Early Explortation on Local Ran Stable-Diffusion Models

I’m in the early stages of exploring AI-powered image generation, one of the most exciting developments has been the ability to run Stable Diffusion models locally, even on average consumer-grade hardware like the NVIDIA GeForce RTX 3060 Ti. Despite not being the top-tier GPU, this graphics card still provides enough power to experiment with high-quality image refinement techniques.

For this demonstration, I started with a low-quality reference image of a race-car (512x512 pixels).

Models

I used this image as the input latent, and upscaled it by a factor of 2 (1024x1024 pixels).

My goal was for the end result to take hints from the original image, but it to be unrecognizable from the source.

Here are the different models I used to enhance and upscale the image (discovered and obtained at https://civitai.com):

  • v1-5-pruned-emaonly
  • dreamshaper_8
  • realisticVisionV60B1_v51HyperVAE
  • glamourCelebrity_v10
  • aniversePonyXL_v20
  • blackMAGICPONY_v25
  • revAnimated_v122EOL
  • mcbsMachinecodesComic_v4
  • waiREALCN_v120
  • animeArtDiffusionXL_alpha3

Settings

  • Steps: 35
  • Guidance: 5
  • Start At Step: 3
  • Seed: 173755800709141

Prompt

photoshoot (masterpiece) of a bright (pink race car:1.25), near the ocean, at sunset, amazing ocean side view from street, 4k, 8k, udh, dslr, soft lighting, high quality, file grain, Fujifilm XT3

Sampler

For each model I genearated 4 images, each a different sampler + scheduler combination:

  • Euler + exponential
  • Euler A + exponential
  • DPMPP 2M SDE + Karras
  • DPMPP 3M SDE + Karras

v1-5-pruned-emaonly


dreamshaper_8


realisticVisionV60B1_v51HyperVAE


glamourCelebrity_v10


aniversePonyXL_v20


blackMAGICPONY_v25


revAnimated_v122EOL


mcbsMachinecodesComic_v4


waiREALCN_v120


animeArtDiffusionXL_alpha3


It’s fascinating to see how even low-quality images can be transformed into something remarkable using Stable Diffusion models, even on mid-range hardware.

Here are a few of the above generations that I think came out especially well:

I hope you enjoyed this exploration of image enhancement. Stay tuned for more.