Noise2Contrast: Noise-Guided Denoising

• Noise2Contrast is a research motif that uses noise as a supervisory signal, contrasting independent noise sources to recover image details and enhance restoration.
• It spans methods from self-supervised tomographic denoising with multi-contrast fusion to low-light enhancement via noise self-regression, each exploiting unique noise structures.
• The framework enables learning transferable noise representations while addressing modality-specific challenges such as registration issues and gray-world biases.

Noise2Contrast is a term with both narrow and broad usage in recent arXiv literature. In its narrowest and most explicit sense, it denotes a self-supervised tomographic denoising framework that exploits independently noisy multi-contrast acquisitions by stacking a denoiser with a domain-transfer operator and training against another contrast without clean targets (Wagner et al., 2022). In a broader, descriptive sense, the term is also used for methods that learn restoration or noise modeling from noise structure itself: low-light enhancement via noise self-regression in NoiSER (Zhang et al., 2022), contrastive noise embedding and single-reference noise transfer in NoiseTransfer (Lee et al., 2023), and, less formally, contrastive estimation of camera-specific noise parameters for realistic RAW noise synthesis (Zou et al., 2022). Across these usages, the recurring theme is that noise is not treated merely as corruption to be removed, but as a supervisory signal, an alignment cue, or an identifiable distributional object.

1. Terminology and scope

The term “Noise2Contrast” is not used uniformly. One paper adopts it as the official method name, whereas others use it as a paradigm label or as an interpretive shorthand for related contrastive or self-regressive ideas.

Context	Paper	Use of the term
Tomographic denoising	"Noise2Contrast: Multi-Contrast Fusion Enables Self-Supervised Tomographic Image Denoising" (Wagner et al., 2022)	Official method name
Low-light enhancement	"Noise Self-Regression: A New Learning Paradigm to Enhance Low-Light Images Without Task-Related Data" (Zhang et al., 2022)	Described as a Noise2Contrast paradigm
Noise generation	"NoiseTransfer: Image Noise Generation with Contrastive Embeddings" (Lee et al., 2023)	Connected to a Noise2Contrast paradigm
RAW noise estimation	"Estimating Fine-Grained Noise Model via Contrastive Learning" (Zou et al., 2022)	Not officially named Noise2Contrast

This variation matters because the underlying technical objects being “contrasted” differ substantially. In tomographic denoising, the contrast is literal imaging contrast across co-registered acquisitions. In NoiSER, the central mechanism is self-regression on pure Gaussian noise, with contrast referring to local differences between adjacent pixels. In NoiseTransfer and the fine-grained RAW noise model paper, contrastive learning is used to separate noise distributions or noise parameters in representation space (Wagner et al., 2022, Zhang et al., 2022, Lee et al., 2023, Zou et al., 2022).

A plausible implication is that “Noise2Contrast” functions less as a single algorithmic family than as a research motif: methods use noise-derived structure to recover image contrast, to exploit multi-contrast independence, or to infer transferable noise representations.

2. Noise self-regression and low-light image enhancement

In low-light image enhancement, NoiSER formulates a Noise2Contrast paradigm in which a compact convolutional neural network is trained only on pure Gaussian noise patches and then applied to real low-light inputs at inference (Zhang et al., 2022). The target task is LLIE, seeking a mapping $f_\theta$ from a low-light image $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$ to a normal-light image $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$ with corrected illumination, preserved structure, and natural color. Training uses no task-related data. Instead, each training sample is i.i.d. Gaussian noise,

$I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$

and the network is optimized by self-regression,

$\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$

with $x=I_n$ and a loss composed of $\ell_1$ reconstruction and TV regularization:

$$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$

The architecture is a lightweight self-regression model with approximately $1$K parameters, two non-linear activation layers, instance normalization, and no shortcuts or residual connections. The role of instance normalization is central. For each instance and channel,

$$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$

with learnable $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$0 and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$1. The reported interpretation is threefold: self-regression reconstructs local contrast between adjacent pixels, instance normalization remediates global illumination, and the zero-mean Gaussian training distribution biases outputs toward the gray-world hypothesis when the image is sufficiently large (Zhang et al., 2022).

