스킬 · 이미지·렌더

미드저니 프롬프트 엔지니어링

미드저니 프롬프트 작성, --sref·스타일 코드 선택, 결과 채점과 반복 개선을 돕는 원작 스킬. 원작자 Justin Perea, MIT 라이선스.

0번 설치됨
스킬스크립트 포함승인됨원작 Justin Perea · MIT · 올림 feeeldstudy · 판 1Claude · ChatGPT · Codex · Cursor
midjourney-prompt-engineering/knowledge/failure-modes.md
# Failure Modes & Refinement Strategies> Common problems encountered in Midjourney generation and their fixes.> Auto-generated from iteration logs and patterns database during reflection.> Last updated: 2026-02-09## How to Use This FileWhen a generation fails or doesn't match intent, check this file for known failure modesthat match the symptoms. Each entry includes:- **Symptom**: What you see in the output- **Cause**: Why it happens- **Fix**: What to change in the prompt/parameters- **Evidence**: Which sessions demonstrated this- **Confidence**: How well-tested the pattern is---## Diagnostic FrameworkWhen results don't match expectations, work through these decision trees by dimension.### Subject Accuracy| Issue | Likely Cause | Fix ||-------|--------------|-----|| Wrong subject entirely | Ambiguous prompt, subject buried in middle | Front-load subject, be more specific || Missing key elements | Too many concepts competing | Simplify, focus on essentials || Extra unwanted elements | MJ interpretation, no exclusions | Add `--no [unwanted elements]` || Wrong style of subject | Style bleeding into subject | Use `--style raw`, separate style from subject |### Composition| Issue | Likely Cause | Fix ||-------|--------------|-----|| Subject in wrong position | No position instructions | Add explicit: "subject at far right" || No text space | "Negative space" misinterpreted | Describe empty areas physically || Too zoomed in | No composition guidance | Add "wide shot", "extreme wide angle" || Too zoomed out | Subject description too brief | Add details that demand closer framing || Collaged/arranged layout instead of isolated elements | Aesthetic/movement keyword triggering layout | Remove "punk zine aesthetic" etc., use material keywords instead || 3/4 angle when frontal wanted | MJ default perspective bias | Add "flat frontal view" + "orthographic perspective" || Element position wrong on object face | Sub-object positioning unreliable in V7 | Generate batches and select, don't iterate on position wording || Continuous pattern instead of isolated forms | Pattern/swirl keywords fill the frame | Remove "all-over pattern", describe "scattered discrete shapes floating in dark void" || Gradient direction wrong (horizontal instead of vertical) | No directional language in prompt | Use "from [color] at top to [color] at bottom". Avoid standalone "vertical" (triggers stripe patterns). |### Mood/Atmosphere| Issue | Likely Cause | Fix ||-------|--------------|-----|| Too cheerful/bright | Lighting keywords wrong | Add "moody", "dramatic", specific lighting || Too dark/ominous | Missing warmth keywords | Add "warm", "inviting", adjust lighting || Generic/flat feeling | Low stylize, no atmosphere | Increase `--s`, add atmosphere terms |### Color| Issue | Likely Cause | Fix ||-------|--------------|-----|| Wrong color palette | Not specified | Explicitly state colors in prompt || Too saturated | MJ default beautification | Use `--style raw`, add "muted" || Clashing colors | Too many color mentions | Limit to 2-3 key colors || Color drifting (blue->teal, warm->yellow) | Generic color names | Use specific pigment names: "cobalt blue", "burnt sienna", "warm ochre" || Low contrast, lines blend into gradient | No explicit ground specified | Add "pure black canvas", "against dark void", "high contrast linework" || Warm color contamination (sunset orange/pink at bottom) | Sky/atmospheric metaphors or MJ default | Add `--no warm, orange, yellow, pink, sunset` |### Quality/Realism| Issue | Likely Cause | Fix ||-------|--------------|-----|| Too artificial/AI-looking | High stylize | Lower `--s`, add `--style raw`, add "photorealistic" || Too realistic when artistic wanted | Low stylize, raw mode | Increase `--s`, remove `--style raw` || Blurry/soft | Quality parameter | Ensure `--q 1`, add "sharp", "detailed" || Photorealistic when flat/graphic wanted | Physics keywords in prompt | Remove "subsurface scattering", "polished smooth surface"; use "fine art print" || 3D fluid ribbons instead of fine linework | Water/fluid simulation keywords | Remove "swirling water vortex"; use "fine flowing lines", "thin ink lines" || Thick