AI Visual Creation

Decision frameworks for AI image and video generation. Not tutorials — corrections, gotchas, and "which tool for which job."

Midjourney: Version Gotchas

V7 Breaking Changes (Critical)

Stylize Scale Migration: V6 --s 100 ≈ V7 --s 300-400 | V6 --s 250 ≈ V7 --s 600-700

V7 workarounds for changed weighting:

• Word order matters (early = more weight)
• Use natural language emphasis
• --no for exclusion
• Repetition for emphasis

V6 prompt: cyberpunk::2 nature::1 dystopian::-0.5 V7 equivalent: cyberpunk city with nature elements, NOT dystopian --no dystopian, grim, dark

Midjourney: Reference Type Decision

Quick Selector

⚠️ V7 Migration: --cref deprecated in V7. Use --oref instead (works for characters AND objects).

Reference Type Deep Dive

Image Prompt (--iw)

• Mental model: Addition (image + text = result)
• Preserves: Composition, layout
• Changes: Details, style via text
• Range: 0-3 (V7), 0-2 (V6)

Style Reference (--sref)

• Mental model: Multiplication (style × subject = result)
• Preserves: Color palette, mood, rendering
• Changes: Subject, composition entirely
• Range: --sw 0-1000

Character Reference (--cref) — V6 ONLY

• ⚠️ Deprecated in V7 — use --oref instead
• CRITICAL: Works best with Midjourney-generated images, NOT real photos
• --cw 0 = face only (max outfit flexibility)
• --cw 100 = everything (face, hair, clothing)
• Cannot preserve: fine freckles, small logos, detailed tattoos

Omni Reference (--oref) — V7 ONLY

• 2x GPU cost
• Only ONE reference allowed
• NOT compatible with inpainting/outpainting or Draft mode
• Competing params: high --stylize needs higher --ow to balance

Common Failures

Model Selection: Images

Decision Matrix

Text Rendering Hierarchy

Best → Worst: Ideogram 3.0 (85-90%) >> GPT Image 1.5 >> Recraft V3 >> Flux 2 (~60%) >> Imagen 4 >> DALL-E 3 >> Midjourney V7 (~15% better than V6, still poor)

Rule: If you need readable text, don't use Midjourney. Use Ideogram, GPT Image, or Flux 2.

Model Selection: Video

Decision Matrix

Key Insight

Audio-visual sync is now a competitive differentiator. Only Kling 2.6 generates video + voiceover + sound effects + ambient audio in a single pass.

Troubleshooting Patterns

"It won't preserve the shape"

1. Use Image Prompt with high --iw (2.0+)
2. Match aspect ratio (input 1:1 → output --ar 1:1)
3. Add --style raw for tighter adherence
4. Lower --stylize (30-50) for more literal interpretation
5. If still failing: Try Imagen 4 or Flux 2 — they preserve shapes more literally than Midjourney

"It keeps making it symmetric"

Midjourney defaults to symmetry. Fight it:

1. Add "asymmetrical" keyword explicitly
2. Use --no symmetric, mandala, radial, mirrored, balanced, centered
3. Add --chaos 6-10
4. Use directional language ("positioned to the left", "stepping diagonally")
5. Material words help ("weathered metal", "carved stone" resist perfect symmetry)

"Style overwhelms subject"

Balance the competing forces:

• Lower --sw (style weight)
• Increase --iw (image weight) if using reference
• Use --style raw
• Simplify text prompt

"Character keeps changing"

V7 (recommended):

1. Use --oref with Midjourney-generated source (2x GPU cost)
2. Start at --ow 100, increase to 200-400 for facial accuracy
3. For many images: Leonardo.ai with custom LoRA

V6 (legacy):

1. Use --cref with Midjourney-generated source (not real photos)
2. --cw 0 for face only, --cw 100 for everything

References

Sources: All claims cite official documentation (docs.midjourney.com, vendor APIs) and benchmarks (Artificial Analysis, LM Arena). Full URLs in reference files.