Stable Diffusion

Generate art images from text Stable Diffusion is a latent text-to-image diffusion model. It creates photo realistic and artistic images

Fast generations

Realistic and artistic images

Open source model

Regularly updated dataset

Up to 1280x1280

Try in Phygital+ for free

Quick navigation. Click on these buttons 👇

Parameters description Custom checkpoints Tips for good results Works gallery

How to work with it

1. Type your idea

Neural network creates images based on your text. This text is called prompt. The more precise your prompt is, the better the results will be

2. Choose seed

Seed is an image from which the generation starts. By changing the seed number you can significantly improve and change the quality of the final AI image. The same seed should give you the similar result

3. Set resolution

With these parameters you can change width and height of the final result.

4. Change steps

Steps define how detailed and how close to the prompt the results will be. The more steps, the better the final image is, but it also will take longer to create one image.

5. Launch

Press the green button to launch the AI magic process. You can download the results immediately or upscale it to achieve higher resolution

Input

Text

Output

Image

Main task

Create images, create art

Neural network

Stable diffusion

Release date

18.08.2022

License

CreativeML Open Rail M

Parameters

Steps

The range: from 1 to 500 It defines how detailed and close to the prompt your image will be. The higher number of steps is, the longer it takes to generate one image. Good results can be achieved in 100-250 steps Usually the more steps, the better, but we recommend to experiment with 100 steps, as it's the most optimal time spent and quality correlation

Seed

The range: from 1 to 4294967296 The generation starts from a particular image (or noise). The same seed number will give you similar results each time, whereas by changing it to another number may dramatically influence the final output. We recommend to use different seed numbers if you're experimenting, and as soon as you find the perfect result

Width and height

Range: from 512 to 1024 (one step is 64 pixels) It defines the width (x) and height (y) of the final image. We recommend to use horizontal orientation for landscapes and environments, vertical for characters and creatures

No (optional)

A text field Negative prompt. It tells neural network what it should NOT generate

Artistic Mode

A checkbox It turns any generation into the artistic masterpiece. Currently available at 4 styles: • General Artstic mode: for anything from objects to personas • Portrait: for personas and any humanoid-like creatures • Full-body: for full body height images of humanoid-like creatures • Landscapes: for cinematic scenes

Scale uncond (CFG)

The range: from 0 to 100 It defines how close the image to the initial seed will be. It also Impacts on image contrast. The higher the number is, the less detailed, more blurry and more vibrant the image will be. We recommend to keep it in the range of 7,5 to 15

Eta

The range: from 0 to 1 Impacts on image depth and color range We recommend to keep it around 0.8 so that the picture is more balanced.

Samplers

There are several samplers It defines how AI will generate the image. here is also almost no difference in the output image depending on which sampler you use. We recommend to change it only if you're an experienced user

Start image skip

Range: from 0 to 0.9 It defines how close to the init_image the final image will be. The higher the number is, the more different to the Start image a generated artwork will be. We recommend to keep it between 0,7-0,8 if you use Start image

Number of images

Range: from 1 to 4 Create several images from one prompt. You can take this one image

Blur mask edges

Range: from 0 to 2 It defines the blendability of mask when you inpaint objects.The less number means the edges will be more sharp. Higher number can give you more blend between edges of inpainted object and Start image Usually the results are not noticeable, so we recommend to change it only if you're an experienced user

Getting bad results?

Try one of these!

Use Artistic Mode

It will turn make a beautiful artistic image from any prompt, including very short ones Try turning it on, if it's not on already

Change resolution

A square image doesnt always give a good result. Try adjusting the size for the image of: • characters: vertical orientation (e.g, 512x768) • landscapes: horizontal orientation (e.g., 768x512)

Use more specific keywords

By adding specific words, you can make the image look better and fit your expectations Stable Diffusion by default generates more realistic images, so it can look a bit creepy. We advise to use words: • art-like pictures: trending on artstation, beautiful artwork, fantasy art, devianart, concept art • for more details: ultra detailed, super realistic, 8k, intricate, high definition, detailed textures • for landscapes: cinematic scene, rule of thirds, picturesque scene • for characters: character design, a portrait of ... by Ilya Repin, full body, character concept, high facial details, glossy eyes, sharp focus, smooth face • for 3D looking characters: octane render, unreal engine Try using lexica.art or promptomania.com for references

Change seed number

The seed has a great influence on final result, because it is the starting point for generation process. Try different numbers, the range for the number is from 1 to 42 million

Change eta

Sometimes changing the eta parameter can give you a better result Try doing 0.6 instead of 0.8

Use 'Do not make' field

Try removing unneccesary details Write down what you do NOT want to generate It can help a lot with results

Submit your work