painterly marks instead of fine lines | "Brushstrokes" keyword | Use "thin ink lines", "hair-thin strokes" instead of "fine delicate brushstrokes" |### Gradients/Abstract| Issue | Likely Cause | Fix ||-------|--------------|-----|| Painted swatches instead of digital gradient | "Color swatch", "wash", "matte finish" keywords | Remove paint-related vocabulary; use "smooth color transition", "abstract minimalist background" || Lens glow/aurora instead of smooth gradient | "Defocused" keyword | Use "smooth", "seamless", "soft blend" instead || Gradient banding/harsh transitions | --s value too low | Increase to --s 75-100 for smoother blending || Unwanted subjects in abstract output | Weak negative prompt | Use comprehensive --no list (15+ items) AND "no subject" in prompt text || Gradient runs wrong direction | No directional language or standalone "vertical" | Use "from [color] at top to [color] at bottom" phrasing || Fabric/silk instead of gradient | Warm fabric-associated color names (peach, coral, champagne, blush, soft gold) | Use cool colors, or add `--no silk, satin, fabric, folds, textile` |### Surrealism in Photorealistic Scenes| Issue | Likely Cause | Fix ||-------|--------------|-----|| Surreal modifier treated as mood, no visual change | Abstract surreal words ("impossible reflections", "melting buildings") ignored by --style raw | Describe a concrete impossible OBJECT (whale, paper cranes, koi fish) instead of an impossible EFFECT || Photorealistic base degraded by surreal technique | "Double exposure" conflicts with --style raw anchor | Remove technique keywords; use concrete impossible objects placed in the scene || Surreal element too subtle / absorbed into atmosphere | Element too small (human-scale) or color-mismatched with scene | Use massive scale or multiples (flock/school). Match element color to scene palette. || Surreal element visible but lacks impact | Single object at human scale | Use dense multiples (hundreds of koi, flock of cranes) or massive scale (whale above rooftops) || Surreal element clashes with color palette | Cool bioluminescent element in warm scene, or vice versa | Choose elements whose natural colors complement the scene palette (golden koi in warm neon scene) |---## Quick Fixes Reference Card| Symptom | Quick Fix ||---------|-----------|| Too busy | `--style raw --s 150 --no clutter` || Not following prompt | Front-load subject, lower `--s` || Wrong position | "subject at far [direction]" || Unwanted text | `--no text, letters, words, writing` || Too generic | `--weird 100`, add specific details || Wrong colors | Specific pigment names (`cobalt blue`, `burnt sienna`), `--no [wrong color]` || Color drifting | Use specific color anchors, not generic names || Warm color contamination | `--no warm, orange, yellow, pink, sunset` || Low contrast on dark ground | Add `pure black canvas`, `high contrast linework` || Too dark | "bright", "well-lit", remove moody terms || Too bright | "moody", "dramatic shadows", "dark" || AI-looking | `--style raw`, "photorealistic", lower `--s` || Not artistic enough | Higher `--s 300+`, remove `--style raw` || Photorealistic when graphic wanted | Remove physics keywords, add "fine art print", "graphic quality" || Collaged layout | Remove aesthetic/movement keywords, use material keywords || Sharp when blur wanted | Remove ALL subject keywords, use "out of focus photography" || 3D fluid instead of linework | Remove water/vortex keywords, describe mark-making: "thin ink lines" || Thick marks instead of fine lines | Replace "brushstrokes" with "thin ink lines", "hair-thin strokes" || Continuous pattern, no breathing room | Remove "all-over pattern", add "scattered discrete shapes", "empty space between" || Photorealistic when no medium specified | Add explicit medium: "spray paint airbrush illustration", "oil painting", etc. || Standard 2 eyes when multi-eye wanted | Use "almond eyes stacked vertically" + `--sref` + `--sw 200`. Never use "slits". || Sref not transferring structure | Increase `--sw 200+`. Default 100 is style-only. || Painted swatches instead of gradient | Remove "color swatch", "wash", "matte finish"; use "smooth color transition" || Lens glow instead of smooth gradient | Remove "defocused"; use "smooth", "seamless soft blend" || Gradient harsh/banding | Increase `--s` to 75-100. `--s 0` produces harsh transitions. || Wrong gradient direction | Use "from [color] at top to [color] at bottom". Avoid standalone "vertical". || Fabric/silk instead of gradient | Use cool color names. Avoid peach, coral, champagne, blush, soft gold. Add `--no silk, satin, fabric, folds, textile`. || Surreal words ignored in photo scene | Replace abstract surreal effects with concrete impossible objects || Double exposure kills photo look | Remove "double exposure"; use concrete surreal objects with --style raw || Surreal element invisible/absorbed | Increase scale (massive) or quantity (hundreds). Match colors to scene palette. |---## When to PivotSigns the current prompt direction isn't working:- 3+ generations all miss the mark- Key element consistently ignored- Results getting worse with refinements- Fighting against MJ's "instincts"Better to:1. Step back and re-read original intent2. Try completely different vocabulary3. Use a reference image (`--sref`)4. Break complex scene into simpler elements---## Session-Learned Failure ModesFailure modes are organized by confidence level: **Medium** (tested 2+ times with logged evidence) down to **Low** (single observation, needs validation).### MEDIUM CONFIDENCE (Well-Tested Patterns)#### Subject Keywords Override Abstract/Blur Effects**Pattern ID:** `auto-icm-subject-keyword-dominance`**Confidence:** Medium (0% success, 3 tests)**Symptom:** Any concrete subject keyword — even softened with "hint of", "faint", "dissolving" — produces sharp rendering instead of blur/abstract/defocus effects.**Cause:** V7 prompt hierarchy: subject/material > style/abstract > perceptual quality. Subject nouns re-anchor output in conventional sharp rendering, overriding blur keywords.**Fix:**- Do NOT attempt to soften subject keywords with qualifiers- Complete subject removal is the only reliable approach- Even "hint of dissolving botanical silhouettes" produced identifiable sharp stems and leaves (iter 7: blur scores dropped from 0.90-0.95 to 0.45-0.55)- If organic forms needed, achieve them through Vary Strong on an abstract parent, not prompt keywords**Evidence:** Strong failure mode from ICM session 365b2d9b. Iter 5: "dissolving flower petals" → 3/4 sharp (0.73 avg, blur 0.45-0.80). Iter 7: "hint of dissolving botanical silhouettes" → all 4 sharp (0.74 avg, blur 0.45-0.55). Iter 1: "motion-blurred flowers" → 0/4 blurred (0.64 avg). Cross-session confirmed in 6 sessions (365b2d9b, 39ce7668, 5ebbd0ef, 799400c2, 822750f3, bf0036e5).---#### Warm Fabric-Associated Color Names → Silk/Satin Rendering**Pattern ID:** `auto-warm-color-triggers-fabric`**Confidence:** Medium (50% success, 4 tests)**Symptom:** Warm color names like peach, coral, champagne, blush pink, dusty rose, soft gold, or warm ivory produce silk/satin fabric rendering instead of smooth abstract gradients.**Cause:** These colors are strongly associated with silk/satin textile training data in MJ V7. The model interprets them as material descriptors rather than pure color specifications.**Fix:**- Avoid warm fabric-associated color names in gradient prompts- Use cool or cool-dominant palettes for reliable gradients- If warm gradient needed, try less fabric-associated names like "warm amber" or "terracotta"- Add explicit anti-fabric keywords: `--no silk, satin, fabric, folds, textile`- A single warm color (e.g., dusty pink) can work when surrounded by dominant cool colors**Evidence:** Session d48e5b3c. Tested 4 times: 2 warm palettes both triggered fabric (iters 2, 4), 2 cool palettes succeeded (iters 1, 3), 1 mixed-cool palette with dusty pink succeeded (iter 5). The cool-to-warm ratio appears to matter.---#### Gallery Framing Trap (Art Movement + Gallery Context)**Pattern ID:** `pat-gallery-framing-trap`**Confidence:** Medium (0% success, 2 tests)**Symptom:** MJ renders a photo of artwork on a gallery wall rather than generating the artwork itself.**Cause:** Combining art movement names (color field painting) with gallery/exhibition context (gallery artwork) triggers meta-rendering.**Fix:**- Use artist inspiration (e.g., "James Turrell inspired") without gallery context words- OR use mood/quality words (contemplative, hypnotic luminosity) instead of art movement labels**Evidence:** Discovered in iteration 3B. "Gallery artwork" triggered meta-rendering of painting in exhibition space.---### LOW CONFIDENCE (Single/Limited Tests, Needs Validation)#### "Color Swatch" / "Wash" / "Matte Finish" → Painted Swatches**Pattern ID:** `auto-color-swatch-triggers-paint`**Confidence:** Low (0% success, 1 test)**Symptom:** Painted swatches on paper with visible brush edges instead of clean digital gradients.