Jump to content

How does your avatar look when enhanced by AI?


Danielle Atheria
 Share

Recommended Posts

This was an interesting experiment. I used Chat GPT to reverse engineer a friend's image, and then the results I got were amusing on the final image with my avatar lol.

Here was the original:

Screenshot2024-03-09171559.thumb.png.f8bc23d7c0afff0c110b81e5a2d96b89.png

Chat GPT said this:

The image displays a grayscale digital artwork depicting a sculpted male figure with a realistic, muscular build, reminiscent of classical marble statues. He has well-defined abs, pecs, and arms, with a subtle shadowing that enhances the three-dimensional effect. The man's face is in profile with chiseled features, sharp jawline, and tousled, medium-length hair. Two doves add life to the composition; one is perched gracefully on his shoulder, wings spread as if it just landed, while the other rests gently in his cupped hand. The background is a plain, light gray, putting the focus on the figure and the birds. The overall aesthetic is peaceful and harmonious.

I fed that prompt with a few changes into Gencraft, and it didn't go quite as planned, but the result amused me lol.

53576624802_77b5bf3832_5k.thumb.jpg.212bb240b250dbcaa72cf6ca9f2ee627.jpg

Honorable mention:

image0_20.thumb.jpg.b5be8cea2776bdab0a383b84ec043dcb.jpg
image0_0.thumb.jpg.96e52195157d97797dd61432027289bf.jpg

  • Like 9
Link to comment
Share on other sites

Creating a titanic themed image for a friend inspired me to create one of my own. In the movie, "Nearer My God To Thee" is being played on the violin as the ship is sinking. The anniversary is April 15th, which will mark 112 years. The film recently also had its 25th anniversary in 2023. This was created in Gencraft with a hyperrealistic model and a watercolor style.

53582017248_85aa2c211a_5k.jpg

  • Like 9
Link to comment
Share on other sites

  • 2 weeks later...

So, I've moved on from easy diffusion to stable diffusion, and have been having a blast. I created this using RevAnimated

00021-2839008783.thumb.png.6fd53fd2283931cd3944bd10011bf068.png

I installed InstantID into Automatic1111 and used this picture (once again) as my face model and started playing around with it.

Snapshot_644.jpg.ea186dc29b52d44a0fae1b16c1805954.thumb.jpg.86d2d8549f3c80504142b2f9fd4eade2.jpg

The below is one of my first tries.

00023-1316236641.thumb.png.b651b552d97de1cfe86c2b52d2f39fdd.png

It kind of merges both faces but I'm working on it.  

Then I thought I would give a statue a try

Snapshot_050.thumb.jpg.b4438a6ace76782e2821f06245c15dec.jpg

I love this statue, anyway, I removed the back ground and used InstantID again

00027-2375213302.thumb.png.a1900b7bcfbc759958c976494b881cf2.png

Meh, it is a work in progress, but still a lot of fun.

  • Like 7
  • Thanks 1
Link to comment
Share on other sites

On 3/25/2024 at 9:20 PM, Istelathis said:

So, I've moved on from easy diffusion to stable diffusion, and have been having a blast. I created this using RevAnimated

00021-2839008783.thumb.png.6fd53fd2283931cd3944bd10011bf068.png

I installed InstantID into Automatic1111 and used this picture (once again) as my face model and started playing around with it.

Snapshot_644.jpg.ea186dc29b52d44a0fae1b16c1805954.thumb.jpg.86d2d8549f3c80504142b2f9fd4eade2.jpg

The below is one of my first tries.

00023-1316236641.thumb.png.b651b552d97de1cfe86c2b52d2f39fdd.png

It kind of merges both faces but I'm working on it.  

Then I thought I would give a statue a try

Snapshot_050.thumb.jpg.b4438a6ace76782e2821f06245c15dec.jpg

I love this statue, anyway, I removed the back ground and used InstantID again

00027-2375213302.thumb.png.a1900b7bcfbc759958c976494b881cf2.png

Meh, it is a work in progress, but still a lot of fun.

Magnificent!

