Why does my AI-generated image have distorted hands and how can I fix it?

[qa_xcn]

I’m using Stable Diffusion to generate portraits, but consistently getting distorted hands in my results. Here’s an example:

[Image: A professional-looking portrait of a woman in business attire, but her left hand has six fingers and the right hand appears twisted at an unnatural angle]

My setup:

• Model: Stable Diffusion 1.5

• Sampling steps: 30

• CFG scale: 7

• Negative prompts: “bad anatomy, malformed hands, extra fingers”

• Size: 512x512 pixels

What I’ve tried:

• Adding “perfect hands, five fingers” to positive prompts

• Increasing sampling steps to 50

• Using hand-related negative prompts

The face and body proportions are good, but hands remain problematic. Is this a model limitation or am I missing something in my approach? What specific techniques or model versions work best for anatomically correct hands?

Tags: ai-image-generation, stable-diffusion, computer-vision, deep-learning, image-quality

[Georgi]

This is a universal model limitation due to the training process. Most of the images available are head and upper body shots and they are not anatomically correct 3D models just 2D images. There are few images just of the hands to train on. Dall-E 3 and GPT-4o seem to be doing the best at present.