What does Image Reverse Analysis do?
Positioning
Reference image → structure, not reference image → nouns.
v2.0.0 merges the previous three duplicated reverse-engineering workflows (this Skill, skills/nai5-prompt-engineering/references/reverse-image-analysis.md, and anima-prompt-compiler § 5). Adapters no longer hold their own reverse-engineering logic.
Two Tasks, Distinguish First
| What the user wants | Output destination |
|---|---|
| Faithfully reproduce this image ("Reverse engineer the prompt for this image") | This Skill → Model adapter. Keep specific objects, composition, color scheme. |
| Create original work referencing this image's feel ("Reference this, but make it original") | This Skill → aesthetic-director-core → character-design-engine. Keep only structure; replace all concrete objects. |
If unclear, ask. Default is the latter (the vast majority of this user's needs).
Analysis Sequence
First look at this specific image the user provided—do not substitute it with a similar image found via web search.
1. Silhouette What large shapes make up the outer contour? What is recognizable in a thumbnail?
2. Mass Where is decoration/density concentrated—head, chest/waist, back, or limbs? Where is it empty?
3. Motif grammar What recurring morphological language appears (snake/bone/flower/line/fabric fold/straight/curve)? In how many places? How does it transform?
4. Garment What are the four layers: base / structural / signature extension / accessory? Where is the cut?
5. Material Which two materials create contrast? On which structure does that contrast land?
6. Palette What are the three levels of color—base / structural / punctum—and where do they fall? How many puncta are there (ideally one)?
7. Pose logic What is she doing? Is there physical contact between the pose and props/ancillary structures? What is the cause-and-effect chain?
8. Light Main light direction, falloff, shadow settling area. Are there lights without a source?
9. Negative space Which areas are deliberately left empty? Why?
10. Presentation Clean plate / decorated key visual / editorial plate / environmental vignette?
11. Mood mechanism How is the mood achieved? (Not "mysterious"—but "the contrast between the creature's eyes and her relaxed demeanor")
Output Contract
## Structural Breakdown
1–11 each in one to two lines, pointing to specific locations in the image.
## What Actually Works (3–5 items)
Principles that remain valid after removing all concrete objects. Example: "A huge circular silhouette makes the outline on white immediately readable" rather than "there is a snake."
## Transferable / Non-Transferable
Transferable: Structural principles.
Non-Transferable: Specific character identity, trademark, exclusive composition, copyrighted elements.
## Next Step
→ Faithful reproduction: Hand off to [adapter] with a locked fact list.
→ Original creation: Hand off to aesthetic-director-core with the "What Actually Works" list as constraints.
Prohibitions
- Outputting words like
white background, anime, detailed, intricate, elegantthat carry no structural information - Using "noun swapping" as originality (snake → dragon)
- Replacing the user's reference image with a "similar image" found via search
- Copying specific character / copyrighted settings unless the user explicitly asks for faithful reproduction
Example
See aesthetic-director-core/references/taste-calibration-pairs.md Pair 5: Snake-coiling reference image → five structural principles → silk maiden original.
References
references/design-language-extraction.md— Key points and reconstruction principles for silhouette / color / fashion structure / emotion extraction.