PHATNet: Physics-guided Haze Transfer Network

• The paper introduces a domain-adaptive framework leveraging latent-space ASM to generate target-specific haze-transfer training pairs.
• It disentangles haze from content using dual-branch encoders and multi-scale fusion to preserve clean image details while injecting realistic haze.
• Optimizing with haze-transfer consistency and content-leakage losses, PHATNet significantly improves PSNR performance across various dehazing models.

Searching arXiv for PHATNet and closely related dehazing papers to ground the article. Physics-guided Haze Transfer Network (PHATNet) is a domain-adaptive framework for real-world image dehazing that transfers haze patterns from unseen target domains to source-domain haze-free images, thereby creating domain-specific fine-tuning sets for adapting an already trained dehazing model without paired target supervision. Rather than treating dehazing as a direct mapping from hazy input to clean output, PHATNet is organized around the premise that estimating haze is generally easier than recovering clean content, and it uses the Atmospheric Scattering Model (ASM) as an inductive bias inside a latent parametric haze-transfer mechanism (Tsai et al., 20 Jul 2025).

1. Problem setting and conceptual scope

PHATNet addresses real-world dehazing under domain shift. In this setting, a dehazing model trained on one dataset often performs poorly on unseen real-world hazy images from a different domain because haze patterns vary across locations, weather conditions, camera devices, and scene content. The method is therefore framed as test-time domain adaptation for dehazing: adapt an already trained model to a new, unseen hazy target domain without requiring paired target supervision (Tsai et al., 20 Jul 2025).

The method is motivated by two observations stated explicitly in the underlying work. First, real-world haze is hard to simulate accurately with synthetic scattering models. Second, paired real hazy/clean images are scarce, so models trained on limited paired datasets fail to generalize well to unseen haze distributions. PHATNet responds by constructing a target-specific fine-tuning set whose haze appearance comes from real target hazy images while scene content comes from source clean images. This yields synthetic but domain-matched hazy/clean pairs suitable for updating downstream dehazing models (Tsai et al., 20 Jul 2025).

PHATNet is therefore not primarily a direct dehazing network. Its primary role is to act as a domain adapter: it generates target-style hazy training pairs for fine-tuning, after which a separate dehazing model performs the final restoration. This differentiates it from direct $I \rightarrow J$ dehazing architectures and from test-time training methods that optimize auxiliary tasks during inference. Because the adaptation is performed offline, inference latency of the final dehazer does not increase (Tsai et al., 20 Jul 2025).

2. Physical formulation and haze transfer principle

The physical prior underlying PHATNet is the classical Atmospheric Scattering Model: $I(x) = J(x)\times t(x) + A\times (1-t(x)),$ with

$t(x) = e^{-\beta d(x)},$

where $I(x)$ is the hazy image, $J(x)$ is the haze-free scene radiance, $A$ is global atmospheric light, $t(x)$ is the transmission map, $\beta$ is haze density, and $d(x)$ is scene depth (Tsai et al., 20 Jul 2025).

PHATNet does not apply this model directly in image space. Instead, it extends the ASM into a latent parametric haze space. The paper’s rationale is that image-domain ASM is not sufficiently flexible for real-world haze because real haze is non-uniform, region-specific, affected by content and depth, and difficult to transfer directly with naive image-to-image translation. PHATNet therefore performs haze transfer in feature space, where haze and content can be disentangled more effectively (Tsai et al., 20 Jul 2025).

The central fusion rule is an ASM-like operation in latent space: $$F^I = F^J \times F^\mathrm{TM} + F^\mathrm{AL} \times (1 - F^\mathrm{TM}), \tag{1}$$ where $I(x) = J(x)\times t(x) + A\times (1-t(x)),$0 is the clean content feature, $I(x) = J(x)\times t(x) + A\times (1-t(x)),$1 is the latent transmission feature, $I(x) = J(x)\times t(x) + A\times (1-t(x)),$2 is the latent atmospheric-light feature, and $I(x) = J(x)\times t(x) + A\times (1-t(x)),$3 is the rehazed feature representation (Tsai et al., 20 Jul 2025).

This formulation mirrors the standard scattering equation while moving the transfer process into learned features rather than raw pixels. The paper argues that this latent-space ASM reduces ghosting artifacts caused by depth mismatch and allows transmission maps to function as haze-aware attention maps. A plausible implication is that PHATNet uses the physics prior not as a rigid forward model, but as a structured constraint on representation disentanglement (Tsai et al., 20 Jul 2025).

