Rugby Variations. After Laocoon Group
I decided to work on one of the images used in the collage to show you my usual process. I wanted to use these images together because they demonstrate the power of Control Nets by combining something old and something new. I also had a plan to create something special with all these images, even if it’s not clear yet.
To make this image, I first created a depth map from an image of the Laocoon Group sculpture. I used the depth map image to drive a depth control net inside Stable Diffusion, and described a scene showing a group of rugby players inside a locker room with steam. Stable Diffusion then adapted my text prompt to the depth information, and created a final image.
Next, I took the image and made it bigger using Topaz Gigapixel. I adjusted the colors in Adobe Lightroom and fixed any areas with issues in Adobe Photoshop. This kind of post-processing can take a long time. Sometimes, I wish I had a team of designers and retouchers to help me with this often very tedious part of the process.
Many people think that working with AI is easy, but it’s actually the opposite. To get the exact result you want, you need to go through many rounds of revisions, both within the AI program and outside of it. This process can often take longer than working with a regular photo. Since AI is not always perfect, some post-processing is needed to get the desired outcome. You don’t have to post-process your AI-generated images, but I always do because I’m used to the photography workflow.
Stable Diffusion, depth control net, Virile Fusion 3 Beta 1.
Technical Details
- Rugby Variations. After Laocoön Group.
- After Series / 2024-01-09 / 2023-01-07 Render
- GD8.OV2.20240107.103424.00224-3982069216-VF3B1.E2
Stable Diffusion txt2img
- Steps: 30
- Sampler: DPM++ SDE Karras
- CFG scale: 7
- Seed: 3982069216
- Size: 512x512
- Model hash: de1aabc4d5
- Model: virileFusion_v30BETA1
- VAE hash: 63aeecb90f
- VAE: vae-ft-mse-840000-ema-pruned.safetensors
- Denoising strength: 0.5
- Clip skip: 2
- Hires upscale: 4
- Hires steps: 30
- Hires upscaler: 8x_NMKD-Superscale_150000_G
- Hires size: 2048x2048
- Automatic1111 Version: v1.7.0
Control Nets
- ControlNet 0: Depth
- CN0 Module: depth_leres
- CN0 Model: control_v11f1p_sd15_depth cfd03158
- CN0 Weight: 0.8
- CN0 Resize Mode: Crop and Resize
- CN0 Low Vram: False
- CN0 Processor Res: 1024
- CN0 Threshold A: 0
- CN0 Threshold B: 0
- CN0 Guidance Start: 0
- CN0 Guidance End: 1
- CN0 Pixel Perfect: False
- CN0 Control Mode: My prompt is more important
ADetailer 1
- ADetailer 1:
- AD1 model: face_yolov8n.pt
- AD1 prompt: …
- AD1 confidence: 0.3
- AD1 dilate erode: 4
- AD1 mask blur: 4
- AD1 denoising strength: 0.4
- AD1 inpaint only masked: True
- AD1 inpaint padding: 32
- AD1 use separate steps: True
- AD1 steps: 28
- AD1 ControlNet model: control_v11f1e_sd15_tile a371b31b
- AD1 ControlNet module: tile_resample
Post-Processing
- Upscale: Topaz Gigapixel HQ 4x
- Size: 8192x8192
- Color: Adobe Lightroom
- Inpaint: Adobe Photoshop
- Collage: Adobe InDesign