피파 한 줄 정리: Image-to-image의 핵심 dial은 *denoising strength*. 0.2 = 미세 조정, 0.5 = style 변환, 0.9 = 거의 새로 생성. 이 슬라이더가 reference 영향력의 핵심.
Mental model: Imagine putting tracing paper over a photograph and drawing on top of it. The photograph provides the structure — where things are, their proportions, the composition. Your pen provides the new style — the artistic interpretation, the color choices, the texture. Image-to-image generation works the same way: you give the model an existing image as a structural foundation, and it reinterprets that structure through the lens of your text prompt.
How Image-to-Image Works
Instead of starting from pure random noise (as in text-to-image), image-to-image starts by adding a controlled amount of noise to your input image, then denoises it according to your prompt. The less noise you add, the more the output resembles the input. The more noise, the more creative freedom the model takes.
Input Image Add Noise (controlled) Denoise with Prompt
┌──────────┐ ┌──────────────────────┐ ┌──────────────────┐
│ Original │ ──→ │ Partially noised │ ──→ │ New image with │
│ photo │ │ version (structure │ │ original layout │
│ │ │ partially preserved) │ │ but new style │
└──────────┘ └──────────────────────┘ └──────────────────┘
▲
"Denoising Strength"
Low (0.2) = subtle changes
High (0.8) = dramatic changes
The Denoising Strength Slider
The most important parameter in image-to-image is denoising strength (sometimes called "creativity" or "transformation" depending on the platform):
- 0.1–0.3 (low): Subtle adjustments. Color shifts, slight texture changes, minor refinements. The output is clearly recognizable as the input.
- 0.4–0.6 (medium): Meaningful transformation. Style transfer, season changes, time-of-day shifts. Structure is preserved but details change significantly.
- 0.7–0.9 (high): Major reimagining. The input is a loose inspiration. Composition might be preserved, but content can change dramatically.
- 1.0: Equivalent to text-to-image — the input is completely noised and the model generates freely.
Common Use Cases
- Style transfer: Photo → oil painting, sketch → rendered illustration, realistic → anime
- Concept variation: "What would this room look like in a different style?"
- Rough-to-refined: Start with a rough sketch or block-out, use image-to-image to add detail and polish
- Season/time changes: Summer scene → winter scene, day → night
- Color palette changes: Warm tones → cool tones while keeping composition
"A cozy cabin in winter, watercolor style" → Generic result, not YOUR cabin
Feed a photo of your actual cabin + prompt "watercolor painting, snowy winter scene, warm window glow" + denoising strength 0.5 → Your cabin, in watercolor, in winter
- Image-to-image preserves structure from an input image while applying new style and details.
- Denoising strength controls the balance: low = subtle change, high = dramatic reimagining.
- Common uses: style transfer, concept variation, rough-to-refined, environmental changes.
- Iterative image-to-image loops are a core professional workflow for progressive refinement.