3 mins read

Generating Realistic Portraits with Stable Diffusion 1.5


Creating AI images that look extremely realistic is not an easy task. Creating AI-generated images that are strikingly realistic, especially when we are talking about realistic portraits photos, can be sometimes challenging. Despite the advancements brought by the latest Stable Diffusion XL, its predecessor, Stable Diffusion 1.5, retains a noteworthy edge in certain aspects, and generating realistic portraits is one of those. This version stands out not only for its ability to produce high-quality images but also for its speed and lower resource requirements, making it a practical and efficient choice for many tasks.

Generating Realistic Portraits with Stable Diffusion 1.5

In this article, I will showcase the resources and techniques I used to generate realistic portraits with Stable Diffusion 1.5, utilizing the automatic1111 GUI. The primary goal here is to produce portraits that don’t just look real but feel as if they were captured directly by a camera. A key objective of these portraits is ensuring that the people in the picture appear ordinary and relatable, steering clear of the over-idealized or artificial looks often associated with supermodels or synthetic characters that Stable Diffusion tends to create, if not instructed differently.

Resources Used

  1. Base Model: Epicphotogasm This innovative base model plays a pivotal role in image generation, offering a foundation that’s both detailed and versatile, setting the stage for nuanced art creation.
  2. Loras: EpiCRealLife and Analogdiffusion These Loras act as enhancements, injecting realism (EpiCRealLife) and a classic, analog feel (Analogdiffusion) into the portraits, elevating them from mere digital creations to works of art.
  3. Negative Embedding: EpiCNegative EpiCNegative is crucial in refining outputs. It helps mitigate common AI art issues like overly dark imagery and ensures a balanced representation in gender portrayal, a noted tendency in this version.

The Base Prompt

Crafting the perfect prompt is an art in itself. A well-structured prompt for Stable Diffusion 1.5 might look like this:

"Close up analog style portrait of a [Nationality] [Man/Woman] ([Profession]), background of [Place], (homely), looking at the camera, f1.4 lens, dramatic composition, cinematic lighting, [Facial Expression], lora:analogdiffusion_Lora300:<0.6> lora:epiCRealLife:<0.3>"

Each element of this prompt, from nationality to profession, background setting, and even camera specifications, plays an important role in defining the details of final output, ensuring each portrait is not just unique but also resonates with a sense of realism and familiarity – basically, I want common people, not models.

anime, drawings, cartoon, cgi, render, illustration, painting, drawing, shadow, deformed eyes, hands, ((cute)), sexy, cleavage, ((unemotional))

Addressing Common Issues with Negative Prompts

Negative prompts are also decisive to obtain the best portraits. In this setup, they counteract unwanted styles like anime, cartoons, and over-rendered CGI looks. You might also need to use them to avoid any gender biases in outputs – while experimenting, I noticed that the model would give me portraits of women, even if I specified man in the positive prompt.

Technical Parameters

About parameters, here are the specifics:

  • Sampling Method: DPM++ 2M Karras
  • Steps: 30
  • CFG Scale: 4 (lower values can yield surprisingly good results)

Showcase of Generated Portraits

Each portrait here is just the raw output from Stable Diffusion. They are not perfect, but they are a good starting point; with face restoration and other edits with more traditional tools, the results can improve drastically. Have a look at Face Detailer and ADetailer to learn more.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.