SG161222

/

Realistic_Vision_V5.1_noVAE

like 178

+1 on SDXL

When will you train realistic vision on SDXL?

+1 on SDXL

Hi. I was going to train RV as soon as SDXL came out, but I've been having problems with kohya_ss that I've been trying to solve, but so far without success.

Hi. I was going to train RV as soon as SDXL came out, but I've been having problems with kohya_ss that I've been trying to solve, but so far without success.

Any specific problems the community might be able to help with?
Also looking forward to RV XL :)

When will you train realistic vision on SDXL?

+1 on SDXL

Hi. I was going to train RV as soon as SDXL came out, but I've been having problems with kohya_ss that I've been trying to solve, but so far without success.

I hope you get all the help you need. Realistic Vision is the BEST sd model there is. A XL version would probably blow minds 🤯

At this point I've made progress on the problem, the model has started to train, but....

during the training of the model image samples are generated, normally we would see something resembling training data, but in my case I get only noise. I am currently working on solving this problem.

we appreciate the efforts.. really we do

At this point I've made progress on the problem, the model has started to train, but....

during the training of the model image samples are generated, normally we would see something resembling training data, but in my case I get only noise. I am currently working on solving this problem.

This is just my guess on why this could be happening.
Is your dataset the same resolutions trained on SDXL? Using anything other than theirs could create artifacts like that. The resolutions are:

• 1024 x 1024
• 1152 x 896
• 896 x 1152
• 1216 x 832
• 832 x 1216
• 1344 x 768
• 768 x 1344
• 1536 x 640
• 640 x 1536

This is just my guess on why this could be happening.
Is your dataset the same resolutions trained on SDXL? Using anything other than theirs could create artifacts like that. The resolutions are:

• 1024 x 1024
• 1152 x 896
• 896 x 1152
• 1216 x 832
• 832 x 1216
• 1344 x 768
• 768 x 1344
• 1536 x 640
• 640 x 1536

I also tried the training at 896x1152 resolution, but the result is the same.

This is just my guess on why this could be happening.
Is your dataset the same resolutions trained on SDXL? Using anything other than theirs could create artifacts like that. The resolutions are:

• 1024 x 1024
• 1152 x 896
• 896 x 1152
• 1216 x 832
• 832 x 1216
• 1344 x 768
• 768 x 1344
• 1536 x 640
• 640 x 1536

I also tried the training at 896x1152 resolution, but the result is the same.
Is it possible for you to share all the parameters you are using with Kohya?

I think the problem's solved. Kohya_ss update has been released which fixed many problems including mine. I trained the model a bit on 1 thousand steps and 20 images from an old dataset at 768x1024 resolution. Watch out for NSFW content! https://postimg.cc/gallery/02pTDf2
The images represent women of legal age, in this case the minimum age is 20 years old.

I think the problem's solved. Kohya_ss update has been released which fixed many problems including mine. I trained the model a bit on 1 thousand steps and 20 images from an old dataset at 768x1024 resolution. Watch out for NSFW content! https://postimg.cc/gallery/02pTDf2
The images represent women of legal age, in this case the minimum age is 20 years old.

It's extraordinary to hear, I look forward to the model's release. 😁

I think the problem's solved. Kohya_ss update has been released which fixed many problems including mine. I trained the model a bit on 1 thousand steps and 20 images from an old dataset at 768x1024 resolution. Watch out for NSFW content! https://postimg.cc/gallery/02pTDf2
The images represent women of legal age, in this case the minimum age is 20 years old.

@SG161222 Good job, that's great to hear! Do you have an idea when you are going to release it?

@SG161222 Good job, that's great to hear! Do you have an idea when you are going to release it?

I won't be able to give a specific timeline. I am currently working on optimizing the tag base and typing dataset for training. Services like RunPod are not available for me, I have to train the model only on my single GPU.

@SG161222 Good job, that's great to hear! Do you have an idea when you are going to release it?

I won't be able to give a specific timeline. I am currently working on optimizing the tag base and typing dataset for training. Services like RunPod are not available for me, I have to train the model only on my single GPU.

Single GPU, that's incredible for a model that's so coherent. But since SDXL is such a bigger model, I guess we have to be patient.

@SG161222 if it helps i would love to donate a one week 4090s instance on vast.ai. Linux based, ssh root access. cuda 12.2 etc. ready to run.

@SG161222 and you can choose the country, ru, vn, tw, us, eu, .. worldwide with +500mb/s ..

or 2 days a 4 x 4090s. or 4 days a 2 x 4090s .. (mainly in island, and us).