Stable Diffusion 3 from Stability AI has now been released and is available for public use. However, they currently only offer a REST API to test these models. Therefore, I tried using this on my local machine using Node.js code.
import fs from "node:fs";
import axios from "axios";
import FormData from "form-data";
const formData = {
prompt: "prompt",
output_format: "png",
aspect_ratio: "16:9",
seed: "645,508,968",
};
const response = await axios.postForm(
`https://api.stability.ai/v2beta/stable-image/generate/sd3`,
axios.toFormData(formData, new FormData()),
{
validateStatus: undefined,
responseType: "arraybuffer",
headers: {
Authorization: `TOKEN`,
Accept: "image/*",
},
},
);
if (response.status === 200) {
fs.writeFileSync("./output.png", Buffer.from(response.data));
} else {
throw new Error(`${response.status}: ${response.data.toString()}`);
}
If you prefer testing it without code, you can check and use your token wisely. Fireworks Playground
What is seed?
In Stable Diffusion, a seed
is a number used to initialize the image generation process. It acts like a starting point for the AI to determine the random noise it uses to build the image. Here's a breakdown of how it works:
Random by default: If you don't specify a
seed
, Stable Diffusion will automatically choose a random seed number for each image generation. This results in completely different images every time you run the program with the same prompt.Replicability with specific seed: However, if you set a specific
seed
value, Stable Diffusion will always use that sameseed
number to generate an image whenever you provide the same prompt along with that seed. This allows you to reproduce the same image again.Making minor adjustments: You can use the same
seed
as a base and then tweak other settings in Stable Diffusion to see how they affect the image while keeping the overall content consistent.Sharing results: Sharing the
seed
value along with your prompt allows others to recreate the exact image you generated, which can be useful for collaboration or tutorials.Seed and image similarity: Interestingly,
seeds
that are close in value tend to generate images that share some similarities. This can be helpful if you're looking for variations on a theme. For example, using aseed
value that's one number higher or lower than another seed might generate an image with a similar composition but with slight differences in details.
In summary, the seed
value in Stable Diffusion acts like a recipe for generating random noise, which the AI then uses to build an image based on your text prompt. By controlling the seed
, you can influence the reproducibility and explore variations of the generated images.
To generate a similar image, we need to utilize the seed
parameter; using the same seed will produce similar images.
Here, I'm trying to create a prompt for a countryside scene across four seasons and also a picture with a weather ambience vibe, where the result is not exactly the same but produces a similar output in terms of how the landscape looks.
Generating Similar Image (4 different season)
seed -> 645,508,968
Spring Prompt
A countryside village with a clear blue sky, surrounded by sakura trees, in an anime art style.
Summer Prompt
A countryside village with a clear blue sky, surrounded by green trees, in an anime art style.
Autumn Prompt
A countryside village with a clear blue sky, surrounded by trees that are falling down in the autumn season, in an anime art style.
Winter Prompt
A countryside village with a clear blue sky, surrounded by a snow, in an anime art style.
Generating Similar Image (ambience vibes)
Seed -> 358,804,591
Default Prompt
In a dark alley illuminated by a neon light, the scene exudes Tokyo city vibes. It's foggy and ultra-realistic in 4K resolution. A cat, not facing the camera, sits in the middle of the alley street, with a bokeh focus on the cat. A vending machine is mounted on the wall.
Snow Ambience Prompt
In a dark alley illuminated by a neon light, the scene exudes Tokyo city vibes. It's foggy and ultra-realistic in 4K resolution, with a snowstorm and a thick layer of snow covering the street. A cat, not facing the camera, sits in the middle of the alley street, with a bokeh focus on the cat. A vending machine is mounted on the wall.
Rain Ambience Prompt
In a dark alley illuminated by a neon light, the scene exudes Tokyo city vibes. It's foggy and ultra-realistic in 4K resolution, with a heavy rainstorm creating a muddy environment. A cat, not facing the camera, sits in the middle of the alley street, sharply in focus with a bokeh effect. A vending machine is mounted on the wall.
Summary
While the output for each image is not exactly the same, it already has the same landscape structure if we use the same seeds and a similar prompt, with some customization added to create a new scene or ambience.
Resources
- Stable Diffusion
- Stable Diffusion 3 API References
- Stable Diffusion 3 Research Paper
- Stable Diffusion in AWS Bedrock
- Fireworks Playground
- Seed Explanation by Segmind blog
- Seed Explanation by Wolfe
- Seed Explanation by getimg.ai
- Seed Explanation from springer
All the images here were generated by Stable Diffusion 3, including the cover image.
Top comments (0)