Seeds
In the context of AI generation models used in Fater, a Seed is a number that initializes the random process used to create an image, asset, or video. Think of it as the starting point for the AI's "creativity."
How Seeds Affect Generation
AI models rely on controlled randomness to generate unique outputs even from the same prompt. The Seed value directly controls this randomness:
Same Prompt + Same Parameters + Same Input Image + Same Seed = Same Result: If you use the exact same text prompt, negative prompt, model settings (like Guidance, Steps, etc.), and the exact same numerical Seed value, the AI will produce virtually the identical output every time.
Same Prompt + Same Parameters + Same Input Image + Different Seed = Different Result: Changing only the Seed value, while keeping everything else the same, will result in a different output that still adheres to your prompt but varies in composition, details, or specific arrangement.
Using Seeds in Fater
Finding the Seed Input: The Seed parameter is typically found in the Left Sidebar of the relevant Editor Mode (Image, Asset, Video) when an AI model is selected, often near the "Number of Generations" control.
Random Seed (-1): By default, the Seed is often set to
-1. This tells Fater to pick a random seed for each generation. This is great when you are exploring initial ideas and want maximum variety with each click of the "Generate" button.Specific Seed (Locking): You can enter any specific positive integer (up to a very large number) into the Seed input field. This locks the seed. (Only when performing a single generation, otherwise it's always random seeds)
Benefits of Locking a Seed
Using a specific, locked seed value is a powerful technique for iterative refinement and control:
Reproducibility: If you generate an output you really like, noting down its specific Seed value allows you to recreate that exact result again later, provided all other settings remain the same. (Fater stores the seed used in the layer/asset/video parameters, see Restoring Layer Generation Configuration).
Controlled Iteration: This is the most common use case. Once you find a generation with a composition or core subject you like (even if other details aren't perfect), lock its seed. Now, you can:
Tweak Prompts: Slightly modify your positive or negative prompts (e.g., add details, change colors, remove unwanted elements). Since the seed is locked, the overall structure should remain similar, but the AI will attempt to incorporate your prompt changes.
Adjust Parameters: Experiment with changing parameters like Guidance, Steps, or model-specific controls. Locking the seed isolates the effect of these parameter changes on the existing composition, rather than generating a completely new image each time.
At each try, don't forget to hide the previous results otherwise the sent image input will be different and so will produce completely different results.
Seed Hunting: Exploring Variations
Seed hunting is the process of systematically exploring different Seeds while keeping your prompt and other parameters constant.
Purpose: When your prompt is good but you haven't found a specific composition or layout you like yet, seed hunting lets you see different random variations that all match your core description.
How:
Set your desired Prompt, Negative Prompt, and AI Parameters.
Generate an image (or multiple images if
Number of Generations> 1) with the Seed set to-1(Random).Review the results. If you see one with potential, find its specific Seed number (stored in its parameters).
Enter that specific Seed into the Seed input field to lock it for further refinement (as described under "Controlled Iteration").
If none of the initial results are quite right, you can either generate again with a random seed, or manually enter different seed numbers (e.g., incrementing the last used seed by 1) to explore nearby variations systematically.
By understanding and utilizing Seeds, you gain significant control over the AI generation process, moving from random exploration to precise refinement and reproducible results.
Last updated