init: project structure with LAION scorer, conda setup, and research docs
This commit is contained in:
@@ -0,0 +1,65 @@
|
||||
# Output Format — JSON
|
||||
|
||||
## Structure
|
||||
|
||||
The scorer outputs a JSON file with an array of results. Each entry represents one photo.
|
||||
|
||||
## Schema
|
||||
|
||||
```json
|
||||
[
|
||||
{
|
||||
"filename": "IMG_0001.jpg",
|
||||
"filepath": "/path/to/wedding-photos/IMG_0001.jpg",
|
||||
"subtype": "aesthetic",
|
||||
"laion_score": 7.2,
|
||||
"laion_version": "v2",
|
||||
"scoring_date": "2026-01-15T10:30:00Z"
|
||||
}
|
||||
]
|
||||
```
|
||||
|
||||
## Fields
|
||||
|
||||
| Field | Type | Description |
|
||||
|-------|------|-------------|
|
||||
| `filename` | string | Base filename (e.g., `IMG_0001.jpg`) |
|
||||
| `filepath` | string | Full absolute path to the image |
|
||||
| `subtype` | string | Classifier subtype — `aesthetic` for LAION (future: `technical`, `blur`, etc.) |
|
||||
| `laion_score` | float | LAION aesthetic score (0-10 scale) |
|
||||
| `laion_version` | string | Version of LAION model used (e.g., `v2`) |
|
||||
| `scoring_date` | string | ISO 8601 timestamp of when scoring was performed |
|
||||
|
||||
## Example
|
||||
|
||||
```json
|
||||
[
|
||||
{
|
||||
"filename": "DSC_0421.jpg",
|
||||
"filepath": "C:/wedding-photos/day1/DSC_0421.jpg",
|
||||
"subtype": "aesthetic",
|
||||
"laion_score": 8.1,
|
||||
"laion_version": "v2",
|
||||
"scoring_date": "2026-01-15T10:30:00Z"
|
||||
},
|
||||
{
|
||||
"filename": "DSC_0422.jpg",
|
||||
"filepath": "C:/wedding-photos/day1/DSC_0422.jpg",
|
||||
"subtype": "aesthetic",
|
||||
"laion_score": 3.4,
|
||||
"laion_version": "v2",
|
||||
"scoring_date": "2026-01-15T10:30:01Z"
|
||||
}
|
||||
]
|
||||
```
|
||||
|
||||
## Thresholds (TBD)
|
||||
|
||||
Thresholds for keep/reject/review are not yet defined. Will be determined after running initial batches and reviewing score distributions.
|
||||
|
||||
Suggested approach:
|
||||
|
||||
1. Score a sample of 100-200 photos
|
||||
2. Analyze the score distribution
|
||||
3. Set thresholds based on the distribution (e.g., top 25% = keep, bottom 25% = reject, middle = review)
|
||||
4. Adjust based on photographer feedback
|
||||
Reference in New Issue
Block a user