Lesson 4. Aspect ratio, seed, and quality parameters
Many users focus all their attention on the prompt and ignore parameters. But --ar (aspect ratio) and --seed are just as important as the prompt itself: they determine the format, reproducibility, and overall look of the result.
Topic breakdown
Aspect ratio is the ratio of image width to height. For an Instagram post — 1:1, for stories — 9:16, for a banner — 16:9, for a portrait — 4:5. The right ratio gives the correct format for the platform.
Seed is a random number, the 'key' to generation. With the same prompt and seed, you always get the same (or very similar) image. It's crucial for iteration and consistency.
Quality parameters (--quality, --stylize in Midjourney; CFG scale, steps in SD) control detailing and how strictly the prompt is followed.
This lesson uses Midjourney as a primary example, but the principles apply to all models.
What you'll learn
- Know aspect ratio parameters for different platforms
- Use seeds for reproducible results
- Adjust quality and stylize parameters according to the task
- Consistently modify parameters during the iteration process
Deep dive
Aspect ratio is not just the size of the picture, but also a composition decision. The 9:16 format pushes the model towards a vertical rhythm, while 16:9 better reveals the scene and atmosphere.
Seed is especially important in the workflow: if a client says 'we are keeping this option, but make the background warmer', you can return to the same composition and refine it, rather than starting over.
There is always a trade-off between quality and stylize. For e-commerce and precise product images, a lower stylize is usually better, while for art posters and atmospheric visuals, a higher one works better.
Lesson plan
Aspect ratio: correct format for the platform
Each platform requires its own format. Instagram feed 1:1 or 4:5, YouTube thumbnail 16:9, Pinterest 2:3. The correct ratio shows the image fully without cropping.
Seed: reproducible results and iteration
With a seed, you introduce minor changes to a single image. The base composition is preserved, only prompt changes affect the result.
Quality and stylize: balance of details and aesthetics
Quality increases detailing (but also generation time). Stylize sets how strictly to follow the prompt: low = literal, high = creative.
Parameter combination strategy
Change only 1-2 parameters simultaneously. This way you understand exactly what affects the result.
Weak vs strong prompt
a product photo of a perfume bottle
luxury perfume bottle on marble surface, studio lighting, reflection, high-end commercial photography, clean white background --ar 1:1 --q 2 --seed 42857
The second prompt sets the correct ratio (1:1 for Instagram), high quality (--q 2), and a seed for reuse.
Ready visual prompt template
Copy and adapt[main visual prompt] --ar [W:H] --seed [number] --v 6 --q 2 --stylize [0-1000]
Why it works
--ar 16:9 (landscape/banner), --ar 9:16 (stories/TikTok), --ar 1:1 (square post), --ar 4:5 (Instagram portrait).
--seed [number]: for consistency with the same prompt. --seed 0 — random, specific number — reproducible.
--q 0.25 (fast, few details), --q 1 (standard), --q 2 (high detail, slower) — Midjourney v5.
--stylize 0 (strictly follows prompt), --stylize 100 (standard), --stylize 1000 (strong model aesthetics).
In SD: steps (20-50), CFG scale (7-12 standard), sampler (DPM++ 2M Karras).
Practice
- Write a simple prompt: 'a mountain lake at sunset, photography'.
- First generate with --ar 1:1, then with --ar 16:9 and compare.
- Get the seed of the first generation (in Midjourney — via the envelope icon).
- Compare the same seed with --stylize 0 and --stylize 750.
Checklist
Common mistakes
- ignoring aspect ratio and getting the wrong format for the platform
- forgetting to save the seed and losing the base for iteration
- setting --stylize too high, straying from the prompt
- changing steps and CFG scale simultaneously — it becomes unclear what influenced the result
Lesson FAQ
How do I set the aspect ratio in DALL·E?
In the ChatGPT interface, you can choose 'square', 'landscape', 'portrait'. Via API, exact dimensions in pixels are set (e.g.: 1792x1024).
How many steps are optimal for Stable Diffusion?
For the DPM++ 2M Karras sampler, 20-30 steps are enough for good quality. Over 50 steps rarely yield noticeable improvement and only waste time.