Tips for Generating Realistic Human Images with Stable Diffusion
Here are some key tips I've learned after months of running SD.
- Prompt: Keywords like 'photorealistic, 8k, detailed skin texture' are essential.
- Negative Prompt: Adding 'cartoon, anime, deformed hands' makes the character look natural.
- Checkpoint: Realistic Vision V4.0 is the best.
- VAE: Switching to kl-f8-anime2 instead of the default improves color tones.
- Sampler: DPM++ 2M Karras offers the best balance of speed and quality.
Final tip: Use ControlNet to set the pose and Img2Img to refine details—the result will look like a real photo.
10 answers
Oh, this is really well-organized, I agree.
Realistic Vision is good, but why is the VAE kl-f8-anime2? Isn't that originally for anime? I find kl-f8-mse to be more versatile.
Putting 'photorealistic' in the prompt is really basic, lol
I've been running SD for 3 months now, and DPM++ 2M Karras is indeed the best. Others are either slow or produce weird results.
Do you have a source? When I put 'deformed hands' in the negative prompt, it fixes the number of fingers, but sometimes the arms get weird. Is it just me?
Wait, does changing the VAE really make that big of a difference? I've only been using the default... I should give it a try.
A friend of mine once tried using ControlNet for pose and then ran Img2Img, but it turned out too different from the original and failed. It would be great if you could share more tips.
Well, I feel like the 'photorealistic' keyword actually makes the results too stiff. If you want something natural, mixing in 'candid' or 'natural lighting' works better.
This is right lol. Thanks especially for the checkpoint recommendation. Try Juggernaut XL instead of Realistic Vision. The color tones are smoother.
I feel the same way. But for beginners, it's easier to just stick with the Sampler DPM++ 2M Karras. If you mess with the others, the results get distorted, so I gave up lol.