Style Reference UX Patterns¶

Comparative analysis of style reference workflows across major AI image generation products. Reference for product design decisions around style input, strength control, browsing, and persistence.

Style Input Methods¶

Ranked by user effort required:

Midjourney --sref Deep Dive¶

Five creation methods: 1. Image URL: --sref <url> in Discord, paperclip icon in web app 2. Numeric code: --sref 2213253170 (billions of internal styles) 3. --sref random: discover style, reveals code post-generation 4. Style Creator: grid selection → generates unique code 5. Moodboard: --profile <id> named image collections

Style strength: --sw 0-1000, default 100. Community-found optimal: 65-175.

Multi-style blending: --sref code1::2 code2::1 - proportional weighting (only ratios matter).

Benchmark grids: 16 creative domains shown with same sref code. Reveals style's range across character, fashion, portraiture, technical drawing, etc.

Community ecosystem spawned: Midlibrary.io, SrefHunt.com (50K+ codes), sref-midjourney.com, PromptsRef.com. All follow: grid cards + one-click copy + category filters + personal bookmarks.

Strength/Influence Control Patterns¶

Style vs Structure Separation¶

Adobe Firefly: two distinct sliders: - Style Strength: how closely output follows the style reference - Visual Intensity: detail/drama of base image before style application

Structure Reference (Firefly) is separate from Style Reference: controls outline and depth matching. Can use both independently.

Leonardo.ai: same image can be Content Reference OR Style Reference. Clear mental model: - Content Reference = WHAT is in the image (subjects, composition) - Style Reference = HOW it looks (colors, artistic style, texture)

Persistence and Sharing¶

Key UX Insights¶

1. Numeric codes are the most viral mechanism. Midjourney's single-number encoding spawned an entire third-party ecosystem. A portable, copy-pasteable aesthetic identifier creates sharing loops that no gallery can match.

2. Two-tier commitment works. Ideogram's "Quick Reference" (temporary) vs "My Styles" (permanent named) pattern serves exploration and production simultaneously. Key: frictionless experimentation, clear path to commitment.

3. --sref random as discovery mechanic. Generate → discover → save code → build collection. Turns randomness into curation. Creates addictive collecting behavior without explicit gamification.

4. Named qualitative modes beat numeric sliders for non-power-users. Freepik's "Faithful/Bold/Dreamy" flavors and Leonardo's "Low/Mid/High/Ultra/Max" are more actionable than 0-1000 scales for users who think in qualities.

5. Benchmark grids show style versatility. 16 variations of one style across different subjects is more informative than 4 variations of the same subject. Reveals style's range and limitations.

6. LoRA training exposed to end users is a UX barrier AND moat. Freepik's 15-30 min wait creates investment/commitment. Users who trained a custom style don't switch platforms.

7. Real-time feedback (Krea, <50ms) changes the paradigm. From "configure then generate" to "continuously sculpt." Fundamentally different interaction model.

8. Style + Character as independent axes. Freepik's separate trainable entities that combine freely is powerful for brand content (same character, multiple styles or vice versa).

Ideogram Pattern: Immutable Custom Styles¶

Reference images in Ideogram's custom styles cannot be changed after creation. Forces intentional curation - you commit to a style before training it. UX trade-off: more friction upfront, cleaner library management.

Quick Reference tab: temporary style for experimenting (no naming, no saving required). Permanent "My Styles" for production use.

Runway Aleph: Post-Generation Style¶

Post-generation style transfer via text prompts without regenerating from scratch. Apply style AFTER generation → non-destructive workflow. Currently only for video.

Implication for product design: style can be a post-processing step, not just a generation parameter.

Implementation Notes for Custom LoRA Systems¶

For a LoRA-based style system (like Freepik's):

Recommended: rank32, 1000 steps, cosine scheduler
Dataset: 5-50 images, trigger_word + content captions (no style descriptors)
Caption dropout: 0.1 (forces style into LoRA weights)
Inference: Gemini-based prompt reformulation to weave style into scene description

Dataset size → quality: | Refs | Quality | Steps | |------|---------|-------| | 1-3 | Poor | N/A | | 5-7 | Marginal | 1500 | | 10-15 | Good | 1000 | | 30+ | Excellent | 500 |

Gemini two-level approach for style injection: - L1 (Style Extraction): analyze reference images → style profile with MUST/ENRICH elements - L2 (Prompt Reformulation): weave style profile + user prompt naturally, resolve conflicts (style wins on visuals, user wins on subject/action)

Gotchas¶

• Building style browsing into the product is critical: the explosion of third-party sref sites shows unmet demand. Users want to discover styles they haven't seen, not just apply ones they already know.
• LoRA at rank128 on small datasets (<10 images) overfit after ~500 steps: rank32 is more robust for small datasets. Use rank128 only with 30+ reference images.
• Cosine scheduler causes "jumps" at midpoint in LoRA training: if LR is still high when scheduler reaches midpoint, model can overshoot and produce inconsistent checkpoints. Use linear warm-down or check-at-500-steps pattern.
• Content leakage with small/homogeneous datasets: a 2-image style dataset that both contain the same person → LoRA learns the person, not the style. Need content diversity within style consistency.

See Also¶

• Diffusion LoRA Training - training style LoRAs
• flux klein style lora system - full style system architecture
• FLUX Klein 9B Inference - LoRA stacking at inference