C.W.K.
Stream
Lesson 09 of 10 · published

Reading Outputs Diagnostically

~16 min · prompting, control, l9

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

피파 한 줄 정리: 실패 진단 4-카테고리: prompt 문제 / model 한계 / parameter 문제 / seed 운. 무작정 prompt 다시 쓰는 게 아니라 *어떤 카테고리인지* 먼저 정해.

Here's a skill that separates beginners from experienced practitioners: the ability to look at a generated image and diagnose what went wrong — and more importantly, whether it's a prompt problem, a model limitation, or a parameter issue. Different root causes require different fixes.

Think of it like being a doctor: a headache could be caused by dehydration (simple fix), a migraine (different treatment), or something more serious (structural problem). Prescribing the right treatment requires the right diagnosis.

The Diagnostic Framework

1. Is it a prompt problem?

Symptoms: The image is well-rendered but doesn't match what you wanted. Wrong subject, wrong mood, wrong composition, wrong style. The model clearly "heard" something different from what you intended.

  • Fix: Rephrase the prompt. Be more specific about what you want. Remove ambiguous terms.
  • Test: Try 3-4 different prompt phrasings with the same seed. If the same problem persists across all phrasings, it's probably not a prompt problem.
❌ Symptom: Wrong Subject

"Apple on a table" → Got an Apple computer on a table

✅ Fix: Disambiguate

"A fresh red Fuji apple on a wooden cutting board, kitchen setting, food photography"

2. Is it a model limitation?

Symptoms: The image has specific, recurring quality issues regardless of prompt changes — distorted hands, illegible text, wrong finger count, objects merging into each other. The model is generating its best attempt but the task is beyond its reliable capability.

  • Fix: Use a different model, use inpainting to fix specific regions, or accept the limitation and work around it.
  • Test: Try the same concept across multiple models. If all models struggle, it's a fundamental limitation of current technology. If only your model struggles, consider switching.

3. Is it a parameter problem?

Symptoms: Colors are oversaturated or washed out. Image looks plasticky/over-refined or too rough/noisy. Composition is okay but the "feel" is wrong.

  • Fix: Adjust guidance scale (lower for more natural, higher for more prompt-adherent), sampling steps (more for refinement, fewer for speed), or try a different sampler.
  • Test: Keep the same prompt and seed, vary one parameter at a time. You'll quickly see what each parameter does.

4. Is it a seed problem?

Symptoms: Sometimes the prompt works great, sometimes it doesn't. The quality is inconsistent across generations.

  • Fix: Generate more variations (batch generation). Not every seed produces a winner — that's expected.
  • Test: Generate 8-16 images with the same prompt but different seeds. If some are great and some are bad, it's just natural variation — curate the good ones.

Common Failure Modes and Their Causes

What You SeeLikely CauseFix
Wrong subject / stylePrompt ambiguityRephrase with specifics
Distorted hands/fingersModel limitationInpaint or use specialized model
Garbled text in imageModel limitationUse FLUX (better text) or add text in post
Oversaturated / plastickyGuidance too highLower CFG scale to 5-7
Doesn't match prompt at allGuidance too low or prompt conflictRaise CFG or check for conflicts
Too generic / boringUnderspecified promptAdd lighting, camera, specific details
Muddled / confused compositionOverspecified promptSimplify to 3-4 key elements
Good sometimes, bad sometimesNatural seed variationBatch generate and curate
Key Takeaways
  • Diagnose failures into four categories: prompt problem, model limitation, parameter issue, or seed variation.
  • Each category requires a different fix — rewriting the prompt isn't always the answer.
  • Consistent failures across seeds = prompt or model issue. Inconsistent results = seed variation (batch generate more).
  • Systematic diagnosis is faster than random prompt iteration.

External links

Exercise

최근 실패 generation 3개 골라. 각각 4-카테고리 진단 (prompt·model·parameter·seed) 돌려. 카테고리와 *first try*했어야 할 specific fix 적기.

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.