C.W.K.
Stream
Lesson 05 of 10 · published

Noise, Randomness, Seeds, and Variation

~16 min · diffusion, latent-space, l5

Level 0Spark
0 XP0/100 lessons0/14 achievements
0/200 XP to next level200 XP to go0% complete

피파 한 줄 정리: Seed = 시작 noise의 DNA. Same seed + same prompt + same params = 같은 image. 이게 reproducibility의 기반이고, batch generation의 무기야.

Here's something that trips up beginners: you type the exact same prompt twice and get two completely different images. Is the model random? Unpredictable? Broken? No — it's working exactly as designed. The randomness is a feature, and once you understand how it works, you can use it strategically.

The Starting Noise Is the "DNA" of the Image

Remember: generation starts from random noise. That initial noise pattern determines which particular image emerges from the denoising process. Think of it like genetics: two genetically different seeds planted in the same garden (same prompt, same model) grow into different but related plants. The seed is the number that determines the initial noise pattern.

Same prompt: "a red fox in snow, oil painting"

Seed 42:     🎲₄₂ ──▶ [Denoise...] ──▶ 🖼️ Fox facing left, soft light
Seed 123:    🎲₁₂₃ ──▶ [Denoise...] ──▶ 🖼️ Fox facing right, dramatic light
Seed 7777:   🎲₇₇₇₇ ──▶ [Denoise...] ──▶ 🖼️ Fox mid-leap, warm tones

Same prompt → different seeds → different images
Same seed → same image (reproducible!)

What Seeds Control

In practice, a seed is just an integer (like 42 or 1234567) that initializes a random number generator. This generator produces the noise pattern. The critical property: the same seed always produces the same noise, which means the same seed + same prompt + same parameters = the same image, every time.

Practical Seed Strategies

  • Random seeds (default): Each generation explores a different point in the output space. Use this for broad exploration when you're looking for inspiration or haven't settled on a direction.
  • Fixed seed + varying prompts: Keep the same seed, tweak the prompt. The overall composition stays similar but details change. Great for iterative refinement: "I like the composition of this image, but I want different lighting."
  • Same prompt + varying seeds: Generates multiple interpretations of the same concept. This is batch generation — produce 4, 8, or 20 versions and pick the best. This is arguably the single most underrated technique in practical image generation.

Why Variation Matters

Variation isn't a problem to solve — it's a creative tool. Professional photographers don't take one shot and go home. They take hundreds and select the best. Generative AI works the same way: each seed is a different "shot" of your concept.

Key Takeaways
  • The random seed determines the starting noise, which determines which specific image emerges.
  • Same seed + same prompt + same parameters = identical output (reproducibility).
  • Batch generation (many seeds, same prompt) is the most practical way to find great results.
  • Variation is a creative tool, not a flaw — generate many, curate the best.

Code

예시 코드·python
# Pseudocode: batch generation with different seeds
prompt = "a cozy bookshop interior, warm lamplight, rain on windows"

for seed in [42, 123, 456, 789, 1024, 2048, 4096, 8192]:
    image = generate(prompt=prompt, seed=seed, steps=30)
    save(image, f"bookshop_seed_{seed}.png")

# Now review all 8 and pick the best composition

External links

Exercise

Seed 하나 고정, generate. Seed 고정한 채 prompt 살짝 변경 → 거꾸로 prompt 고정한 채 seed 변경. 어느 dimension이 더 dramatic? 어디서 unrecognizable해지나?

Progress

Progress is local-only — sign in to sync across devices.
이 페이지에서 버그를 발견하셨거나 피드백이 있으세요?문제 신고

댓글 0

🔔 답글 알림 (로그인 필요)
로그인댓글을 남기려면 로그인해 주세요.

아직 댓글이 없어요. 첫 댓글을 남겨보세요.