Creator encompasses both Text2Image and Image2Image functionalities.
As the name suggests, Text2Image allows for the generation of images from text, instructing the AI on how to create your image based on the prompts you provide. On the other hand, Image2Image involves generating images from existing pictures, where you provide both an uploaded image and textual prompts to guide the AI in generating your desired image.
Text2Image
Prompt
This is where you describe the content you want in the generated photo. For example, if you want to generate an image of "a red-haired girl standing in a sunflower field," your prompt could be written as:
1girl, red hair, dark brown eyes ,little freckle, jacket with flower pattern, sunflowers field
ou can notice that besides describing the girl's hair, I've added more detailed descriptions like dark brown eyes, little freckle, and jacket with flower pattern. You are free to modify these descriptions, change their colors, or add more details as per your preference.
Negative Prompt
These are the prompts that specify what you do not want in the generated image. For instance, if you want to avoid images of low quality (worst quality, low quality) and you want to keep the generated images safe for work (NSFW), and you don't want poorly rendered hands (bad hand, extra hands, extra fingers, too many fingers, fused fingers), your negative prompt could be:
worst quality, low quality, NSFW, bad hand, extra hands, extra fingers, too many fingers, fused fingers
Base Model
This is primarily used to control the style of image generation. For example, you can select the "Marvel comics" model. Other options include "Realistic" (tending towards a realistic style), "CG" (computer-generated style), "Ghostmix" (excellent for generating detailed textures on characters' clothing), and "Others" (for exploration).
Avatar Model
Trained LORA models will be displayed here.
(Denoising) Strength
This parameter controls the degree to which elements from the original image are retained. A higher value makes the generated image less faithful to the original, while a lower value makes it more faithful. The extreme values represent complete retention and complete non-retention.
We can use the default value of 0.75 here.
Classifier-Free Guidance (Scale)
This represents the intensity of the prompts, indicating how much the AI should adhere to the provided prompts. Lower values make the AI follow prompts closely, while higher values lead to more random image generation.
We can use the default value of 7.5 here.
Number of Inference Step
This indicates the number of steps required to generate the image. More steps result in longer generation times. Typically, for good images, 25-35 steps are suitable, depending on the model or sampler used.
Size
This determines the output image dimensions.
Width: Width of the output image, in multiples of 64.
Height: Height of the output image, in multiples of 64.
💡 This option does not apply to the Image2Image feature.
Seed
This is the seed for generating the image. If the seed, Negative Prompt, and Prompt are the same, there's a high probability of generating nearly identical images. Keeping parameters the same will result in identical images.
Seed: 101565998
Randomize: Enabling this option generates a new seed automatically, ensuring each generated image is different.
Sampler
The sampler determines the method used for generating images.
Select DDIM
K_Eular: Suitable for obtaining relatively simple results.
DPMSolverMultistep: Aims for high-quality images without concerning whether the image converges.
After configuring the parameters as shown above, you should be able to find the image you just generated in your "Gallery".
If you've filled in all the parameters as per the example above, you should obtain an image similar to the one I have.
Image2Image
The parameters for Image to Image are the same as those for Text to Image. You can refer to Text to Image for guidance.
To start, we need to upload a reference image and let's use the following image as an example:
Prompt: ust like in the Text2Image, fill in your prompts according to your idea. For example, you can use the prompts below:
flowers, light green hair, brown eyes, laugh
Negative Prompt: Since there's not much details in the reference image (there is no hands, no background etc.), we only need to avoid low-quality results and ensure the image is safe for work:
worst quality, low quality
Base Model: Marvel comics
(Denoising) Strength: 0.75
Classifier-Free Guidance (Scale): 7.5
Number of Inference Steps: 30
Seed: 20366705
Sampler: DDIM
Once you've filled in the parameters as described above, please wait for a moment, and you should be able to see the image below: