Human Korvo but It's Real Dan Stevens (Made with Stable Diffusion/Photoshop/MTXX)
(Here is a screenshot of Solar Opposites S4E11 to compare)
First of all, I understand that there is some controversy on AIGC currently... However, I still decided to post this work here. Yes, I call it “work” because I had spent at least 5 hours to make this image, and I went through hundreds of small steps to achieve this effect. It's not perfect, of course, since I basically know nothing about drawing and photography…but I tried my best.
To make this post more "meaningful", here is the complete detailed workflow (FYI: It's VERY LONG. Here is a 1-minute summary video I made, in case you wanna save some time:)
Preparation stage:
①A base model (or base model+LoRA/Embedding, depending on your need). Normally people use LoRA for a specific person’s face, but I’m not using LoRA for Dan because small models don’t reconstruct his face very well. I’m using a base model (in my case, Serenity V1.0) trained on his photos with Dreambooth. I trained this model 3 weeks ago so it’s ready to use.
(If you’re wondering, “why not using Midjourney to generate celebrities?”, the answer is, Midjourney does a terrible job generating Dan’s face. Also it strictly forbids NSFW. Even just mild NSFW.)
②An OpenPose map and a Scribble image, to use in ControlNet (an extension on SD WebUI). I made a screenshot of the posed dummy from Magic Poser, and used this image to make an OpenPose map, and trace the outline of the dummy on Photoshop to make a Scribble image. Why do I need to make a Scribble? Because from my experience, all SD1.5 base models have a hard time generating “man wearing unbuttoned shirt”. So, only an OpenPose map is not enough. Gotta let it know what shape I want.
Base image making stage:
Step 1: Write a prompt.
I’m not good at writing prompt but it does the job:
Positive prompt:
full color photo of RLDAN, (wearing_unbuttoned_black_long-sleeve_glitter_shirt), BREAK, RLDAN, (long light blond hair), BREAK, RLDAN, (holding a martini glass), in a bar, looking to the right, (flat lighting), sharp focus, raw photo, 8k uhd, high resolution, DSLR, high quality, Fujifilm XT3
About the prompt, I didn’t use any negative embeddings because sometimes they affect my trigger word; I add “muscular” in negative prompt because if I don’t, it’ll give me super buff man and it’ll look weird; “BREAK” is for separating groups of words.
Step 2: Choose a decent “raw” image
Editing the prompt and adjusting the ControlNet weights according to the results, and cherry-picking the better ones. The style of the clothing is the most important feature to imitate the original animation, so it has to be accurate enough. After cherry-picking the first okay-ish image, I reused its seed, and put this image into Lineart-realistic preprocessor to use the Lineart ControlNet model. The second cherry picked image looked more natural on lighting and background, so I saved this one for future use. I then used ADetailer (an extension) to swap face roughly, not good enough but that’s far from final.
Step 3: Upscale this image
Since SD1.5 was based on 512 resolution, my initial image size was 682 x 512, but that’s too small to make a higher quality face. Hence I used img2img+Tiled Diffusion+ControlNet Tile Resample to upscale x 2. So I get a 1364 x 1024 image to work on. I furthur upscaled it with MTXX, just 3 clicks.
Step 4: Inpaint the hair and face
I wasn’t satisfied with the hairstyle in the previous image. It doesn’t look natural and there’s no tendril near ear, so I gotta fix it. I masked the hair area and edited my prompt, and used a 0.55 denoising to get a better hair, and then used inpaint sketch to draw a blond strand near his ear, and used a 0.75 denoising to add a tendril. There were some artifacts in the image so I used Photoshop to remove those. Finally I used Adetailer again to change the face, this time it looked much more similar to the real Dan.
Post-processing(?) stage:
The Stable Diffusion part was complete. Now the Photoshop time. I changed the background (remove the wall-like thing and replace it with night view), martini glass bottom (it was really messed up), the shirt buttons (there weren’t any before), and of course the hand (I normally use OpenPose hand map to control it but this pose is rather complex, and I don’t like using depth/canny models so…) with Generative Fill. Also plenty of times of remove tool (and patch tool, color replacement tool, burn tool, liquify…yada yada yada). Then I used Neural Filters-Harmonization to match the color tone of the torso to the tone of face. Finally, I used MTXX to make some small adjustments on face and body.
the tag has been on fire! ive had a hard time keeping up since school started again, but I'll catch up soon 🖖 in the meantime enjoy some week-old fanart!
It was such an honor to be one of the Co-Art Directors of Season 4 of Solar Opposites! I had such a blast with my talented team, continuing to explore and expand on the Solar Opposites universe! It was also really fun push the atmosphere of everything and come up with wacky colors!
titsay
janainaart
germansingergirl
ellievsbear
discoholicmusic
dailybridgerton
gummygunk
blakekathryn
2087
