HOME | DD

Description
At some point in the past, I did a few experiments with using somebody's Sims 4 artwork as a base for AI image gen. Now I'm returning with some new tools to try out to see what can be done!

Since this is the first, I'll quickly run through the basics of what I did to get the image bases to use. I used Sims 4 to make all the bases, with the most notable mods being zel1302 's better in-game weightgain mod to help "enhance" the figures of my Sim models-to-be (though actually, haven't really used it to its full extent for most of what I'm going to share :v), and Pose Player to help in the few times I actually posed the sims. For cheats I used "headlineeffects off" to hide the speech bubbles/plumbobs, because they're annoying for pictures Previously in a guide I'd suggested extracting Depth + Normals from a game to use as ControlNet inputs, which I wanted to finally try out, so I installed Reshade along with the DisplayDepth effect for that. Then in-game, I could set up a scene, enter camera mode (TAB), and open the Reshade menu (which meant I couldn't then accidentally move the camera again). With the Reshade menu open, you can use Print Screen to take a screenshot of the game (which was more convenient with the base game's (c)amera option), so I took one screenshot with no effects on, one with DisplayDepth's Depth effect on, and one with DisplayDepth's Normals effect on. For depth effects I'd usually play around with the far plane setting, getting a few pictures at different depths in hopes one might improve the features I want. I do remember having to change some settings to get DisplayDepth fully working, but don't remember what. My Depth screenshots are also in the opposite format for ControlNet, so I had to use them with the Invert preprocessor.

So, as for this picture... This was my first, and probably worse documented, attempt. I was mostly figuring out what ControlNets would be best to use, and what settings would work (generally, I settled on ~0.5 for all their weights). Lower guidance attempts seemed to struggle with the concept of looking away - even with OpenPose, the model's biases fought against it. It wasn't until I started using a line ControlNet that I could mostly constrain the pose in there. Upscales/higher denoising had a tendency to eliminate all detail from the wall, so I had to composite together two images - one with a good character result, and one with a good background. While I wasn't fully intending to, I couldn't help myself but inpaint a bit to touch things up xP

The techniques/models I settled on for this image were:
Using the screenshot for an img2img base (various denoising strengths). This helped lock in the colours, which was somewhat needed to help the complex scene be separated out well. Also because I wanted to.
Canny/SoftEdge to help maintain the pose, the brickwork pattern that liked to get lost, and generally all the complex stuff in the scene.
Depth, to provide more context to what the lines mean, since the scene is quite busy and features depth.
Normals, because sometimes Depth and Canny would fail to interpret things correctly, such as adding holes in the wall or weird features at ambiguous parts. It still was a bit wonky with this, but less so. Note that I screwed up the Normals screenshot by accidentally leaving the game display option on - I don't know how well the resulting image actually worked, if at all.
It was only until after that I realised how fake the setting feels - like it's a photoshoot set. The in-game grass definitely adds to that, looking like astroturf, or green concrete, and the flat lighting doesn't help either (partially it's inherited from the game screenshot, and partially because I didn't prompt for it). You can also really see the polygons in that shirt there - in a later picture I edit around that, but I thought it'd be important to showcase a more default result here

Model: epiCPhotoGasm - Last Unicorn [62BB78983A]

This was the base image:

You can find the rest of the base images involved here .

Description
At some point in the past, I did a few experiments with using somebody's Sims 4 artwork as a base for AI image gen. Now I'm returning with some new tools to try out to see what can be done!

Since this is the first, I'll quickly run through the basics of what I did to get the image bases to use. I used Sims 4 to make all the bases, with the most notable mods being zel1302 's better in-game weightgain mod to help "enhance" the figures of my Sim models-to-be (though actually, haven't really used it to its full extent for most of what I'm going to share :v), and Pose Player to help in the few times I actually posed the sims. For cheats I used "headlineeffects off" to hide the speech bubbles/plumbobs, because they're annoying for pictures Previously in a guide I'd suggested extracting Depth + Normals from a game to use as ControlNet inputs, which I wanted to finally try out, so I installed Reshade along with the DisplayDepth effect for that. Then in-game, I could set up a scene, enter camera mode (TAB), and open the Reshade menu (which meant I couldn't then accidentally move the camera again). With the Reshade menu open, you can use Print Screen to take a screenshot of the game (which was more convenient with the base game's (c)amera option), so I took one screenshot with no effects on, one with DisplayDepth's Depth effect on, and one with DisplayDepth's Normals effect on. For depth effects I'd usually play around with the far plane setting, getting a few pictures at different depths in hopes one might improve the features I want. I do remember having to change some settings to get DisplayDepth fully working, but don't remember what. My Depth screenshots are also in the opposite format for ControlNet, so I had to use them with the Invert preprocessor.

So, as for this picture... This was my first, and probably worse documented, attempt. I was mostly figuring out what ControlNets would be best to use, and what settings would work (generally, I settled on ~0.5 for all their weights). Lower guidance attempts seemed to struggle with the concept of looking away - even with OpenPose, the model's biases fought against it. It wasn't until I started using a line ControlNet that I could mostly constrain the pose in there. Upscales/higher denoising had a tendency to eliminate all detail from the wall, so I had to composite together two images - one with a good character result, and one with a good background. While I wasn't fully intending to, I couldn't help myself but inpaint a bit to touch things up xP

The techniques/models I settled on for this image were:
Using the screenshot for an img2img base (various denoising strengths). This helped lock in the colours, which was somewhat needed to help the complex scene be separated out well. Also because I wanted to.
Canny/SoftEdge to help maintain the pose, the brickwork pattern that liked to get lost, and generally all the complex stuff in the scene.
Depth, to provide more context to what the lines mean, since the scene is quite busy and features depth.
Normals, because sometimes Depth and Canny would fail to interpret things correctly, such as adding holes in the wall or weird features at ambiguous parts. It still was a bit wonky with this, but less so. Note that I screwed up the Normals screenshot by accidentally leaving the game display option on - I don't know how well the resulting image actually worked, if at all.
It was only until after that I realised how fake the setting feels - like it's a photoshoot set. The in-game grass definitely adds to that, looking like astroturf, or green concrete, and the flat lighting doesn't help either (partially it's inherited from the game screenshot, and partially because I didn't prompt for it). You can also really see the polygons in that shirt there - in a later picture I edit around that, but I thought it'd be important to showcase a more default result here

Model: epiCPhotoGasm - Last Unicorn [62BB78983A]

This was the base image:

You can find the rest of the base images involved here .