alexgenovese/seesr-upscale-and-stable-diffusion-turbo 🔢🖼️✓📝 → 🖼️

▶️ 50 runs 📅 Aug 2025 ⚙️ Cog 0.13.6 🔗 GitHub 📄 Paper
image-restoration image-upscaling

About

Image super-resolution model based on SEESR, designed to reconstruct low-resolution images with sharper details and accurate colors.

Example Output

Output

Example output

Performance Metrics

25.39s Prediction Time
151.39s Total Time
All Input Parameters 
{
  "seed": 231,
  "image": "https://replicate.delivery/pbxt/NWALlB7XUGNjhOOM8oUSDVf0kOXtJO9ILt35TgER3WR2cr8X/ComfyUI_00021_.png",
  "use_kds": true,
  "bandwidth": 0.1,
  "cfg_scale": 1,
  "user_prompt": "a woman taking sun",
  "scale_factor": 4,
  "num_particles": 10,
  "negative_prompt": "dotted, noise, blur, lowres, oversmooth, longbody, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality",
  "positive_prompt": "clean, high-resolution, 8k, best quality, masterpiece",
  "latent_tiled_size": 320,
  "num_inference_steps": 4,
  "latent_tiled_overlap": 4
}
Input Parameters
seed Type: integerDefault: 231Range: 0 - ∞
Random seed for reproducibility
image (required) Type: string
Input image to upscale
use_kds Type: booleanDefault: true
Use Kernel Density Steering
bandwidth Type: numberDefault: 0.1Range: 0.1 - 0.8
Bandwidth for KDS
cfg_scale Type: numberDefault: 1Range: 0.5 - 1.5
Classifier Free Guidance Scale (SD Turbo optimized at 1.0)
user_prompt Type: stringDefault:
User prompt to guide the super-resolution
scale_factor Type: integerDefault: 4Range: 1 - 6
Super-resolution scale factor (max 4x recommended for Replicate)
num_particles Type: integerDefault: 10Range: 1 - 16
Number of particles for KDS
negative_prompt Type: stringDefault: dotted, noise, blur, lowres, oversmooth, longbody, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality
Negative prompt
positive_prompt Type: stringDefault: clean, high-resolution, 8k, best quality, masterpiece
Positive prompt
latent_tiled_size Type: integerDefault: 320Range: 128 - 480
Diffusion tile size for memory management
num_inference_steps Type: integerDefault: 4Range: 1 - 8
Number of inference steps (SD Turbo optimized for 1-4 steps on Replicate)
latent_tiled_overlap Type: integerDefault: 4Range: 4 - 16
Diffusion tile overlap
Output Schema

Output

Type: string • Format: uri

Example Execution Logs 
Seed set to 231
INFO:predict:Input image size: 1216x1600
0it [00:00, ?it/s]
0it [00:00, ?it/s]
ERROR:predict:Fallback processing failed: stack expects a non-empty TensorList
INFO:predict:Super-resolution complete. Output saved to /tmp/tmpro4j7o4k/upscaled.png
INFO:predict:Final image size: (4864, 6400)
Version Details
Version ID
e42821945faaf99e263f538b786063875c56cd3ffc5e33ee357dd9ed6f4017d8
Version Created
August 11, 2025
Run on Replicate →