The method’s reported behavior depends on variance choice and stopping time. Standard Gaussian training with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$2 is used in NoiSER-FC and NoiSER-ES, while NoiSER-Var3 uses $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$3 for improved visual effect. Inputs and outputs are normalized to $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$4, where $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$5 corresponds to the mid-tone $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$6 in $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$7. Optimization uses Adam with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$8, $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$9, learning rate $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$0, batch size $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$1, and $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$2 noise patches. Full convergence uses $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$3 iterations, whereas an early-stopping variant uses approximately $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$4 iterations to trade slight quantitative performance for better visual clarity (Zhang et al., 2022).

On LOL, the reported results are: NoiSER-FC with PSNR $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$5, SSIM $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$6 and NIQE $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$7; NoiSER-ES with PSNR $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$8, SSIM $$I_{\text{norm}} \in \mathbb{R}^{H \times W \times 3}$$9 and NIQE $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$0; and NoiSER-Var3 with PSNR $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$1, SSIM $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$2 and NIQE $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$3. Training and inference costs are also reported: NoiSER-FC requires approximately $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$4 min training time, NoiSER-ES approximately $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$5 min, inference is approximately $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$6 ms, and the model has approximately $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$7 parameters. Cross-dataset evaluation on LSRW reports Huawei SSIM $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$8 for NoiSER-FC and Nikon PSNR $I_n(u,v,c) \sim \mathcal{N}(0,\sigma^2),$9 for NoiSER-ES, with NoiSER-Var3 achieving Nikon SSIM $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$0. The same mapping is also reported to suppress overexposure on SICE, because the gray-world pull and instance-normalization rescaling move extreme brightness toward more moderate outputs (Zhang et al., 2022).

The principal limitations follow directly from the mechanism. Scenes that violate the gray-world hypothesis may be over-neutralized, very small images weaken the law-of-large-numbers argument on channel means, heavily sensor-noisy inputs may have noise emphasized rather than removed, and full convergence at $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$1 can introduce slightly gray veiling. The paper’s practical guidance is correspondingly conservative: retain instance normalization, keep the compact no-shortcut design, include TV regularization, and use early stopping or $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$2 when visual crispness is more important than maximal PSNR or SSIM (Zhang et al., 2022).

3. Multi-contrast fusion for self-supervised tomographic denoising

In its most specific usage, Noise2Contrast is a self-supervised denoising framework for tomographic imaging that exploits routine multi-contrast acquisitions such as MRI T1, T2, and FLAIR or dual-energy CT/X-ray microscopy (Wagner et al., 2022). The formal setting assumes images

$\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$3

with zero-mean noise and independence across contrasts for $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$4:

$\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$5

The framework introduces a denoiser $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$6 that maps a noisy image to a denoised image in the same contrast, and a domain-transfer operator $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$7 that maps clean content from contrast $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$8 to contrast $\arg\min_\theta \mathbb{E}_x \{L(f_\theta(x),x)\},$9. The self-supervised loss is

$x=I_n$0

with MSE as the loss function.

The key methodological point is separation of denoising from contrast transfer. Two denoising realizations are studied. The first uses three stacked trainable bilateral filter layers with learnable spatial and range kernel widths; by construction, these filters cannot perform contrast transfer and therefore preserve task separation. The second uses a network operator-based denoiser such as a U-Net, but this is trained sequentially with a frozen domain-transfer network to avoid the denoiser “cheating” by altering contrast rather than denoising (Wagner et al., 2022).

The domain-transfer network is a standard U-Net with about $x=I_n$1 million parameters. Training uses Adam with learning rate $x=I_n$2 and proceeds until self-supervised validation loss or training loss converges, depending on modality. The reported MRI pairings are T1$x=I_n$3T2, T2$x=I_n$4T1, and T2$x=I_n$5FLAIR; the reported XRM pairings are LE$x=I_n$6HE and HE$x=I_n$7LE. No masking or blind-spot mechanism is required, because supervision comes from an independently noisy target contrast rather than from pixel-wise independence assumptions (Wagner et al., 2022).

