SL Avatars in RL.

[Madi Melodious]

Have you ever wondered what your avatar would look like if you met them in RL?  Let's admit that even the most realistic looking SL avatar looks cartoonish.  Using Stable Diffusion SDXL I created a model of my AV to see if I could work it out.  Someone told me that SL avatars don't translate well to RL images.   This is what I got after few hours working on the process. 

 

 

Edited September 2 by Madi Melodious
image issue

[Stephanie Misfit]

You might be interested to see what other have done to their avatars with AI in this thread.

[Madi Melodious]

Oh cool.  Thanks.  I never noticed that thread.  I wish we could delete threads now.

Edited September 3 by Madi Melodious

[Amata Karas]

5 minutes ago, Madi Melodious said:

Oh cool.  Thanks.  I never noticed that thread.  I wish we could delete threads now.

Nooo, I like this one. The other is for all sort of AI expression of avs. Keep this one specifically to get as realistic a look as possible.

I like hearing the process and tools used, too. After the disaster that was my contribution to the other thread, I'm low-key glad to hear your results took hours. Mine was a few minutes of changes to prompts and meh, that didn't work, I quit.  

[Madi Melodious]

10 minutes ago, Amata Karas said:

 like hearing the process and tools used, too. After the disaster that was my contribution to the other thread, I'm low-key glad to hear your results took hours. Mine was a few minutes of changes to prompts and meh, that didn't work, I quit.  

I'm more than happy to tell you how I did it.  I'll even share the tools and workflow.   Basically, I made a custom lora of my avatar using kohya_ss.    I took about 50 pictures of my head, and another 50 of my body from various angles and poses.  Then using kohya I captioned them and created the lora.   That took almost 2 hours.   

Once I created the model and verified that it worked.  I spent another few days figuring out which base model seemed to work the best.   This image was made using swiperxl.fp16.  The lora itself was trained against the base sdxl model.  I didn't use a refiner.

I combined the body and head lora in a stack to create that image.   I used a custom comfyui workflow for the results with 35 steps, a cfg of 4.2.  dpmpp_2m_sede_gpu with a karras scheduler.

[Amata Karas]

12 minutes ago, Madi Melodious said:

I'm more than happy to tell you how I did it.  I'll even share the tools and workflow.   Basically, I made a custom lora of my avatar using kohya_ss.    I took about 50 pictures of my head, and another 50 of my body from various angles and poses.  Then using kohya I captioned them and created the lora.   That took almost 2 hours.   

Once I created the model and verified that it worked.  I spent another few days figuring out which base model seemed to work the best.   This image was made using swiperxl.fp16.  The lora itself was trained against the base sdxl model.  I didn't use a refiner.

I combined the body and head lora in a stack to create that image.   I used a custom comfyui workflow for the results with 35 steps, a cfg of 4.2.  dpmpp_2m_sede_gpu with a karras scheduler.

I am impressed! I only dabble, I've been trying to learn enough to come up with basic visuals that aren't people with eighteen fingers coming out of their hip 🤣

Seriously, thank you for writing that out. I didn't understand 3/4 of what you wrote on first read. But looking up the unfamiliar terms then going back to them in the context of your workflow gives me a way to try to approach something that, I will admit, is not at all intuitive to me ♥

 

[BillFletcher]

On 9/2/2024 at 9:01 PM, Madi Melodious said:

Oh cool.  Thanks.  I never noticed that thread.  I wish we could delete threads now.

Or we could use this one to talk about those moments when you see some stranger and think maybe you know them, then realize they look like the RL version of one of your SL friends. Or maybe that only happens to me.

[JeromFranzic]

This is from last year. I won't say who did for me, but she'll recognize it. I was so flattered... and I mean really flattered on so many unspeakable levels lol. Who wouldn't want to spend a weekend with her... *hiss*

 

[Zrina Nebula]

I wish someone would do that with my Avi. I tried it (picture in the other thread) but I was never able to load up a picture of my Avi and tell some AI to make it look real.

[Madi Melodious]

2 hours ago, Zrina Nebula said:

I wish someone would do that with my Avi. I tried it (picture in the other thread) but I was never able to load up a picture of my Avi and tell some AI to make it look real.

I'll do it for you.   Contact me in world or with a message here and we can set up a time and I'll tell you what I need to do.   My process does take a few days to complete, I don't just type a bunch of words into a prompt and hope for the best.   But if you are interested give me a shout.