Stable Diffusion - Generate images locally with prompt examples

With this post I want to give you an idea how a prompt can be designed to generate nice images with the chilloutmix_NiPrunedFp32Fix.safetensors and modernDisney_v1.ckpt net. It is a suggestion and of course can be customized. Writing the prompts is not so easy at the beginning to generate something neat as a result and needs some practice. But once you have noted down the appropriate commands and combined them, it is fun. But it’s really fun when the waiting time for the individual attempts or images is not too long and an NVIDIA GPU is used that has the power to generate the images quickly. I always recommend a GPU with at least 12GB RAM and maybe in the class of a RTX30XX.

Table of Contents

Why do generated hands usually look funny?

Now, of course, there are countless trained networks that can also all generate beautiful images if the prompt fits. But now it is also so that these beautiful pictures always show e.g. from a person a portrait but rarely the variety of finders these can represent. Our fingers are so mobile and when we grasp, show these move then in the pictures rarely all are to be seen, they grasp partially over each other or into each other etc. and exactly this makes it difficult with the training of neural networks. There are simply too few pictures of hands in everyday situations that are also complex to learn for a neural network.

Here is an example of an image of the game character Duke Nukem 3D that I generated. The left hand holding the gun just doesn’t look good and shows too many fingers. You will notice that as soon as you look at the picture. Otherwise, I would say that the character and the background are very well done.

Duke Nukem 3D bad fingers

A very nice video brings it in a few minutes exactly to the point what is the problem here. I insert it here once and recommend to watch it. Because then we already know why the following prompts deal in part strongly with fingers and their correct generation. For us humans, it is simply that we quickly perceive misalignments in fingers or hands as disturbing and therefore it is still difficult to show fingers and hands in generated images.

AI gnerated hands

Generate images

But now to the actual topic of my contribution here the prompt write. What is a good, right, bad… prompt depends on your own expectations. But I try to give you a little help how to write prompts.

Note: It is important to understand that the prompts behave depending on the version of a network or generally depending on the used network. So it is best to save the tried prompts for the respective net in a table and evaluate them subjectively. This is the quickest way to achieve good results that meet your expectations and you can also generate them repeatedly.

The positive prompt: This describes exactly what you want to see.
The negative prompt: This describes what you don’t want to see.

Military soldier

Here is the configuration or parameterization of the neural network used.

Description:	Value:
Stable Diffusion checkpoint:	chilloutmix_NiPrunedFp32Fix.safetensors
Sample steps:	25
Sampling method:	DPM++ SDE Karras
Barch count:	4
Batch size:	1
CFG	12
Seed:	-1

Here is a sample image that I was able to achieve with the prompt described below and the settings mentioned above.

Stable Diffusion – military officer (male)

Positive Prompt:

male officer in military uniform, (masterpiece, best quality:1.2), ultra high res, (photorealistic:1.4), detailed skin, muscular, perfect male figure, full body, cinematic lighting, friendly, intelligent, conversation engaging, happy, kind, energetic, cheerful, creative, with sparkling eyes and a contagious smile, gun

Negative Prompt:

leather, net lining, mesh lining, paintings, big head, sketches, naked, (worst quality:2), (low quality:2), (normal quality:2), lowres, bad anatomy, bad hands, normal quality, ((monochrome)), ((grayscale)), futanari, full-package_futanari, newhalf, collapsed eyeshadow, multiple eyeblows, pink hair, ng_deepnegative_v1_75t, skin spots, acnes, skin blemishes, age spot, glans, extra fingers, ((extra arms)), (extra legs), mutated hands, (fused fingers), (too many fingers), (long neck:1.3)

Duke Nukem 3D

A game I used to play a lot was Duke Nukem 3D and of course I followed the stories and projects around this game figure on the internet for the last 20 years. So I found it very exciting how this action figure will look like when I generate it.

Description:	Value:
Stable Diffusion checkpoint:	modernDisney_v1.ckpt
Sample steps:	25
Sampling method:	DPM++ SDE Karras
Barch count:	4
Batch size:	1
CFG	12
Seed:	-1

Duke Nukem 3D

Positive Prompt:

modern disney style of Duke Nukem 3D, (masterpiece, best quality:1.2), granad launcher, ultra high res, (photorealistic:1.4), detailed skin, muscular, perfect male figure, full body, cinematic lighting, angry face, energetic

Negative Prompt:

duplicate, bad anatomy, bad hands out of frame, two heads, totem pole, several faces, extra fingers, mutated hands, (poorly drawn hands:1.21), (poorly drawn face:1.21), (mutation:1.331), (deformed:1.331), (ugly:1.21), blurry, (bad anatomy:1.21), (bad proportions:1.331), (extra limbs:1.21), cloned face, mutated legs, deformed legs, bad legs, out of frame, two heads, extra hands, extra feet, distorted fingers, mutilated, poorly drawn hands, poorly drawn face, deformed, bad proportions, extra limbs, cloned face, disfigured, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, (disfigured), (bad art), (deformed), (extra limbs), (disfigured), (deformed), (extra limbs), ((out of focus body)), ((out of focus face)), ((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), [out of frame], extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), out of frame, ugly, extra limbs, (bad anatomy), (malformed limbs), ((missing arms)), ((missing legs)),(((missing feet))), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), ugly, missing fingers, extra digit, fewer digits, out of focus, wrong eyes, wrong iris, wrong eyelids

Cartoon style princess

Of course, my little daughter also loves cartoons and princesses. With the right prompts, you can of course generate images here as well. I don’t want to discuss their value for a small child and how useful that is here. But she was happy of course and the more colorful the hair was the better the pictures were received.

Here is the configuration or parameterization of the neural network used.

Description:	Value:
Stable Diffusion checkpoint:	modernDisney_v1.ckpt
Sample steps:	25
Sampling method:	DPM++ SDE Karras
Barch count:	4
Batch size:	1
CFG	12
Seed:	-1

Here is a sample image that I was able to achieve with the prompt described below and the settings mentioned above.

beautiful princess

Positive Prompt:

A beautiful picture of a Princess, masterpiece, (photorealistic:1.4), perfect female figure, full body, from front, tropic background, long hair, slim, A beautiful, perfect face, perfect eyes, cinematic female flying archangel queen marilyn monroe face, fantasy sea landscape, fantasy magic, short aqua blue black fade hair, dark light night, intricate, elegant, sharp focus, illustration, highly detailed, digital painting, concept art, matte, art by Dali, masterpiece

Negative Prompt:

paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, ((monochrome)), ((grayscale)), Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, double head, deformed face, ugly face, disfigured face, mutated hands and finger, out of frame, ((((mutated hands and fingers)))), (((out of frame)))

Summary

The beginning to generate beautiful images was not easy. But if something is beautiful or not is in the eye of the beholder. But with time and the collection of prompts you get better and better into the topic and the results get better. Very good prompts can always be found on the site Civitai if you click through the pictures and look at their configuration. Because I didn’t come up with all these combinations myself but looked at many pictures and their prompts. Here again the link to the page. I really enjoy the topic of stable diffusion and now I would like to go a step further and in one of my next posts I will talk about how such a network can be further trained with your own images.