SenseCraft: Unified Perceptual Feature Loss Framework

A PyTorch framework providing various perceptual loss functions and evaluation metrics for image processing tasks including super-resolution, image restoration, style transfer, and more.

Features

• Compound Loss System: SenseCraftLoss for easy multi-loss configuration with automatic value range handling
• Evaluation Metrics: PSNR, SSIM, MS-SSIM, LPIPS with functional API and dB scale support
• Multiple Perceptual Loss Types: ConvNext, DINOv3 (ConvNext & ViT), LPIPS
• Frequency Domain Losses: FFT and Patch-FFT losses with configurable normalization
• Edge & Structure Losses: Sobel, Laplacian, Gradient, Structure Tensor losses
• Video/3D Losses: Temporal SSIM, 3D SSIM, Frame Difference losses
• General Losses: Charbonnier, SSIM, MS-SSIM, Gaussian noise-aware losses
• Self-Supervised Features: DINOv3 models provide better generalization than supervised features
• Flexible Configuration: Layer selection, normalization options, Gram matrix support
• Gradient Flow: Proper gradient handling for training neural networks

Installation

# Basic installation
pip install sensecraft

# With DINOv3 support (requires transformers >= 4.56.0)
pip install sensecraft[dinov3]

# Full installation with all optional dependencies
pip install sensecraft[full]

For development:

git clone https://github.com/KohakuBlueleaf/SenseCraft.git
cd SenseCraft
pip install -e ".[full]"

Quick Start

Using SenseCraftLoss (Recommended)

The easiest way to use multiple losses is through SenseCraftLoss:

import torch
from sensecraft.loss import SenseCraftLoss

# Simple configuration with {name: weight} format
loss_fn = SenseCraftLoss(
    loss_config=[
        {"charbonnier": 1.0},    # Main reconstruction loss
        {"sobel": 0.1},          # Edge preservation
        {"ssim": 0.05},          # Structural similarity
        {"lpips": 0.1},          # Perceptual quality
    ],
    input_range=(-1, 1),  # Your data's value range
    mode="2d",            # "2d" for images, "3d" for video
)

# Create sample images
predicted = torch.randn(1, 3, 256, 256)
target = torch.randn(1, 3, 256, 256)

# Compute all losses at once
losses = loss_fn(predicted, target)
print(losses["loss"])        # Total weighted loss (for backprop)
print(losses["charbonnier"]) # Individual loss values
print(losses["sobel"])

Typed Configs for Complex Losses

For losses with many parameters, use typed config classes:

from sensecraft.loss import (
    SenseCraftLoss,
    DinoV3LossConfig,
    LPIPSConfig,
    PatchFFTConfig,
)

loss_fn = SenseCraftLoss(
    loss_config=[
        {"charbonnier": 1.0},
        DinoV3LossConfig(
            weight=0.1,
            model_type="small_plus",
            loss_layer=-4,
            use_gram=False,
            use_norm=True,
        ),
        LPIPSConfig(weight=0.05, net="alex"),
        PatchFFTConfig(weight=0.05, patch_size=16),
    ],
    input_range=(-1, 1),
)

Monitoring Losses (weight=0)

Losses with weight=0 are computed under torch.no_grad() for efficiency but still returned:

loss_fn = SenseCraftLoss(
    loss_config=[
        {"charbonnier": 1.0},
        {"ssim": 0.0},      # Computed for logging, not in gradient
        {"ms_ssim": 0.0},   # Same here
    ],
    input_range=(0, 1),
)

losses = loss_fn(pred, target)
# losses["loss"] only includes charbonnier
# losses["ssim"] and losses["ms_ssim"] available for logging

3D/Video Mode

For video data (B, T, C, H, W), use mode="3d". 2D-only losses are applied frame-by-frame:

loss_fn = SenseCraftLoss(
    loss_config=[
        {"charbonnier": 1.0},
        {"sobel": 0.1},           # Applied per-frame
        {"temporal_gradient": 0.1},  # 3D-specific loss
        {"stssim": 0.05},         # Spatio-temporal SSIM
    ],
    input_range=(-1, 1),
    mode="3d",
)

video_pred = torch.randn(2, 8, 3, 64, 64)  # (B, T, C, H, W)
video_target = torch.randn(2, 8, 3, 64, 64)
losses = loss_fn(video_pred, video_target)

Automatic Value Range Handling

SenseCraftLoss automatically converts inputs to the required range for each loss:

• UNIT [0, 1]: SSIM, MS-SSIM, edge losses (Sobel, Laplacian, etc.)
• SYMMETRIC [-1, 1]: Perceptual losses (LPIPS, DINOv3, ConvNext)
• ANY: Charbonnier, MSE, L1, FFT losses

You just specify your data's range via input_range, and the conversion is handled automatically.