  • Like 1
  • Thanks 1
Link to comment
Share on other sites

Someone asked me how I made an earlier image in the thread and truthfully, I can't remember the precise parameters I used last time as it has been a while. The second best thing I can do is walk through the general process of it. I am going to re-use the same base image like I used before and turn it into an Anime picture (cursed hands included, thank you Stable Diffusion :P). 

I'm fully aware that there are now ready made solutions for this kind of deal but this is for the folks that want to do things by hand or understand what goes on behind the scenes and don't feel like forking over money. Base requirements: Either A1111 webui or SD.Forge webui. The latter recommended. The following extensions: ADetailer, SegmentAnything (optional, you can just do the masking by hand).

aiwitch.png.ccd4aa90a68d7f1c1e98e337afbecbff.thumb.png.b53b7cfefa56ab4fbb1ef850992bfe91.png00062-2332835005.thumb.png.ab0752cf206c275d95d3826ed967f5dd.png

 

That last one is what I want to start with. With the grey background, we can easily use SegmentAnything to create a mask for us. Red spot = I want this gone. Segment Anything will create a mask based on positive and negative points.

image.thumb.png.6d417329cd6aca833b9c1b3dbae094b8.png

No settings needed yet, just fire away, get your mask and put it into "Inpaint Upload". Now is when we start to configure things. To start with, doesn't matter whether you are using a 1.5 or SDXL checkpoint. What's important is the resolution. 1.5 was made for resolutions of 512x512 and SDXL for 1024x1024. You can stretch these somewhat but deviate too far and you'll get a lot of halucinations. Rule of thumb: Start low resolution, upscale finished picture. For my purposes, the image was already cropped to 823x1216 from the last time I did this.

Then I set the denoise to a value fo 1. Yes, 1. For the prompt I gave it something like the usual prompt vomit people use: "concept art forest . digital artwork, illustrative, painterly, matte painting, highly detailed" and a negative prompt of "photo, photorealistic, realism, ugly".

This gave me the following result:

image.thumb.png.d89e6486e3be043429d0f9bb9f1938bd.png

Not very anime yet but we've transplanted my character into a painted anime background. There are visible artefacts around the hair and if I were to inclined, I could clean those up in photoshop. From there it was just a matter of running a fairly medium denoise on the whole picture. Thus I sent the result to Image 2 Image and used the deepbooru autotagging to give me a description, which I then further tweaked:

"1girl, freckles, against tree, bamboo, bamboo forest, bare tree, breasts, brown hair, cleavage, collarbone, crow, denim, earrings, foliage, forest, glasses, grass, green eyes, jewelry, jungle, lips, looking at viewer, medium breasts, nail polish, nature, outdoors, palm tree, park, plant, short hair, skirt, solo, standing, tree, concept art forest . digital artwork, illustrative, painterly, matte painting, highly detailed" with "photo, photorealistic, realism, ugly" as the negative prompt.

Denoise was at... honestly I forgot already. It was likely somewhat around 0.45 or 0.35, depending on how much I wanted the image to change. I then turned on Adetailer. In the default configuration, it will make a second pass on the face, fixing most of the issues in one go. From there, I ran a few generations until I got one that was good enough.

If I wanted to, I could then use inpainting to fix the hands, maybe with the meshgraphformer model as that creates 3d replica of hands to use as a basis. It's all a bit fiddly and time consuming and with a pose like this, Stable Diffusion gets veeeery thirsty. After a few attempts that resulted in dangerously airborne skirts, I figured it's good enough. Usually I'm much more likely to just throw it into photoshop and do it by hand quick. Which I didn't do here to demonstrate the result. Cursed fingers included.

00062-2332835005.thumb.png.ab0752cf206c275d95d3826ed967f5dd.png

And that's it. In short:

Mask out the background. Generate background with prompts. Run an Img2Img pass on the result with prompts for background AND character as well as a mid level denoise. Hope this was helpful. The earlier picture from me using the same base was made much in a similar way.

Edited by ValKalAstra
  • Like 6
  • Thanks 1
Link to comment
Share on other sites

Please sign in to comment

You will be able to leave a comment after signing in



Sign In Now
 Share

×
×
  • Create New...