3. Architecture: PHDT and multi-scale transfer

The core module of PHATNet is Parametric Haze Disentanglement and Transfer (PHDT), a dual-branch network. One branch extracts haze features from a hazy image $I(x) = J(x)\times t(x) + A\times (1-t(x)),$4; the other extracts content features from a clean image $I(x) = J(x)\times t(x) + A\times (1-t(x)),$5. After latent-space fusion, a Rehazing Encoder reconstructs the haze-transferred image $I(x) = J(x)\times t(x) + A\times (1-t(x)),$6 (Tsai et al., 20 Jul 2025).

Component	Input	Function
ALE	$I(x) = J(x)\times t(x) + A\times (1-t(x)),$7	Extracts atmospheric-light-related features
TME	$I(x) = J(x)\times t(x) + A\times (1-t(x)),$8	Extracts transmission-related features
CE	$I(x) = J(x)\times t(x) + A\times (1-t(x)),$9	Extracts clean content features
RE	$t(x) = e^{-\beta d(x)},$0	Maps fused latent features back to image space

In the haze branch, the Atmospheric Light Encoder (ALE) produces

$t(x) = e^{-\beta d(x)},$1

with $t(x) = e^{-\beta d(x)},$2. ReLU is applied before the exponential normalization to ensure nonnegative outputs. The Transmission Maps Encoder (TME) produces

$t(x) = e^{-\beta d(x)},$3

with $t(x) = e^{-\beta d(x)},$4. The exponential constrains values to $t(x) = e^{-\beta d(x)},$5, consistent with transmission semantics. In the content branch, the Content Encoder (CE) extracts

$t(x) = e^{-\beta d(x)},$6

with $t(x) = e^{-\beta d(x)},$7 (Tsai et al., 20 Jul 2025).

After fusion, the Rehazing Encoder maps the latent rehazed feature back into RGB space: $t(x) = e^{-\beta d(x)},$8 The output should preserve the content of $t(x) = e^{-\beta d(x)},$9 while inheriting haze patterns from $I(x)$0 (Tsai et al., 20 Jul 2025).

To address spatially non-homogeneous haze, PHATNet uses a multi-scale formulation: $I(x)$1

$I(x)$2

$I(x)$3

where $I(x)$4 denotes downsampling by $I(x)$5, $I(x)$6 denotes downsampling by $I(x)$7, and $I(x)$8 denotes upsampling by $I(x)$9. The reported ablation shows that a 3-stage design performs better than 2-stage and 1-stage variants, supporting the claim that haze is multi-scale and non-uniform (Tsai et al., 20 Jul 2025).

4. Disentanglement objectives

PHATNet is optimized using two losses designed to separate haze from content: the Haze-Transfer-Consistency Loss $J(x)$0 and the Content-Leakage Loss $J(x)$1. The total objective is

$J(x)$2

These losses replace adversarial training with explicit constraints on transferred haze consistency and content isolation (Tsai et al., 20 Jul 2025).

For a paired source sample, the transferred image is

$J(x)$3

The Haze-Transfer-Consistency Loss is defined as

$J(x)$4

where $J(x)$5 indexes the three scales. This loss enforces that the transferred output remain consistent with the haze style of the hazy reference image, thereby encouraging meaningful haze representations in the haze branch (Tsai et al., 20 Jul 2025).

The Content-Leakage Loss is built from two clean images $J(x)$6 and $J(x)$7, $J(x)$8, using

$J(x)$9

Because the first input is clean and therefore contains no haze, the output should ideally equal the content image: $A$0 The resulting loss is

$A$1

Its function is to prevent the haze encoders from extracting scene content instead of pure haze. The paper reports that adding $A$2 consistently improves PSNR across datasets, especially on O-Haze, which contains sparser haze and is more prone to content interference (Tsai et al., 20 Jul 2025).

This objective design is closely tied to the method’s non-GAN character. The paper explicitly contrasts PHATNet with GAN-based haze synthesis and domain translation methods, arguing that those methods may suffer from mode collapse, unstable optimization, poor handling of non-uniform haze, and weak interpretability. PHATNet instead uses a physically guided transfer mechanism and disentanglement-specific losses (Tsai et al., 20 Jul 2025).

5. Domain adaptation procedure and operational workflow

PHATNet operates in two phases. First, it is trained on source-domain paired hazy/clean images: $A$3 Second, it is used to transfer haze from target hazy images $A$4 onto source clean images $A$5, generating

$A$6

This produces a synthetic adaptation set

$A$7

with $A$8 hazy/clean pairs (Tsai et al., 20 Jul 2025).

This dataset is then used to fine-tune an existing dehazing model. The paper evaluates this procedure with FocalNet, Dehamer, MITNet, and SGDN. During testing, each model is fine-tuned for one epoch on the target-specific fine-tuning set. Because the generated set is constructed offline, the adaptation procedure does not increase inference cost for the final dehazer (Tsai et al., 20 Jul 2025).