Using Individual Losses

You can also use losses directly:

from sensecraft.loss import (
    ConvNextDinoV3PerceptualLoss,
    CharbonnierLoss,
    PatchFFTLoss,
    SSIMLoss,
)

# Perceptual loss with DINOv3 ConvNext
loss_fn = ConvNextDinoV3PerceptualLoss(
    loss_layer=-1,      # Use last layer
    use_norm=True,      # L2 normalize features
    use_gram=False,     # Direct MSE loss
    input_range=(0, 1), # Input value range
)
loss = loss_fn(predicted, target)

# Charbonnier loss (smooth L1)
charbonnier = CharbonnierLoss(eps=1e-6)
loss = charbonnier(predicted, target)

# Patch FFT loss
fft_loss = PatchFFTLoss(patch_size=8, loss_type="l1")
loss = fft_loss(predicted, target)

# SSIM loss (returns 1 - SSIM for minimization)
ssim_loss = SSIMLoss(data_range=1.0)
loss = ssim_loss(predicted, target)

Evaluation Metrics

SenseCraft provides evaluation metrics separate from loss functions. Metrics return actual quality values (not loss formulations).

Functional API (Recommended)

The functional API is simple and auto-manages resources like LPIPS models:

from sensecraft.metrics import psnr, ssim, ms_ssim, rmse, mae, mape, lpips

# Compute metrics (tensors should be in [0, 1] range for most metrics)
pred = torch.rand(1, 3, 256, 256)
target = torch.rand(1, 3, 256, 256)

# PSNR (always in dB, higher is better)
print(f"PSNR: {psnr(pred, target):.2f} dB")

# SSIM (0-1, higher is better)
print(f"SSIM: {ssim(pred, target):.4f}")

# SSIM in dB scale (higher is better)
print(f"SSIM: {ssim(pred, target, as_db=True):.2f} dB")

# MS-SSIM (requires ~160x160 minimum image size)
print(f"MS-SSIM: {ms_ssim(pred, target):.4f}")

# Error metrics (lower is better)
print(f"RMSE: {rmse(pred, target):.4f}")
print(f"MAE: {mae(pred, target):.4f}")
print(f"MAPE: {mape(pred, target):.4f}")  # Mean Absolute Percentage Error

# LPIPS (lower is better, expects [-1, 1] range)
pred_sym = pred * 2 - 1
target_sym = target * 2 - 1
print(f"LPIPS: {lpips(pred_sym, target_sym):.4f}")

Data Range

For PSNR and SSIM metrics, data_range specifies the dynamic range:

• data_range=1.0 for images in [0, 1]
• data_range=2.0 for images in [-1, 1]
• data_range=255.0 for images in [0, 255]

# For [-1, 1] normalized images
psnr_val = psnr(pred, target, data_range=2.0)
ssim_val = ssim(pred, target, data_range=2.0)

dB Scale for SSIM/MS-SSIM

SSIM and MS-SSIM can be returned in dB scale using as_db=True:

# dB scale: -10 * log10(1 - ssim_value)
# Gives values like 15-25 dB for typical quality ranges
ssim_db = ssim(pred, target, as_db=True)  # e.g., 18.5 dB
ms_ssim_db = ms_ssim(pred, target, as_db=True)  # e.g., 22.1 dB

LPIPS Auto-Caching

The lpips() function automatically caches models and moves them to the correct device:

# First call loads the model
val1 = lpips(pred1, target1, net="alex")  # Loads AlexNet model

# Subsequent calls reuse the cached model
val2 = lpips(pred2, target2, net="alex")  # Uses cached model

# Different network types are cached separately
val3 = lpips(pred3, target3, net="vgg")  # Loads and caches VGG model

Class-based API

For repeated use with the same settings:

from sensecraft.metrics import PSNR, SSIM, MSSSIM, RMSE, MAE, MAPE

psnr_metric = PSNR(data_range=1.0)
ssim_metric = SSIM(data_range=1.0, as_db=True)

# Use like nn.Module
psnr_val = psnr_metric(pred, target)
ssim_val = ssim_metric(pred, target)

Available Metrics

Function	Class	Description	Range	Better
psnr()	PSNR	Peak Signal-to-Noise Ratio	dB	Higher
ssim()	SSIM	Structural Similarity	0-1 or dB	Higher
ms_ssim()	MSSSIM	Multi-Scale SSIM	0-1 or dB	Higher
rmse()	RMSE	Root Mean Squared Error	0+	Lower
mae()	MAE	Mean Absolute Error	0+	Lower
mape()	MAPE	Mean Absolute Percentage Error	0+	Lower
lpips()	LPIPSMetric	Learned Perceptual Similarity	0+	Lower

