README.md

---
library_name: onnx
tags:
  - computer-vision
  - image-enhancement
  - photo-retouching
  - computational-photography
  - color-grading
  - onnx
  - mobile
  - lightweight
  - multi-task-learning
datasets:
  - Phitran21/adaptive-photo-retouching-6style
---

# AdaptivePhotoNet

**AdaptivePhotoNet** is a lightweight, scene-aware neural network for
automatic photo retouching.

Instead of generating a new image pixel-by-pixel, the model analyzes a
low-resolution preview of the photograph and predicts a compact
**21-dimensional retouching recipe** that can be applied to the original
full-resolution image by a deterministic image-processing pipeline.

The model also predicts **10 scene attributes** to provide auxiliary
scene understanding.

AdaptivePhotoNet contains **5,876,943 parameters** and supports six
retouching styles:

- Natural
- Vivid
- Cinema
- Portrait
- Film
- Moody

The model is exported to **ONNX** and designed with lightweight desktop,
mobile, and edge inference in mind.

---

## Model Concept

AdaptivePhotoNet separates **visual understanding** from
**full-resolution image processing**.

```text
                     ┌─────────────────┐
                     │  Original Image │
                     └────────┬────────┘
                              │
                       resize / preview
                              │
                              ▼
                    ┌──────────────────┐
                    │ RGB 224 × 224    │
                    └────────┬─────────┘
                             │
             ┌───────────────┴───────────────┐
             │                               │
        Image Features                  Style ID
             │                         0 ... 5
             └───────────────┬───────────────┘
                             ▼
                    ┌──────────────────┐
                    │ AdaptivePhotoNet │
                    │  5.88M params    │
                    └────────┬─────────┘
                             │
                  ┌──────────┴──────────┐
                  ▼                     ▼
          21D Retouch Recipe      10 Scene Scores
                  │
                  ▼
        Deterministic Retouching
                Pipeline
                  │
                  ▼
       Full-Resolution Output Image

The neural network therefore does not need to reconstruct the full-resolution photograph.

It predicts how the photograph should be adjusted, while the final rendering is performed by conventional image-processing operations.

This design has several practical advantages:

low neural-network inference cost;

processing is independent of the original image resolution at the model stage;

deterministic full-resolution rendering;

compact ONNX deployment;

interpretable adjustment parameters;

selectable photographic styles;

suitable for mobile and edge applications.



---

Model Specifications

Property	Value

Model	AdaptivePhotoNet
Parameters	5,876,943
Model input resolution	224 × 224
Image format	RGB
Tensor layout	NCHW
Image dtype	float32
Image range	[0.0, 1.0]
Style input	int64
Number of styles	6
Recipe output	21 dimensions
Scene output	10 dimensions
Runtime format	ONNX



---

Inputs

Image

name: image
shape: [1, 3, 224, 224]
dtype: float32
layout: NCHW
color: RGB
range: 0.0 - 1.0

The original photograph should be converted to RGB, resized to 224 × 224, converted to float32, normalized to [0, 1], and arranged in NCHW format.

The 224 × 224 image is used for analysis only.

The final retouching operations can be applied separately to the original full-resolution photograph.

Style

name: style_id
shape: [1]
dtype: int64

ID	Style

0	Natural
1	Vivid
2	Cinema
3	Portrait
4	Film
5	Moody


Changing style_id instructs the same model to predict a different retouching direction for the input photograph.


---

Outputs

AdaptivePhotoNet produces two outputs.

1. Retouching Recipe

shape: [1, 21]

The 21-dimensional vector describes the photographic adjustments that should be applied by the retouching engine.

Linear Parameters

Dimensions 0–16 represent:

Dim	Parameter	Range

0	Exposure EV	-2.0 → 2.0
1	Temperature	-1.0 → 1.0
2	Tint	-1.0 → 1.0
3	Shadows	-1.0 → 1.0
4	Highlights	-1.0 → 1.0
5	Contrast	-1.0 → 1.0
6	Tone Curve 0	0.0 → 1.0
7	Tone Curve 1	0.0 → 1.0
8	Tone Curve 2	0.0 → 1.0
9	Tone Curve 3	0.0 → 1.0
10	Tone Curve 4	0.0 → 1.0
11	Shadow Tone Strength	0.0 → 0.3
12	Highlight Tone Strength	0.0 → 0.3
13	Saturation	-1.0 → 1.0
14	Vibrance	-1.0 → 1.0
15	Fade	0.0 → 1.0
16	Vignette	0.0 → 1.0


Hue Parameters

Hue is represented circularly using sine/cosine pairs rather than a single scalar value.

17, 18 → shadow_tone_hue    [sin, cos]
19, 20 → highlight_tone_hue [sin, cos]

This avoids the discontinuity that occurs when representing circular hue values directly near the angle boundary.

The complete interpretation is defined in:

recipe_schema.json


---

2. Scene Probabilities

shape: [1, 10]

The auxiliary scene head predicts ten visual attributes:

Index	Scene Attribute

0	Human
1	Face Visible
2	Skin Visible
3	Portrait
4	Indoor
5	Outdoor
6	Night
7	Low Light
8	Backlit
9	High Dynamic Range


These attributes provide additional scene understanding alongside the retouching prediction.

They can also be useful for debugging, analysis, UI features, or future scene-aware processing logic.


---

Why Predict a Recipe Instead of Pixels?

Many neural photo-enhancement systems directly generate a complete output image.

AdaptivePhotoNet takes a different approach.

