Detail-Aware Expert Module (DAEM)

• DAEM is a decoder-side module that adaptively refines high-frequency textures by leveraging sparse top-1 expert routing and multi-scale receptive fields.
• It fuses decoder latent features with skip-connected encoder cues to overcome VAE compression losses and mitigate iterative sampling blurring.
• Empirical results show that incorporating DAEM boosts PSNR and perceptual scores, demonstrating its effectiveness in preserving fine details.

Searching arXiv for the specified paper and closely related expert/detail-aware works. The Detail-Aware Expert Module (DAEM) is a decoder-side mixture-of-experts refinement mechanism introduced in the unified image restoration framework UniLDiff to address a specific weakness of latent diffusion restoration: high-frequency detail loss caused by VAE compression and by the progressive nature of iterative sampling, which can yield blurred textures, incomplete structures, and texture hallucinations. In its exact arXiv usage, DAEM denotes a module that leverages decoder features, skip-connected encoder cues, sparse expert routing, and a shared global branch to enhance texture and fine-structure recovery, particularly under spatially heterogeneous degradations (Cheng et al., 31 Jul 2025). In a broader research sense, the term also serves as a useful organizing concept for a family of expert-style designs that selectively preserve subtle local cues, dynamically weight specialized branches, or adapt expert capacity according to fine-grained structure, although many of those neighboring methods do not use the exact name “DAEM.”

1. Definition, scope, and conceptual boundaries

In the precise terminology of UniLDiff, DAEM is a decoder-side detail refinement module inside a latent diffusion restoration system. Its stated role is to “selectively enhance high-frequency structures and texture fidelity across spatially diverse regions,” to “retrieve fine-grained cues from encoder skip connections,” and to adapt restoration behavior through a mixture-of-experts design (Cheng et al., 31 Jul 2025). The module is therefore neither a generic MoE block nor a diffusion-step conditioning mechanism. It belongs specifically to the reconstruction stage of the model, after latent denoising.

Two recurring misunderstandings are clarified by the source material. First, DAEM is not the same component as UniLDiff’s Degradation-Aware Feature Fusion (DAFF) module. DAFF operates in the diffusion UNet and addresses degradation perception and adaptation, whereas DAEM is applied in the decoder to enhance high-frequency textures and preserve fine-grained structural details (Cheng et al., 31 Jul 2025). Second, DAEM is not defined by explicit expert-count scaling or sparse token routing in the transformer-MoE sense. Its defining property is detail-aware adaptive reconstruction using expert branches with varied receptive fields and a router that selects among them according to local context.

A broader literature reading suggests two useful scopes. In the strict scope, DAEM refers only to UniLDiff’s named module. In the extended scope, DAEM can denote an Editor’s term for architectures that combine detail-sensitive priors, specialized branches, and adaptive fusion to preserve local artifacts, structures, or semantic granularity. This broader usage is interpretive rather than nominal; several closely related systems are DAEM-like in function but not in name.

2. Motivation: why detail-aware expert refinement is needed

UniLDiff motivates DAEM from three concrete limitations of latent diffusion restoration. The first is that latent diffusion models “suffer from detail loss due to the high compression of VAE encoders,” which weakens or discards high-frequency signals. The second is that the progressive nature of iterative sampling can produce “blurred textures and incomplete structures,” and later “texture hallucinations.” The third is that, in all-in-one restoration settings, “spatially heterogeneous degradations challenge the expressiveness of a single reconstruction path” (Cheng et al., 31 Jul 2025).

These points define the functional problem that DAEM addresses. Restoration quality is not exhausted by globally plausible content; it depends on recovering textures, edges, small structures, text, and facial contours. A uniform decoder path is insufficient when one spatial region requires local texture recovery and another requires broader structural reconstruction. DAEM is introduced precisely because the paper treats decoder refinement as the stage where those losses become most visible.