The reported results show that Noise2Contrast with bilateral filters outperforms the noisy baseline and the compared self-supervised alternatives. On brain MRI, T1$x=I_n$8T2 improves from baseline PSNR/SSIM $x=I_n$9 and $\ell_1$0 to $\ell_1$1 and $\ell_1$2; T2$\ell_1$3T1 improves from $\ell_1$4 and $\ell_1$5 to $\ell_1$6 and $\ell_1$7; T2$\ell_1$8FLAIR improves from $\ell_1$9 and $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$0 to $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$1 and $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$2. The paper summarizes these as gains of $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$3–$$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$4 in PSNR and $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$5–$$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$6 in SSIM over the noisy baseline. On dual-energy XRM, LE$$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$7HE improves from $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$8 and $$\mathcal{L}_{\text{total}}=\mathcal{L}_{\ell_1}(f_\theta(I_n),I_n)+\mathcal{L}_{tv}(f_\theta(I_n)).$$9 to $1$0 and $1$1, while HE$1$2LE improves from $1$3 and $1$4 to $1$5 and $1$6, corresponding to reported gains of $1$7–$1$8 in PSNR and $1$9–$$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$0 in SSIM (Wagner et al., 2022).

The method’s assumptions are explicit. Accurate alignment between contrasts is important; imperfect registration pushes the transfer network to absorb geometric compensation. The fidelity of $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$1 matters, because an imperfect transfer operator can bias the gradients seen by the denoiser. The method also requires at least two contrasts per subject. Its advantage over blind-spot methods is that it does not rely on pixel-wise independent noise, and the XRM results are presented precisely in a regime where real correlated noise undermines that assumption (Wagner et al., 2022).

4. Contrastive noise embeddings and single-reference transfer

NoiseTransfer extends the broader Noise2Contrast idea from denoising to noise generation (Lee et al., 2023). Its objective is not to denoise directly, but to learn a single generative model that can synthesize noisy images with multiple different noise distributions and can transfer the noise characteristics of a single reference noisy image to any clean image. The method learns a $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$2-dimensional noise embedding with a contrastive encoder branch $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$3, using positives from the same noise distribution and negatives from different distributions. A momentum key encoder $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$4 and a MoCo-style queue of size $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$5 support InfoNCE training with temperature $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$6:

$$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$7

The full architecture contains a discriminator with two branches, $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$8 and $$y_c=\gamma_c \cdot \frac{x_c-\mu_c}{\sqrt{\sigma_c^2+\epsilon}}+\beta_c,$$9, and a generator $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$00 conditioned on a clean image $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$01 and the reference-noise embedding $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$02. The generator outputs a synthesized noisy image $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$03. Training combines contrastive, adversarial, feature-matching, and Gaussian-filtered reconstruction losses. The discriminator objective is

$$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$04

and the generator objective is

$$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$05

with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$06 (Lee et al., 2023).

Training uses SIDD-Medium, SIDD validation, SIDD+, DND, DIV2K, and BSDS500. Synthetic noise includes Gaussian with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$07, Poisson with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$08, and Poisson-Gaussian mixtures over those ranges. Mini-batches contain $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$09 patches of size $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$10, with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$11 real SIDD patches and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$12 synthetic patches. Optimization uses Adam with learning rate $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$13, $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$14, $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$15, and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$16 regularization $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$17. The model is trained for $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$18 epochs, with discriminator and generator each updated $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$19 times per epoch, taking approximately one week on two Tesla V100 GPUs (Lee et al., 2023).

The reported results concern both noise-generation fidelity and downstream denoising. On SIDD validation, NoiseTransfer achieves AKLD/KS of $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$20, compared with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$21 for DANet and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$22 for CycleISP; on SIDD+, it reports $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$23. When RIDNet is trained on NoiseTransfer-generated pairs, the reported denoising performance is $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$24 on SIDD validation, $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$25 on SIDD+, and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$26 on DND, approaching the corresponding ground-truth-trained references and exceeding CycleISP in the reported comparisons (Lee et al., 2023).

The broader connection to Noise2Contrast lies in what is being contrasted: noisy images from the same versus different noise distributions. This differs from the tomographic setting, where the contrast is a paired acquisition of the same anatomy under different imaging settings, and from NoiSER, where the operative object is local adjacent-pixel contrast reconstructed through self-regression. The paper also reports failure of ablations that remove the discriminator-side noise contrastive loss: removing $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$27 causes divergence by epoch $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$28, while adding $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$29 and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$30 stabilizes training and yields the best AKLD/KS (Lee et al., 2023).

