Controlling Image Generation
Getting a good image from a generative model is not just about writing a prompt. You can control the output using parameters like negative prompts, seed, guidance scale, and more. This chapter explains the most important controls.
Prompt (The Main Input)
The prompt describes what you want. Be specific: "a red fox jumping over a log in a snowy forest, photorealistic" works better than "fox". Include style cues: "oil painting", "cinematic lighting", "minimalist".
Negative Prompt
Tell the model what to avoid. For example, if you want a clean human face, you might add: "blurry, deformed, extra limbs, bad anatomy, ugly". Negative prompts are very effective at removing artifacts.
Guidance Scale (CFG Scale)
This controls how strongly the model follows your prompt. Range is typically 1–20. Low (1‑4): creative, may ignore prompt. Medium (5‑9): good balance. High (10‑15): follows prompt strictly but may look over‑processed. For most uses, 7 is a good starting point.
Seed
A seed is a number that initializes the random noise. Same seed + same settings = same image. Change the seed to get a different variation. Use a fixed seed if you want reproducible results.
Sampling Steps and Sampler
Steps: number of denoising iterations. More steps = higher quality but slower. 20–50 steps are typical. Sampler: algorithm for denoising (e.g., Euler a, DPM++). Euler a is fast and good for most uses.
Image Size and Aspect Ratio
Stable Diffusion is trained on 512×512 images. You can generate larger (e.g., 768×768) but quality may drop. For wide images, use square then outpainting or generate with a model fine‑tuned for aspect ratios.
Example Settings for Beginners
- Prompt: "a cute cat sitting on a chair, cartoon style"
- Negative prompt: "blurry, ugly, deformed"
- CFG scale: 7
- Steps: 30
- Sampler: Euler a
- Seed: -1 (random)
Two Minute Drill
- Write clear prompts with style and detail.
- Use negative prompts to avoid unwanted features.
- Guidance scale (7 is good) controls prompt adherence.
- Seed fixes randomness for reproducibility.
- Adjust steps (20‑50) and sampler (Euler a) for quality.
Need more clarification?
Drop us an email at career@quipoinfotech.com