ControlNet Strength: Controls how closely the output follows source images' lines and shapes
IP-Adapter Strength: Controls how much color/style transfers from reference painting
Influence Steps: Number of refinement iterations (quality)
Guidance Scale: How strongly AI follows all inputs
Seed: Random seed for reproducibility
Base: stabilityai/stable-diffusion-xl-base-1.0 --> Core image generation engine
CLIP: openai/clip-vit-base-patch32 (22 landscape classes) --> Semantic understanding of sketch regions
ControlNet: lllyasviel/sd-controlnet-softedge --> Preserves sketch structure during generation
SAM: ViT-B checkpoint --> Precise object boundary detection
IP-Adapter: h94/IP-Adapter XL variant --> Style/color transfer from reference image
NVIDIA RTX 6000 Blackwell
CUDA 12.9
Reference
Original Gaussian Splat
100% Style Transfer Intensity
50% Style Transfer Intensity
0% Style Transfer Intensity
Gaussian Splat looks blurry after receive a style transfer althought the transfer intensity is 0%
The style transfer missing in some area
The transfer process takes about 10 minutes
Use OpenUSD to store multiple variants of style of the same gaussian splat object
Testing original and stylized gaussian splat in Unreal
Refine Gaussian splat style AI transfer tool