
Why this works
ClipVision (aka “PhotoMaker” in many UIs) encodes the identity, pose, framing, and palette of your source image into a compact embedding. ControlNet then conditions diffusion with that embedding, keeping the output faithful while your text prompt and LoRA swap in the Sega Saturn aesthetic. GIMP closes the loop: AI text is usually nonsense, so we white-out and retype. All models shown are open source at civitai.com.
What you’ll need
- Stable Diffusion WebUI (ReForge fork of A1111) with ControlNet extension enabled.
- SDXL anime checkpoint (e.g., Hassaku/Illustrious).
- PS1-style LoRA (character/environment).
- GIMP (for textbox patching).
- Source still: a main-menu close-up of Gold Ship.
At-a-glance config
Checkpoint | SDXL anime (Hassaku/Illustrious) |
---|---|
Sampler / Steps | Euler a · 25–40 steps |
CFG | 4.0–5.5 |
Resolution | 768×768 (downscale to 320×240 later for Saturn feel) |
LoRA | <lora:ps1_character:0.8–1.0> |
ControlNet Unit 0 | Preprocessor: clip_vision (PhotoMaker) Model: ControlNetXL Weight: 0.65–0.85 · Pixel Perfect: ON · Resize: Just Resize · Control Mode: Balanced |
Post | GIMP textbox + optional CRT layer |
Step-by-step: Gold Ship → Saturn Gold Ship
0) Capture the source still
From Uma Musume’s menu, grab a clean Gold Ship shot. Crop so she fills the frame but keep hair/accessories intact.

1) Load the base model
In Forge/ReForge, load your SDXL checkpoint. Suggested: 768x768 resolution, sampler Euler a, CFG 4.5.
2) Wire up ControlNet
- Enable Unit 0.
- Input: your Gold Ship still.
- Preprocessor:
clip_vision
(PhotoMaker). - Model: ControlNetXL.
- Weight: ~0.7. Adjust if drift occurs.
3) Add the PS1-style LoRA
Drop LoRA in models/Lora
. Prompt: <lora:ps1_character:0.9>
. It brings blocky textures, vertex wobble, low color depth—the right retro 3D vibe.
4) Prompting the Saturn look
Positive prompt:
1girl, gold ship (umamusume), close-up portrait, smug expression,
sega saturn aesthetic, 1996 retro 3D, low poly look, 400-800 polygon look,
affine texture mapping, warped quads, vertex wobble, dithered shadows,
8-bit color banding, flat lighting, transparency artifacts, slight bilinear blur,
stadium backdrop bokeh, <lora:ps1_character:0.85>
Negative prompt:
watermark, logo, text, signature, nsfw, extra fingers, extra limbs, distorted eyes, deformed hands
5) Generate
Run several seeds. Adjust CFG and LoRA weights for retro fidelity. Keep ControlNet weight higher if identity drifts.

6) Patch the textbox in GIMP
- Open the image.
- Rectangle Select + Fill: cover the garbled text with a white box.
- Use Text Tool: Comic Sans, MS Gothic, or Pixel Operator fonts.
- Optional: apply a slight pixelate/blur filter for harmony.

7) Optional Saturn finishing touches
- Downscale to 320×240, then upscale ×4 with nearest neighbor.
- Add scanlines/dither at 10–15% opacity.
- Add chroma noise for composite vibes.

Troubleshooting
Face off-model?
Raise ControlNet weight or let it guide more.
Too HD?
Lower CFG, raise LoRA, add “low color depth” tokens.
Framing drift?
Pixel Perfect ON; crop tighter.
Text messy?
Always fix in GIMP—faster and authentic.
Copy-paste quick start
// Positive
1girl, standing, anime, lowpoly, soccer stadium, pointing at viewer, full body,Gold Ship, umamusume, tracen academy uniform, grinning, holding trophy, retro 3d style, sega saturn aesthetic, warped quadrilateral polygons, affine texture distortion, layered sprites, blocky texture, low resolution, pixelated shading, dithered shadows, transparency artifacts, VDP1 rendering, early 3d, vertex wobble, low color depth, model2 era, flat lighting, ps1_character style, track field
<lora:ps1_character:0.9>
// Negative
watermark, triangle polygons
// ControlNet Unit 0
Preprocessor: ClipVision (PhotoMaker)
Model: ControlNetXL
Weight: 0.7 · Pixel Perfect: OFF · Resize: Crop and Resize
// Sampler/Size
Euler a · Steps: 35 · CFG: 3.7 · Size: 768x768
Notes on Saturn vs PS1
Saturn and PS1 render differently, but a PS1-trained LoRA plus prompt cues (affine quads, dithering, vertex wobble) suggest Saturn-era 3D convincingly. The aim isn’t a strict hardware simulation, just the nostalgic aesthetic.