khaled.uno

April 30, 2026 · by Khaled Uno

Prompt engineering glossary

Terms, techniques, and traps of the trade when writing prompts for modern text-to-image models.

prompt-engineeringglossaryai

If you've recently started writing prompts for text-to-image models, here's a quick glossary of the terms you'll see everywhere.

Negative prompt

List of concepts to exclude from the image. On SDXL and SD3 it works wonders: blurry, watermark, deformed hands, low quality removes many artifacts. On FLUX schnell the effect is marginal — the model is barely "guidable."

CFG scale (Guidance scale)

How strictly the model adheres to the prompt. Low values (1-3) produce more creative but less faithful images; high values (8-15) follow the prompt literally but cause unnatural saturation. Sweet spot for most models is 6-7.

Seed

Number that initializes the random generator. Same prompt + same seed = same image. Useful when you find a composition and want to tweak only one detail: change the prompt and keep the seed.

Inference steps

Number of denoising passes. More steps = more time + more detail. FLUX schnell is optimized for 1-4 steps; SDXL prefers 30-50.

Style suffixes

Standard phrases to append to the prompt to "push" the style:

  • cinematic, 35mm, anamorphic — film look
  • studio photography, soft light, depth of field — pro photo
  • oil painting, impasto, brush strokes visible — oil painting
  • concept art, trending on ArtStation — fantasy/sci-fi concept
  • octane render, unreal engine 5 — 3D render

Token budget

Models have a token limit (FLUX accepts ~256). Trimming redundancies ("very very beautiful gorgeous stunning") often improves coherence.

Composition and camera

Terms the model understands surprisingly well:

  • wide shot, close-up, medium shot, extreme close-up
  • low angle, high angle, dutch angle, bird's eye view
  • rule of thirds, leading lines, symmetric composition

ControlNet (advanced)

Lets you guide generation with a reference image (sketch, depth map, openpose). Not available on standard FLUX schnell, but some Replicate variants support it.

LoRA

Lightweight adapters that modify a base model's behavior. There are LoRAs for specific styles, characters, objects. FLUX has a growing ecosystem.

Going back to fundamentals always helps: concrete, descriptive, hierarchical prompt (subject → environment → style → technique). Everything else is detail.