I screwed up big time and uploaded the wrong file recently. This one is the correct one with all fixes implemented. I'm very sorry about this.

PRESS START

This is a big project of mine that lapsed thru multiple days and took many time and nerves to finish. I always push for quality over quantity and in this project went absolutely all my effort and skills I currently have.
For your greater entertainment there’s a timelapse of me creating this picture (actually there are two parts):
1. https://t.me/oaf40_pub/753 (~4 hours down to 8 minutes)
2. https://t.me/oaf40_pub/760 (~3 hours down to 5 minutes)
Please keep in mind ~20% of footage that should have been between two parts was lost and could not be recovered (I am a clown because at some point I shutdown my computer without clicking the Save button in the recorder program) and you could notice some sharp transitions here and there. I will explain the whole thing below.

Each girl symbolizes a ghost from the game Pac-Man. I generated each one separately and did few fixes such as removing the text from their t-shirts, fixing their claws/nails and eyes. I used the loractl plugin to combine many LoRAs at different timepoints of generation and to synthesize an interesting style that looked somewhat 3D’ish with just a little bit of celshading. Then I used the rembg plugin for A1111 to create a transparent background. The plugin works quite well but it tends to erode the dark parts near the edges so I fixed that by overlaying the original image and drawing the mask to restore the accidentally erased parts. I find this approach much more convenient because IMO it’s easier to run rembg and restore details from the inside using a mask rather than rely on the «magic wand» tool and fix black jagged edges from the outside. Yes I know Photoshop has that cool AI-assisted selection tool but I’m on Linux and GIMP is my only way to work.

Thorough the 1st timelapse you could see me realizing I’ve fucked up the color of Fennec’s clothes and trying to fix it using the «Colorize» tool. In the end I ditched it because it kept screwing up the sharp lines and making them very bright no matter what settings and blend modes I tried. After all I could always generate a new Fennec in yellow clothes, right? :P
To make three-dimensional Bonus Fruits I came up with a trick where I took the original game assets and ran them thru a free online PNG-to-STL/OBJ converter which allowed me to easily «extrude» the image to a desired depth. The in-browser model viewer has let me enable a solid background and rotate the model however I want.

The background seen in the end of the first timelapse is very simple and doesn’t look quality, and once all 4 characters were done I started working on improving the overall scene to replace the terrible background. Just prompting could not give me the desired background so I made my own sketch in multiple steps. GIMP has a very nice tool to render grids that I gladly abused to create separate ultra-high-res textures for the floor and the wall accordingly. The base texture size was 8000x8000. Each texture had the neon glow effect made by duplicating the grid, decreasing it’s exposure and applying Gaussian Blur. Atop of the floor I rendered and colored a weak (but still noticeable) Perlin Noise to help the AI create a well-textured floor tiles at the img2img stage. Both textures were imported into the new project file where I used the Free Transform tool to create the perspective. The basic scene was done.

I fired up txt2img to generate some arcade cabinets using the DucHaiten PonyXL model. I’ve got a really nice front-view picture and one slightly turned angle of an acrade cabinet. For the 4 cabinets in the center I just copy-pasted the same frontview picture and changed its tint, for 2 cabinets on the sides I flipped them horizontally and also changed their tint so they would have different colors after img2img. Then I added Pac-Man chasing Score Orbs on the wall with a small glow effect to imitate decorative lights.

Bonus Fruits also received a fix: I wrote the scores of each bonus on the PNG and extruded them as well.

The most painful part for me was Shadowing and Lighting. I experimented with Blend Modes, Blurring and Transparency levels to achieve a plausible look. I manually drew shadows for each character.

I tried IC-Light and actually did my best to get it working but unfortunately it’s max working resolution is still far below my working resolution, i.e. one character is 1024x1400 but IC-Light can only do up to 1024x1024 with terrible cropping. Yes it does work at lower resolutions but it screws up the characters themselves especially on the edges.

My next goal was to create a decent Depth-of-Field effect, I ran the Depth module of ControlNet against the background scene but somewhy GIMP could not pick up the generated depth map for automatic blurring. So I resorted to using masked Gaussian Blur to imitate gradual loss of focus starting close to the viewer (i.e. right behind the girls and bonus fruits) and getting stronger near the background wall.

At around 40 layers in GIMP it was starting to get really confusing to work so for the final merge I exported two halves with the background and foreground correspondingly.

My last step was creating a plausible light in foreground manually because IC-Light has let me down. First I made the color temperature a bit colder (the characters and bonus fruits were much «warmer» than the background) and then experimented a lot with the «Render Light» tool in GIMP.

What you see here on e6 is the final result :P I will allow myself a few days of rest before I could ever think about starting a new project because this one was real exhausting :D

Models used (all are available on civitai):
1. Indigo Furry Mix XL v2.0 (characters)
2. DucHaiten PonyXL something something no score tags v4.0 (background)
3. Midgard THL Hybrid v1.3 (background)
LoRAs used for the characters (note that I also use the loractl plugin):
<loractl:Drako_Style_Lora:0.5@0,[email protected],[email protected],0@1> https://civitai.com/models/571719/drako-style-lora?modelVersionId=637260
<loractl:LadyShalirin:1@0,[email protected],[email protected],0@1> https://civitai.com/models/378865/ladyshalirin-or-pdxl-or-style-lora?modelVersionId=423012
<loractl:incase-ilff-v3-4:0.6@0,[email protected],[email protected],0@1> https://civitai.com/models/300005/incase-style-ponyxl?modelVersionId=436219
<loractl:PDXL_artist_tags_v2_big:0@0,[email protected],[email protected],0.9@1> https://civitai.com/models/317578/pdv6xl-artist-tags?modelVersionId=389962
<loractl:3D_Animation_Diffusion_Pony_style:0@0,[email protected],[email protected]> https://civitai.com/models/405143/3d-animation-pony-style?modelVersionId=451707