**Cause:** These keywords activate physical medium rendering in MJ V7.**Fix:**- Avoid paint-related vocabulary when seeking clean digital gradients or color fields- Use atmospheric/environmental language instead- Describe color transition directly without medium references**Evidence:** Session 7345a6e1 iter 5. Complete failure: all 4 images showed painted swatches. Scores dropped to avg ~0.31.---#### "Defocused" → Lens Glow/Aurora Effects**Pattern ID:** `auto-defocused-triggers-lens-effects`**Confidence:** Low (0% success, 1 test)**Symptom:** Lens glow, aurora, and light flare effects instead of smooth gradient softness.**Cause:** "Defocused" activates lens simulation rather than simple softness.**Fix:**- Use "smooth", "seamless", or "soft blend" instead of "defocused" for gradient smoothness- For actual intentional defocus/blur, use "out of focus photography" (tested separately as reliable)**Evidence:** Session 7345a6e1 iter 6. "Defocused" sabotaged 3/4 images with lens glow/aurora. Image 2 (without lens effects) was one of the best gradient results, proving the rest of the prompt worked.---#### Water/Fluid Keywords → 3D Fluid Simulation**Pattern ID:** `auto-water-vortex-triggers-3d-fluid`**Confidence:** Low (0% success, 1 test)**Symptom:** 3D fluid simulation rendering with thick ribbons instead of painted/illustrated linework.**Cause:** Water/fluid motion keywords (swirling water vortex, fluid patterns) activate MJ's 3D fluid training data, overriding painting/illustration style anchors like "oil painting".**Fix:**- Avoid water simulation keywords when seeking painted/illustrated linework- Describe visual mark-making instead: "fine flowing lines", "delicate strokes", "thin ink lines"---#### "Fine Delicate Brushstrokes" → Thick Painterly Marks**Pattern ID:** `auto-fine-brushstrokes-ignored`**Confidence:** Low (0% success, 1 test)**Symptom:** Thick painted marks instead of fine/thin lines.**Cause:** MJ interprets "brushstrokes" as thick painterly marks regardless of "fine" qualifier.**Fix:**- Describe line quality directly: "hair-thin strokes", "thin ink lines", "fine flowing lines"- Avoid "brushstrokes" keyword entirely- "Traditional brush drawing on black ground" was more effective than "fine delicate brushstrokes"---#### "Torn Paper Edge Texture" → Physical Paper Frame**Pattern ID:** `auto-torn-paper-creates-frame`**Confidence:** Low (0% success, 1 test)**Symptom:** Literal torn paper page sitting on background, creating unwanted frame/border effect.**Cause:** MJ interprets this as a semantic object (paper) rather than just a texture quality.**Fix:**- Avoid "torn paper edge texture" when element isolation is critical- Use "distressed edges" or "rough ink edges" instead- Adding "frame" to --no list does not fully prevent this**Evidence:** Session c2f5cce9. Appeared in 2/4 images in iteration 2 despite --no frame.---#### "Double Exposure" in Photorealistic Prompts**Pattern ID:** `auto-double-exposure-degrades-photo`**Confidence:** Low (0% success, 1 test)**Symptom:** Output shifts from photography to artistic/abstract treatment.**Cause:** "Double exposure" conflicts with --style raw which anchors photorealism.**Fix:**- Avoid "double exposure" when photorealistic base is desired- Use concrete impossible objects instead for surrealism within photorealism**Evidence:** Session 822750f3 iter 2. Double exposure degraded Fujifilm photorealistic quality. User rejected the artistic shift.---#### Pattern/Swirl Keywords → Continuous All-Over Patterns**Pattern ID:** `auto-continuous-vs-isolated-forms`**Confidence:** Low (0% success, 1 test)**Symptom:** Continuous pattern filling the frame instead of isolated floating forms with void between them.**Cause:** MJ V7 defaults to continuous all-over patterns when given pattern/swirl keywords.**Fix:**- Remove "all-over pattern"- Describe spatial structure explicitly: "scattered discrete shapes floating in dark void", "empty space between form clusters"- The pattern-filling instinct is strong and persists even with "negative space between strokes"---#### Generic 2D/3D Material Contrast Description Fails**Pattern ID:** `material-contrast-needs-physical-desc`**Confidence:** Low (0% success, 1 test)**Symptom:** Face and body rendered in similar material quality despite keywords about different materials.