Available Losses

Registered Loss Names

All losses can be used with SenseCraftLoss via their registered names:

Category	Names	Value Range
Basic	mse, l1, huber, charbonnier	ANY
Regression	rmse, mape, smape, log_cosh, quantile	ANY
Robust	cauchy, geman_mcclure, welsch, tukey, wing	ANY
FFT	fft, patch_fft, gaussian_noise	ANY
Edge	sobel, laplacian, canny, gradient, high_freq, multi_scale_gradient, structure_tensor	UNIT [0,1]
SSIM	ssim, ms_ssim	UNIT [0,1]
Perceptual	lpips, convnext, dino_convnext, dino_vit	SYMMETRIC [-1,1]
Video/3D	ssim3d, stssim, tssim, fdb, temporal_accel, temporal_fft, patch_fft_3d, temporal_gradient	varies

Config Classes

For complex losses, use typed config classes:

Config Class	Loss	Key Parameters
GeneralConfig	Any loss	name, weight, **kwargs
DinoV3LossConfig	dino_vit	model_type, loss_layer, use_gram, use_norm
ConvNextDinoV3LossConfig	dino_convnext	model_type, loss_layer, use_gram, use_norm
LPIPSConfig	lpips	net ("vgg", "alex", "squeeze")
SSIMConfig	ssim	win_size, win_sigma
MSSSIMConfig	ms_ssim	win_size, win_sigma, weights
PatchFFTConfig	patch_fft	patch_size, loss_type, norm_type, use_phase

Loss Functions

Perceptual Losses

ConvNextPerceptualLoss

Uses ImageNet-pretrained ConvNext models from torchvision.

from sensecraft.loss import ConvNextPerceptualLoss
from sensecraft.loss.convnext import ConvNextType

loss_fn = ConvNextPerceptualLoss(
    model_type=ConvNextType.SMALL,      # TINY, SMALL, BASE, LARGE
    feature_layers=[2, 4, 8, 14],       # Layer indices to extract
    use_gram=False,                      # True for style/texture loss
    input_range=(-1, 1),                # Expected input range
    layer_weight_decay=1.0,             # Weight decay for layers
)

ConvNextDinoV3PerceptualLoss

Uses DINOv3 self-supervised ConvNext models (requires transformers >= 4.56.0).

from sensecraft.loss import ConvNextDinoV3PerceptualLoss
from sensecraft.loss.convnext_dinov3 import ConvNextType

# Single-layer mode (recommended)
loss_fn = ConvNextDinoV3PerceptualLoss(
    model_type=ConvNextType.SMALL,
    loss_layer=-1,                      # -1 for last layer
    use_norm=True,                      # L2 normalize features
    use_gram=False,                     # MSE on normalized features
    input_range=(0, 1),
)

# Multi-layer mode
loss_fn = ConvNextDinoV3PerceptualLoss(
    model_type=ConvNextType.SMALL,
    feature_layers=[2, 4, 8, 14, 20],   # Multiple layers
    feature_weights=[1.0] * 5,          # Optional explicit weights
    use_gram=True,                      # Gram matrix loss
)

ViTDinoV3PerceptualLoss

Uses DINOv3 Vision Transformer models for sequence-based perceptual loss.

Note: When using use_norm=True and use_gram=False, this is equivalent to the DINO perceptual loss described in NA-VAE.

from sensecraft.loss import ViTDinoV3PerceptualLoss
from sensecraft.loss.gram_dinov3 import ModelType

loss_fn = ViTDinoV3PerceptualLoss(
    model_type=ModelType.SMALL_PLUS,    # SMALL, SMALL_PLUS, BASE, LARGE
    use_norm=True,                       # L2 normalize features
    use_gram=True,                       # Gram matrix for texture
    loss_layer=-4,                       # Layer index (supports negative, default -4)
    input_range=(0, 1),
)

LPIPS

Learned Perceptual Image Patch Similarity from Zhang et al.

from sensecraft.loss import LPIPS

loss_fn = LPIPS(
    net_type="vgg",     # "vgg", "alex", "squeeze"
    version="0.1",      # "0.0" or "0.1"
)

Frequency Domain Losses

FFTLoss

Global FFT loss operating on the entire image. Computes loss on real and imaginary parts separately.

