Update README.md

4d6332d verified 2 months ago

11.6 kB

---
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
---

	---
	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


	---