Bard, Bing Image Creator, and DALL-E are exceptional tools for image generation. However, when it comes to creating realistic human faces, they often yield results with a plastic-like appearance in the skin texture. This tutorial provides a comprehensive guide on how to address this issue using Automatic1111, aiming to enhance the skin texture for a more lifelike look.
Examples of plasticky faces
Here are the examples of plasticky faces for each platform and prompts used for generation:
We will be using the DALL-E image as an example in below steps. We will also show the result of repair for the remaining two images at the end.
Steps
Required Tools
We will be using:
- Automatic1111
- ControlNet Extension for Automatic1111
- ControlNet tile preprocessor and model
If you have never used Automatic1111, check out Getting started with Automatic1111. If you are not familiar with how to use ControlNet with Automatic1111, checkout How to use ControlNet in Automatic1111 Part 2: Installation, which covers both installing the extension and ControlNet models. If you have never used ControlNet Tile, you need to download and install the control_v11f1e_sd15_tile
model following the instructions listed in this tutorial.
Setting up parameters
Start Automatic1111 and nagivate to img2img tab.
For the main model, load a model that can produce high-quality images. In this tutorial, epicrealism_naturalSin is used. For prompts, enter words that will push the model towards generating skin details (prompts are included in full generation parameters listing at the end of steps).
Drag the source image to the canvas below img2img tab.
Make sure that image resolution is set to the size of the source image. Set Denoising strength
to 0.45 but adjust this value higher if you don’t see enough skin details, or lower if you see too much details.
Expand the ControlNet
section. Select Enable, Low VRAM, Pexel Perfect
. Select Tile/Blur
as Control Type
. Make sure that tile_resample
is selected for Preprocessor
and control_v11f1e_sd15_tile
is selected for Model
.
Now press Generate. You should see an image with more realistic skin texture generated.
If you don’t get the desired result, you may need to iterate a few times by adjusting skin-related words in positive prompt and Denoising Strength.
Generation parameters
beautiful 25-year-old woman,
skin details, skin imperfections, (skin pores:1.2), natural skin, skin details, textured skin, skin texture, red lipstick,
photorealistic, masterpiece, best quality, hires, 4k, 8k, uhd, highly detailed, high resolution, extreme details, sharp focus,
canon 5d mark iv, dslr
Negative prompt: worst quality, low quality, 3d, drawing, fake, plastic skin, rubbery skin, smooth skin, ugly, disfigured
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 3612753737, Size: 1024x1024, Model hash: bff4610d23, Model: epicrealism_naturalSin, VAE hash: 235745af8d, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.45, ControlNet 0: "Module: tile_resample, Model: control_v11f1e_sd15_tile [a371b31b], Weight: 1, Resize Mode: Crop and Resize, Low Vram: True, Threshold A: 1, Guidance Start: 0, Guidance End: 1, Pixel Perfect: True, Control Mode: Balanced, Hr Option: Both, Save Detected Map: True",
Here are the before and after images:
Dall-E Image
Bing image
Bard image
Congratulations on mastering the technique to overcome the plastic-like appearance of skin textures. This knowledge will enable you to fully leverage the robust image creation capabilities of these generators, effectively overcoming a significant challenge in generating realistic faces.
Appendix
Fixed Bing image generation parameters
Note that for this image, Inpaint was used instead of img2img in order to mask the fingers to keep them from getting distorted during processing.
beautiful 25-year-old woman,
skin details, (skin imperfections:1.2), (skin pores:1.4), natural skin, skin details, textured skin, skin texture, red lipstick,
photorealistic, masterpiece, best quality, hires, 4k, 8k, uhd, highly detailed, high resolution, extreme details, sharp focus,
canon 5d mark iv, dslr
Negative prompt: worst quality, low quality, 3d, drawing, fake, plastic skin, rubbery skin, smooth skin, ugly, disfigured
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 1922648393, Size: 1024x1024, Model hash: bff4610d23, Model: epicrealism_naturalSin, VAE hash: 235745af8d, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.35, Mask blur: 4, ControlNet 0: "Module: tile_resample, Model: control_v11f1e_sd15_tile [a371b31b], Weight: 1, Resize Mode: Crop and Resize, Low Vram: True, Threshold A: 1, Guidance Start: 0, Guidance End: 1, Pixel Perfect: True, Control Mode: Balanced, Hr Option: Both, Save Detected Map: True",
Fixed Bard image Generation parameters
headshot of a 30-year-old man wearing a business suit,
skin details, skin imperfections, skin pores, natural skin, skin details, textured skin, skin texture,
photorealistic, masterpiece, best quality, hires, 4k, 8k, uhd, highly detailed, high resolution, extreme details, sharp focus,
canon 5d mark iv, dslr
Negative prompt: worst quality, low quality, 3d, drawing, fake, plastic skin, rubbery skin, smooth skin, ugly, disfigured
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 526843340, Size: 1536x1536, Model hash: bff4610d23, Model: epicrealism_naturalSin, VAE hash: 235745af8d, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.45, ControlNet 0: "Module: tile_resample, Model: control_v11f1e_sd15_tile [a371b31b], Weight: 1, Resize Mode: Crop and Resize, Low Vram: True, Threshold A: 1, Guidance Start: 0, Guidance End: 1, Pixel Perfect: True, Control Mode: Balanced, Hr Option: Both, Save Detected Map: True", Pad conds: True