from sensecraft.loss import FFTLoss
from sensecraft.loss.general import NormType

loss_fn = FFTLoss(
    loss_type="mse",                # "mse", "l1", "charbonnier"
    norm_type=NormType.LOG1P,       # NONE, L2, LOG, LOG1P
)

PatchFFTLoss

Patch-based FFT loss for local frequency analysis.

from sensecraft.loss import PatchFFTLoss
from sensecraft.loss.general import NormType

loss_fn = PatchFFTLoss(
    patch_size=8,                   # 8x8 or 16x16 patches
    loss_type="l1",                 # "mse", "l1", "charbonnier"
    norm_type=NormType.LOG1P,       # Normalization for FFT real/imag parts
)

Normalization Types:

• NormType.NONE: No normalization (may produce very large values)
• NormType.L2: L2 normalization per patch
• NormType.LOG: Sign-preserving sign(x) * log(|x| + eps)
• NormType.LOG1P: Sign-preserving sign(x) * log(1 + |x|) (recommended)

Note: FFT losses automatically cast inputs to fp32 to avoid ComplexHalf warnings with mixed precision training.

General Losses

Standard Regression Losses

from sensecraft.loss import (
    MSELoss,           # Mean Squared Error (L2)
    L1Loss,            # Mean Absolute Error (L1)
    HuberLoss,         # Smooth L1 / Huber loss
    CharbonnierLoss,   # Differentiable L1 variant
    RMSELoss,          # Root Mean Squared Error
    MAPELoss,          # Mean Absolute Percentage Error
    SMAPELoss,         # Symmetric MAPE (more stable near zero)
    LogCoshLoss,       # Log-Cosh (smooth L1/L2 hybrid)
    QuantileLoss,      # Pinball loss for quantile regression
)

# Huber loss (smooth L1)
huber = HuberLoss(delta=1.0)  # Quadratic when |x-y| < delta, linear otherwise

# Charbonnier loss (differentiable L1)
charbonnier = CharbonnierLoss(eps=1e-3)  # L(x,y) = sqrt((x-y)^2 + eps^2)

# Log-Cosh (behaves like L2 for small errors, L1 for large)
log_cosh = LogCoshLoss()

# Quantile loss (q=0.5 is equivalent to L1/median regression)
quantile = QuantileLoss(quantile=0.5)

Robust Regression Losses (Outlier-resistant)

These losses are designed to be robust against outliers:

from sensecraft.loss import (
    CauchyLoss,        # Cauchy/Lorentzian loss
    GemanMcClureLoss,  # Bounded loss, very robust
    WelschLoss,        # Bounded with exponential decay
    TukeyBiweightLoss, # Completely ignores large outliers
    WingLoss,          # High precision for small errors (facial landmarks)
)

# Cauchy loss: log(1 + ((x-y)/scale)^2)
cauchy = CauchyLoss(scale=1.0)

# Geman-McClure: bounded loss that saturates for large errors
geman = GemanMcClureLoss(scale=1.0)

# Welsch: 1 - exp(-0.5 * ((x-y)/scale)^2)
welsch = WelschLoss(scale=1.0)

# Tukey: completely ignores outliers beyond threshold c
tukey = TukeyBiweightLoss(c=4.685)

# Wing loss: high precision for small errors, linear for large
wing = WingLoss(width=5.0, curvature=0.5)

GaussianNoiseLoss

Noise-aware loss for denoising tasks.

from sensecraft.loss import GaussianNoiseLoss

loss_fn = GaussianNoiseLoss(
    sigma=0.1,                      # Fixed noise sigma
    sigma_range=(0.01, 0.2),        # Or random range
    loss_type="l1",                 # "mse", "l1", "charbonnier"
)

# Can add noise to target during training
loss = loss_fn(predicted, target, add_noise_to_target=True)

Edge and Structure Losses

Losses for preserving edges and structural details:

from sensecraft.loss import (
    SobelEdgeLoss,
    LaplacianEdgeLoss,
    GradientLoss,
    MultiScaleGradientLoss,
    StructureTensorLoss,
)

# Sobel edge loss
sobel = SobelEdgeLoss(loss_type="l1")  # "l1", "mse", "charbonnier"
loss = sobel(predicted, target)

# Multi-scale gradient for coarse-to-fine edge matching
msg = MultiScaleGradientLoss(num_scales=3)
loss = msg(predicted, target)

# Structure tensor for texture/orientation
st = StructureTensorLoss(window_size=5, sigma=1.0)
loss = st(predicted, target)