The PHATNet training configuration reported in the paper is: 1000 epochs, batch size $A$9, input resolution $t(x)$0, Adam optimizer, and initial learning rate $t(x)$1 cosine annealed to $t(x)$2. The framework itself has 26M parameters and requires 0.153 seconds to generate a $t(x)$3 haze-transferred image (Tsai et al., 20 Jul 2025).

The datasets used span seven real-world dehazing benchmarks: NH-Haze20, NH-Haze21, HD-NH-Haze, DenseHaze, I-Haze, O-Haze, and RTTS. Two source-domain settings are reported. In Setting 1, the source domain is NH-Haze20, trained on the first 50 pairs and validated on the last 5. In Setting 2, the source domain is HD-NH-Haze, trained on the first 35 pairs and validated on the last 5. Target domains include NH-Haze21, HD-NH-Haze, DenseHaze, I-Haze, O-Haze, and RTTS in Setting 1, and NH-Haze20, NH-Haze21, DenseHaze, I-Haze, O-Haze, and RTTS in Setting 2 (Tsai et al., 20 Jul 2025).

6. Empirical performance, limitations, and relation to adjacent physics-guided methods

On paired datasets, PHATNet is evaluated with PSNR and SSIM; on RTTS, which has no clean references, it is evaluated with NIQE and BRISQUE. The reported average PSNR gains in Setting 1 are $t(x)$4 dB for FocalNet, $t(x)$5 dB for Dehamer, $t(x)$6 dB for MITNet, and $t(x)$7 dB for SGDN. In Setting 2, the average PSNR gains are $t(x)$8 dB for FocalNet, $t(x)$9 dB for Dehamer, $\beta$0 dB for MITNet, and $\beta$1 dB for SGDN. On RTTS, PHATNet improves both NIQE and BRISQUE across all tested models and both source domains (Tsai et al., 20 Jul 2025).

Qualitatively, the haze-transferred images are reported to preserve clean image content while carrying realistic haze patterns from the target domain, including region-specific haze rather than uniform synthetic haze. The dehazed outputs obtained after fine-tuning on PHATNet-generated data show fewer residual haze artifacts and better contrast and color. The paper also reports that PHATNet achieves the best PSNR overall in comparisons with TMD, PTTD, HTFANet, and D4+ (Tsai et al., 20 Jul 2025).

Ablation studies clarify the functional role of PHATNet’s components. ALE or TME alone helps, but ALE + TME gives the best PSNR, and TME appears especially important. Performance improves as more PHDT stages are used, with the 3-stage model performing best. Adding the Content-Leakage Loss improves PSNR consistently, especially on O-Haze. These results support the paper’s claim that both parametric haze disentanglement and explicit content suppression are necessary for effective haze transfer (Tsai et al., 20 Jul 2025).

The reported limitations are also specific. Haze augmentation in parametric space is said to remain limited by incomplete prior knowledge of real-world haze distributions. The paper notes that while gamma correction and flipping can augment transfer, a richer physics-based augmentation space remains future work. PHATNet also still relies on a source-domain paired dataset to train the transfer model (Tsai et al., 20 Jul 2025).

Within the broader literature, PHATNet occupies a distinct position among physics-guided dehazing methods. PANet is a physics-guided parametric augmentation framework that learns haze density and atmospheric light maps and resamples them pixel-wise to generate additional realistic hazy/clean training pairs, primarily to enrich dehazing datasets rather than to perform target-specific adaptation at test time (Chang et al., 2024). The variational Bayesian dehazing framework in "Deep Variational Bayesian Modeling of Haze Degradation Process" treats both the clean image and transmission map as latent variables under the scattering model and couples dehazing and transmission networks via a joint variational objective; this suggests a PHATNet-style affinity at the level of physics-guided factorization, but its emphasis is posterior inference rather than domain-adaptive haze transfer (Im et al., 2024). DGFDNet, by contrast, is a prior-guided dual-domain dehazing network that injects a dark channel prior into spatial and frequency processing with closed-loop prior correction; this places it near PHATNet in the broader class of physics-guided haze-aware methods, but its objective is direct dehazing rather than construction of target-specific fine-tuning pairs (Zheng et al., 15 Jul 2025).

A common misconception is to treat PHATNet as simply another haze synthesis network. The reported formulation is narrower and more specific: PHATNet is a domain adaptation mechanism whose synthetic outputs are not an end in themselves, but a means of updating a dehazer for a new target domain. Another misconception is to view it as a GAN-based translation model. The method is explicitly positioned against GAN instability and instead relies on ASM-guided latent transfer plus $\beta$2 and $\beta$3 for disentanglement and transfer control (Tsai et al., 20 Jul 2025).