Similar Image Search
Tool name: find_similar_by_image
Upload an image, and Nephele will find the most visually similar references in your Eagle library. It's perfect for the "I have a sketch / screenshot / inspiration image and want to find pieces in a similar style within my own collection" scenario.
When to use it
- You drew a sketch and want to dig up finished-work references in a similar style from your library
- You saw an inspiration image and want to check whether your Eagle library holds anything similar
- You want to use one image in Eagle as a baseline and find more pieces in the same vein
How to use it
Through Agent chat
- "This is my sketch — help me find references in a similar style"
- "Find images in Eagle with a vibe similar to this screenshot"
- "Use this image as a reference and find more pieces in the same style"
The Agent automatically recognizes whether you mean a local file or an item in Eagle, then calls the tool. Two query methods are supported:
| Method | Description |
|---|---|
| Local file path | Pass the full path to a local image |
| Eagle item ID | Pass the ID of an item in Eagle to use that image directly as the query |
技巧
Similar image search is currently only available in Agent chat — there is no standalone entry in the toolbox.
How it works
Similarity is computed as a weighted blend of two signals:
| Signal | Weight | What it measures |
|---|---|---|
| EVA02 visual embedding | 85% | Overall composition, color tone, mood, and art style |
| Tag overlap | 15% | The makeup of the scene's elements (characters, settings, objects) |
The visual embedding is the primary basis: every image generates a 1024-dimensional EVA02 vector as a byproduct of the tagging process. The smaller the angle between two images' vectors, the more visually similar they are. This set of vectors is persisted alongside the library index, so no extra computation is needed. Tag overlap serves as a supplement, guarding against the "similar mood but completely unrelated content" case.
By default it returns the Top 30 results, and results with too low a similarity are filtered out. If the library hasn't yet generated vectors for the relevant images, the tool falls back to pure tag-matching mode.
Reading the results
Returned results are sorted from highest to lowest similarity. Each result image shows:
- Similarity score — the higher, the more alike; results that are too low aren't shown
- Source — which item it comes from in Eagle
- Tags and star rating — to help you judge whether this is the result you wanted
提示
Result quality depends directly on how complete your library index is. The more thorough the index coverage, the more complete the EVA02 visual vectors and tags, and the more accurate the recommendations. If you haven't indexed yet, results will be heavily limited.
Limitations and notes
- It only searches the Eagle library; it won't go online to find external images
- When unindexed, it falls back to pure tag matching, and result relevance drops noticeably
- Results with tag overlap below 12% are treated as noise and excluded from ranking