5. Fine-grained RAW noise estimation via contrastive learning

The paper "Estimating Fine-Grained Noise Model via Contrastive Learning" is not officially named Noise2Contrast, and the data explicitly notes that the nickname is not used by the authors (Zou et al., 2022). Nevertheless, it is closely related conceptually because it uses contrastive learning to isolate noise statistics rather than scene content and then synthesizes realistic noise for downstream denoiser training.

The method operates in RAW space and models noise with a fine-grained, physics-based parameterization

$$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$31

combining Poisson shot noise, Gaussian readout noise with non-zero mean, Gaussian row noise, and a multiplicative gain. With clean RAW $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$32, the generative form is

$$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$33

where $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$34 and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$35. Under a heteroscedastic Gaussian approximation, the noise level function becomes

$$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$36

To emphasize noise statistics, the input noisy RAW image is first transformed by a Haar wavelet transform $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$37. A ResNet encoder and MLP projection head are trained with InfoNCE-style contrastive learning, where positives share the same noise parameters but differ in scene content, and negatives use different parameter tuples. A regression head predicts $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$38, and the full objective combines weighted MSE on transformed parameters with a contrastive term of weight $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$39. Training uses Adam with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$40, learning rate $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$41 decayed by $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$42 every $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$43 epochs, batch size $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$44, and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$45 epochs on an NVIDIA RTX 3090 (Zou et al., 2022).

The estimator is then applied to noisy test images from a target camera, without calibration frames or paired data, to fit camera-level relationships such as

$$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$46

and optionally $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$47 with ISO $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$48. These inferred relationships define a camera-specific sampling distribution for realistic RAW noise synthesis. The intended use is to generate synthetic noisy/clean pairs for denoiser training, or to provide a non-blind denoiser with an estimated noise level function (Zou et al., 2022).

The reported synthesis fidelity on SIDD is strong: average KL divergence between noise histograms is $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$49 for the proposed method, compared with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$50 for AWGN, $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$51 for Poisson-Gaussian, $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$52 for NoiseFlow, and $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$53 for CANGAN. On SIDD S6 denoising with a U-Net trained on synthesized pairs, the reported “All ISOs” result is $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$54, compared with $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$55 for paired real data; the ablation without contrastive loss drops to $$I_{\text{low}} \in \mathbb{R}^{H \times W \times 3}$$56 (Zou et al., 2022).

6. Common principles, distinctions, and recurring limitations

Taken together, these works suggest that Noise2Contrast is best understood as a family resemblance across different imaging problems rather than a single canonical recipe. The unifying principle is that noise contains exploitable structure. In NoiSER, the structure is the abundance of local differences in Gaussian noise coupled with instance normalization and a gray-world bias (Zhang et al., 2022). In tomographic denoising, the structure is the independence of noise across co-registered imaging contrasts, which turns another noisy contrast into a valid self-supervised target (Wagner et al., 2022). In NoiseTransfer and the RAW noise model paper, the structure is distributional: noisy images or synthesized parameter-matched samples can be contrasted so that the representation becomes sensitive to the underlying noise process rather than scene semantics (Lee et al., 2023, Zou et al., 2022).

The distinctions are equally important. NoiSER is a compact model trained once on synthetic Gaussian noise and then applied broadly, not a per-image fitting procedure. The tomographic Noise2Contrast method is not a generic contrastive-learning framework but a denoiser-plus-transfer architecture whose validity depends on cross-contrast independence and anatomical alignment. NoiseTransfer is a generative model for producing realistic noisy images from a single noisy reference. The fine-grained RAW method is a parameter estimator and synthesis pipeline rather than a restoration network (Zhang et al., 2022, Wagner et al., 2022, Lee et al., 2023, Zou et al., 2022).

The recurring failure modes are also modality-specific. Gray-world assumptions can over-neutralize genuine color casts in low-light enhancement; misregistration and imperfect domain transfer can bias tomographic denoising; rare or highly structured artifacts may be underrepresented in learned noise embeddings for noise generation; and RAW-domain physics-based synthesis does not model the ISP or non-Bayer CFA variants (Zhang et al., 2022, Wagner et al., 2022, Lee et al., 2023, Zou et al., 2022). A plausible implication is that future uses of the term will continue to diverge unless authors specify precisely what is being contrasted: image contrasts, adjacent-pixel contrasts, noisy-distribution embeddings, or noise-parameter classes.