Pixel-to-pixel model:

Full Image → Neural Network → Full Image


AdaptivePhotoNet:

Small Preview → Neural Network → 21 Parameters
                                    ↓
Original Full-Resolution Image → Retouching Engine → Output

For photographic retouching, much of the desired transformation can be expressed through global or structured photographic controls.

Predicting these controls instead of millions of output pixels allows the neural network to remain relatively small.

It also keeps the transformation interpretable.

For example, an application can inspect whether the network requested:

Exposure        +0.32 EV
Temperature     -0.08
Highlights      -0.21
Contrast        +0.14
Saturation      +0.07
Vignette         0.11
...

rather than receiving only an opaque generated image.


---

Multi-Style Retouching

AdaptivePhotoNet uses a separate style_id input rather than requiring six independent models.

The same photograph can therefore be analyzed under different retouching directions:

┌─ Natural
                    ├─ Vivid
Input Photograph ───├─ Cinema
                    ├─ Portrait
                    ├─ Film
                    └─ Moody

The style determines the intended aesthetic direction while the image content determines the actual adjustment recipe.

This means that Film, for example, is not intended to represent one fixed preset applied identically to every photograph.

Two photographs using the same style may receive different exposure, tone, color, curve, and other adjustments according to their visual characteristics.


---

Training Dataset

AdaptivePhotoNet was developed together with:

Adaptive Photo Retouching 6-Style Dataset

https://huggingface.co/datasets/Phitran21/adaptive-photo-retouching-6style

The dataset contains original photographs paired with six adaptively retouched variants:

Original
├── Natural
├── Vivid
├── Cinema
├── Portrait
├── Film
└── Moody

The target transformations were generated adaptively for individual images rather than by applying six globally fixed presets.

See the dataset card for details about dataset generation, source data, licensing, and limitations.


---

Inference Pipeline

A typical application pipeline is:

1. Load the original image
             ↓
2. Create 224 × 224 RGB preview
             ↓
3. Normalize to float32 [0, 1]
             ↓
4. Convert HWC → NCHW
             ↓
5. Select style_id
             ↓
6. Run AdaptivePhotoNet
             ↓
7. Decode the 21D recipe
             ↓
8. Apply recipe to original-resolution image
             ↓
9. Produce final retouched photograph

The original full-resolution image does not need to pass through the neural network.


---

Minimal ONNX Runtime Example

import numpy as np
import onnxruntime as ort
from PIL import Image

STYLE = {
    "natural": 0,
    "vivid": 1,
    "cinema": 2,
    "portrait": 3,
    "film": 4,
    "moody": 5,
}

image = Image.open("photo.jpg").convert("RGB")
preview = image.resize((224, 224))

x = np.asarray(preview, dtype=np.float32) / 255.0
x = np.transpose(x, (2, 0, 1))
x = np.expand_dims(x, axis=0)

style_id = np.asarray([STYLE["film"]], dtype=np.int64)

session = ort.InferenceSession("AdaptivePhotoNet.onnx")

recipe_vector, scene_probs = session.run(
    None,
    {
        "image": x,
        "style_id": style_id,
    },
)

print("Recipe:", recipe_vector)
print("Scene probabilities:", scene_probs)

The resulting recipe_vector must then be interpreted according to recipe_schema.json and applied by the corresponding image-retouching pipeline.


---

Android / ONNX Runtime

Recommended execution-provider configuration:

FP32

XNNPACKExecutionProvider
        ↓ fallback
CPUExecutionProvider

INT8

CPUExecutionProvider

Actual performance depends on device hardware, ONNX Runtime version, thread configuration, quantization method, and preprocessing pipeline.


---

Intended Use

AdaptivePhotoNet is intended for experimentation and development in:

automatic photo retouching;

computational photography;

adaptive color grading;

scene-aware image enhancement;

mobile photo editing;

lightweight computer vision;

ONNX Runtime applications;

edge inference;

non-destructive image adjustment prediction.



---

Limitations

AdaptivePhotoNet predicts photographic adjustments from a 224 × 224 representation of the image.

Fine details that disappear during resizing may therefore not influence the predicted recipe.

The model may also perform less reliably on images significantly outside its training distribution, including unusual lighting, extreme exposure, uncommon photographic styles, or heavily degraded images.

Retouching quality is inherently subjective. Different users may prefer different photographic interpretations of the same image.

The six supported styles represent only six predefined aesthetic directions and should not be interpreted as exhaustive photographic styles.

The model predicts retouching parameters rather than reconstructing or generating image content. It therefore cannot perform tasks such as object removal, image inpainting, semantic image editing, or generative relighting.


---

Related Resources

Training Dataset

Adaptive Photo Retouching 6-Style Dataset

https://huggingface.co/datasets/Phitran21/adaptive-photo-retouching-6style

Source Code and Demo

https://github.com/phiiggfdg/adaptive-retouch-6m-onnx


---

Author

Trần Phi

Hugging Face:
https://huggingface.co/Phitran21

GitHub:
https://github.com/phiiggfdg

Website:
https://toren.io.vn


---