results are the same. i think the result is the same as without any the loss

preprocessed: 46000:

[bennyguo]

Could you get better results with stable-dreamfusion image-to-3D? They did a lot of experiments on this.

[andorxornot]

Could you get better results with stable-dreamfusion image-to-3D? They did a lot of experiments on this.

after 50 epochs with stable-dreamfusion i got it:

[bennyguo]

I suggest you try Make-It-3D. It can produce results well aligned with the input image. By the way, Make-It-3D will come to threestudio soon, please stay tuned :)

[andorxornot]

yeah, i will try. it would also be good to add an image normal condition for fantasia3d, in my tests this model shows good results. unfortunately i am not very familiar with lightning

[andorxornot]

I suggest you try Make-It-3D. It can produce results well aligned with the input image. By the way, Make-It-3D will come to threestudio soon, please stay tuned :)

trying Make-It-3D without coarse stage. nothing interesting, the same result as stable-dreamfusion