Stable Diffusion 3 is out. This is what we discovered.

Perfect the images you create
with Stable Diffusion

Magicflow lets you experiment with hundreds of prompts, and evaluate thousands of images to discover optimal settings for aesthetic visuals

Evaluating Stable Diffusion 3 (SD3) Large and Medium, and SDXL with a Body Positions Prompt Challenge

In this article we evaluate how well SD3 Large, SD3 Medium and SDXL models handle different body positions. We're rating both the the anatomical correctness and the overall quality of the image.

TLDR: Our results show that for anatomical correctness none of the models perform well, but SD3 Large was slightly better. For overall image quality, SD3 Large is the clear winner.
sd3 sdxl sd1.5
Weekly insightful articles on new models tested against 100+ prompts

To conduct the test we used ChatGPT-4o to generate prompts for different body positions across various categories: sports, dance, everyday activities, fitness, martial arts, entertainment, and leisure.

We tested over 30 prompts, each prompt tests a specific challenge. We then rated all images, gathered insights and here are the results.

Challenge: Body Positions Prompt Challenges

Anatomical correctness
SD3 Large
SD3 Medium
SDXL
Overall quality
SD3 Large
SD3 Medium
SDXL
While none of the models does a great job, SD3 Large in the winner for anatomical correctness. SD3 Medium is arguably worse than SDXL. For overall image quality, SD3 Large is the clear winner.
The challenge was to generate images of people in basic and complicated body positions across various categories: sports, dance, everyday activities, fitness, martial arts, entertainment, and leisure.
Image Size:
__model
prompt
Stable Diffusion 3 Large
Stable Diffusion 3 Medium
Stable Diffusion XL
A Taekwondo practitioner in a high front kick, one leg extended upwards
A ballerina in a pirouette, one leg raised, arms gracefully curved
A band on stage, each member playing an instrument, lead singer mid-song
A basketball player in mid-air slam dunk, arms stretched towards the hoop
A boxer in a defensive stance, fists up, body slightly turned
A chef chopping vegetables, one hand holding the knife, the other holding the food
A couple in a salsa dance dip, one partner leaning back supported by the other
A cyclist riding a bike, bent forward, hands on the handlebars, legs pedaling
A family gathered around a table, each person engaged in conversation, some reaching for food
A fitness class doing aerobics, all participants moving in sync, arms and legs in motion
A football player mid-kick, leg extended, with focused expression
A football team huddled together, strategizing with intense expressions
A gardener kneeling on one knee, planting flowers with a small shovel
A group of dancers performing a synchronized routine, mid-jump, arms and legs
A group of friends having a picnic, some sitting on a blanket, others standing and chatting
A hip-hop dancer mid-breakdance move, hand on the ground, legs in the air
A martial artist mid-karate kick, one leg extended high, arms in guard position
A musician standing and playing an electric guitar, strumming with one hand
A person in a plank position doing a push-up, arms bent, body parallel to the ground
A person in a yoga warrior pose, one leg forward, arms extended parallel to the floor
A person sitting cross-legged on the floor, engrossed in a book
A runner sprinting, one leg extended forward, arms pumping
A singer holding a microphone with one hand, the other hand extended outwards
A swimmer diving into a pool, body arched, arms extended forward
A weightlifter in a squat position, holding a barbell across their shoulders
A woman lying in the grass
An actor in a dramatic pose, one hand on their heart, the other extended upwards
Two martial artists sparring, one mid-kick, the other dodging with raised arms
Two people standing side by side, one flexing their arm to show muscle definition, the other relaxed
To explore the entire image set we generated  visit the public folder page of this article.
* The SD3 images were generated using Stability.ai API (SD3 Large). The SDXL images were generated using the SD3 model hosted on Replicate (stability-ai/stable-diffusion-3). The SDXL images were generated using the SDXL model hosted on Replicate (stability-ai/SDXL).

Perfect your Gen-AI image quality with Magicflow

Use Magicflow to generate and review images in bulk, compare and optimize Gen AI pipelines with the necessary human touch for perfection.