Video/3D Losses

Losses for temporal consistency in video:

from sensecraft.loss import (
    STSSIM,
    TSSIM,
    TemporalGradientLoss,
    TemporalFFTLoss,
    FDBLoss,
)

# Spatio-temporal SSIM
stssim = STSSIM(spatial_weight=0.5, temporal_weight=0.5)
loss = stssim(video_pred, video_target)  # (B, T, C, H, W)

# Temporal gradient loss (frame differences)
tg = TemporalGradientLoss(loss_type="l1")
loss = tg(video_pred, video_target)

# Temporal FFT for frequency consistency over time
tfft = TemporalFFTLoss()
loss = tfft(video_pred, video_target)

Comparison: When to Use Which Loss

Loss Type	Best For	Characteristics
MSE	Pixel-accurate reconstruction	Simple, can be blurry
L1	General reconstruction	Less blurry than MSE
Huber	Balanced reconstruction	Quadratic near zero, linear for large errors
Charbonnier	Restoration tasks	Smooth L1, differentiable everywhere
Log-Cosh	Smooth regression	Like L2 for small, L1 for large errors
Cauchy/Welsch/Tukey	Noisy data	Robust to outliers, bounded influence
MAPE/SMAPE	Relative error	Scale-invariant, good for varied magnitudes
SSIM/MS-SSIM	Structural quality	Window-based, perceptually motivated
LPIPS	Perceptual similarity	Learned, correlates with human perception
ConvNext	Content matching	Multi-scale features
DINOv3 ConvNext	Semantic matching	Self-supervised, better generalization
DINOv3 ViT	Global structure	Transformer-based, sequence features
FFT	Frequency content	Captures textures, patterns
PatchFFT	Local frequency	Better for high-frequency details
Sobel/Gradient	Edge preservation	First-order derivatives
Laplacian	Fine details	Second-order derivatives
Structure Tensor	Texture orientation	Captures local anisotropy
Temporal losses	Video consistency	Frame-to-frame coherence

Example: Testing Distortions

The package includes an example script to compare loss and metric behavior under various distortions:

# Run the distortion test
python examples/test_distortions.py --device cuda

# Test specific image
python examples/test_distortions.py --image path/to/image.png

# Skip DINOv3 losses (faster, no transformers needed)
python examples/test_distortions.py --no-dinov3

This generates plots in results/{image_name}/:

Loss plots (losses/):

• Loss values vs distortion level
• Gradient norms vs distortion level

Metric plots (metrics/):

• PSNR, SSIM, MS-SSIM (dB scale) vs distortion level
• SSIM, MS-SSIM (0-1 scale) vs distortion level
• RMSE, MAE, MAPE, LPIPS vs distortion level

Combined plots:

• all_distortions_losses.png - Grid of all loss plots
• all_distortions_metrics.png - Grid of all metric plots

Distortion types tested:

• JPEG compression (quality 5-100)
• WebP compression (quality 5-100)
• Gaussian noise (sigma 0-0.3)
• Gaussian blur (sigma 0-7)

API Reference

Common Parameters

All perceptual losses share these parameters:

Parameter	Type	Description
input_range	Tuple[float, float]	Expected (min, max) of input values
use_gram	bool	Use Gram matrix (L1) vs direct features (MSE)
use_norm	bool	L2 normalize features before loss

DINOv3 Models

Available model types for DINOv3 losses:

ConvNext:

• ConvNextType.TINY: ~28M params
• ConvNextType.SMALL: ~50M params (recommended)
• ConvNextType.BASE: ~89M params
• ConvNextType.LARGE: ~198M params

ViT:

• ModelType.SMALL: ~22M params
• ModelType.SMALL_PLUS: Larger hidden dim
• ModelType.BASE: ~86M params
• ModelType.LARGE: ~307M params

Requirements

• Python >= 3.10
• PyTorch >= 2.0
• torchvision
• numpy

Optional:

• pytorch-msssim (for SSIM/MS-SSIM metrics and losses)
• transformers >= 4.56.0 (for DINOv3 losses)
• scikit-image (for color space conversions)
• matplotlib (for example scripts)
• Pillow (for example scripts)

License

Apache License 2.0

Citation

If you use SenseCraft in your research, please cite:

@software{sensecraft,
  author = {Shih-Ying Yeh (KohakuBlueleaf)},
  title = {SenseCraft: Unified Perceptual Feature Loss Framework},
  url = {https://github.com/KohakuBlueleaf/SenseCraft},
  year = {2024}
}

Acknowledgments

• ConvNext Perceptual Loss by sypsyp97
• LPIPS by Richard Zhang et al.
• DINOv3 by Facebook Research