**Cause:** Describing 2D/3D material contrast as "outline edge" or "separation" is too abstract.**Fix:**- Describe material contrast physically using opposing material adjectives for each part- Example: "flat matte black" for void face vs "glossy luminous rainbow" for body- Name materials explicitly as different substances rather than relying on edge/outline separation---#### Border + Fine Art Print → Physical Print Rendering**Pattern ID:** `border-triggers-print-framing`**Confidence:** Low (0% success, 1 test)**Symptom:** MJ renders the image as a physical print on paper/surface instead of generating the artwork itself.**Cause:** Combining "painted white border" or "painted border" with "fine art print" creates physical object framing. Variant of Gallery Framing Trap.**Fix:**- Remove border and fine art print keywords when they co-occur- For edge separation, describe it as illustration content: "white edge glow", "glowing edge where dark meets light"- For style anchoring without photorealism, "spray paint airbrush illustration" alone is sufficient---#### MJ UI Style Reference Panel Auto-Selection**Pattern ID:** `auto-icm-sref-ui-unreliable`**Confidence:** Low (0% success, 1 test)**Symptom:** Unintended visual qualities contaminate style transfer; unwanted textures appear.**Cause:** MJ UI style reference panel auto-selects multiple recent images from your library instead of single selected image.**Fix:**- Avoid MJ UI style reference panel for precise style transfer- Use direct `--sref URL` in prompt text for single-image control- Or rely on prompt-only approach**Evidence:** ICM session 365b2d9b iter 8. UI auto-selected 2 refs, introduced unwanted organic fibers. Batch avg dropped from 0.89 (prompt-only iter 6) to 0.78. Needs more testing.---#### Editor Inpainting Introduces Unwanted Environmental Elements**Pattern ID:** `ed01`**Confidence:** Low (0% success, 1 test)**Symptom:** MJ editor inpainting regenerates backgrounds with walls, corners, surfaces even when prompt specifies flat/seamless backdrop.**Cause:** The preserved figure implies a physical space that MJ fills with environmental context.**Fix:**- Editor inpainting is not suitable for creating flat seamless backgrounds- Use prompt-only approaches for flat backdrops instead**Evidence:** Session 17bbeab3 iter 15. Smart Select segmentation worked but regenerated area had walls and tonal gradients instead of flat gray.---#### Editor Edit Creates Grain/Texture Mismatch on sref-Assisted Images**Pattern ID:** `ed02`**Confidence:** Low (0% success, 1 test)**Symptom:** Preserved area has coarse sref-transferred grain; regenerated area has different/smoother grain.**Cause:** Editor edit regenerates only the selected area, which loses the sref grain transfer applied to the full image.**Fix:**- When using editor edit on sref-assisted images, the regenerated area will have different grain character than the preserved area- This is especially visible with heavy film grain- Consider whether the texture discontinuity is acceptable before using editor edit**Evidence:** Session 17bbeab3 iter 15. Preserved figure had coarse sref-transferred grain; regenerated background had smoother different grain.---## Patterns Requiring Further ValidationThe following patterns have been observed once and need cross-session validation before increasing confidence level:- **auto-color-swatch-triggers-paint** — Needs at least 2 more tests confirming painted swatches in clean gradient attempts- **auto-defocused-triggers-lens-effects** — Needs validation that "smooth" consistently outperforms "defocused" across multiple gradient sessions- **auto-water-vortex-triggers-3d-fluid** — Needs tests comparing water keywords vs alternative linework descriptions- **auto-fine-brushstrokes-ignored** — Needs comparison tests of "brushstrokes" vs "hair-thin strokes"- **auto-torn-paper-creates-frame** — Needs tests with different edge descriptors- **auto-double-exposure-degrades-photo** — Single photo session; needs validation in other photorealistic contexts- **auto-continuous-vs-isolated-forms** — Pattern/void tradeoff needs testing across 2+ sessions- **material-contrast-needs-physical-desc** — Needs tests comparing abstract vs physical material descriptions- **border-triggers-print-framing** — Needs tests of border-only, print-only, and combined variations---**Total Patterns:** 18 failure-mode patterns (3 Medium, 15 Low)**Last Generated:** 2026-02-09