A related but distinct argument appears in DR.Experts for blind image quality assessment. That work does not introduce a DAEM under the exact name, but it argues that unified image-feature regression is insensitive to “subtle distortion cues,” partly because distortion evidence is weak, local, heterogeneous, and easily entangled with semantics. Its solution couples a Distortion-Saliency Differential Module (DSDM) with a Dynamic Distortion Weighting Module (DDWM), yielding what the paper explicitly describes as a distortion-aware, saliency-refined, dynamically weighted expert framework (Fu et al., 10 Feb 2026). This suggests a broader DAEM principle: detail-aware expert behavior often emerges when a model first isolates subtle local evidence and then lets specialized channels influence the final prediction according to estimated importance.

The same general design pressure appears in other domains. DeRainMamba argues that Mamba-based deraining has “limited ability to capture fine-grained details” and therefore pairs a frequency-aware state-space branch with MDPConv, a multi-directional detail-enhancement branch that restores local structures by capturing anisotropic gradient features (Zhu et al., 8 Oct 2025). TDATR, in document understanding, argues that end-to-end table recognition benefits when the model first learns table details through auxiliary structure/content tasks and only then fuses them into HTML generation (Qin et al., 24 Mar 2026). Across these settings, the motivating pattern is consistent: when subtle, local, or semantically entangled evidence is decisive, a single homogeneous computation path becomes insufficient.

3. Canonical DAEM architecture in UniLDiff

DAEM in UniLDiff is an MoE-based decoder refinement module built from five stated ingredients: decoder latent features, skip-connected encoder features, a lightweight routing function, multiple expert branches built from NAFBlocks with varied receptive fields, and a shared global branch using transposed self-attention (Cheng et al., 31 Jul 2025). The figure caption describes it as an architecture “which uses experts with varied receptive fields to adaptively refine details from skip-connected encoder features.”

The input structure is central. DAEM “leverages decoder features and skip-connected encoder cues,” and the skip path is explicitly used to “retrieve early-stage, high-resolution cues that are otherwise lost in the VAE compression process” (Cheng et al., 31 Jul 2025). The paper does not provide exact tensor sizes or the exact merge operator between decoder and skip features, but it is explicit that decoder features alone are not treated as sufficient.

The routing mechanism is sparse. The paper defines

$$\mathrm{Router}(x)=\mathrm{top}\text{-}k\left(\mathrm{Softmax}(Wx+\boldsymbol{\xi})\right),$$

where $x$ is the DAEM input feature, $W$ is the router weight matrix, and $\boldsymbol{\xi}\sim\mathcal{N}(0,\sigma^2)$ is Gaussian noise (Cheng et al., 31 Jul 2025). The authors state that, in experiments, $k=1$. In operational terms, DAEM therefore uses top-1 routing: only one expert branch is activated for a routed feature.

Each expert $E^i$ is built using lightweight NAFBlocks with varied receptive fields, which the paper says allows “multi-scale perception of both fine and coarse structures” (Cheng et al., 31 Jul 2025). The exact number of experts, kernel sizes, dilation rates, and NAFBlock counts are not provided in the supplied text. What is explicit is the intended specialization: different experts correspond to different receptive-field regimes, so routing chooses among local-detail versus broader-structure processing modes according to local context.

DAEM also includes a shared global branch $S(\cdot)$ with transposed self-attention. Its output modulates expert responses as

$\hat{y}_E^i = E^i(x)\otimes S(x),$

with $\otimes$ denoting element-wise multiplication (Cheng et al., 31 Jul 2025). This is an important architectural distinction: DAEM is not merely a bank of local convolutional experts. It combines local expert specialization with a global branch intended to maintain semantic coherence. The paper does not print a final residual fusion equation, so the safest characterization is that the modulated expert output is fed onward in the decoder.

This architecture yields a specific interpretation of “detail awareness.” It is not only the presence of small receptive fields. It is the combination of: retrieval of high-resolution skip cues, sparse expert selection, receptive-field diversity, and global modulation that constrains local enhancement.

4. Mathematical formulation and training dynamics

The two explicit DAEM equations in UniLDiff are the router and the globally modulated expert output: $$\mathrm{Router}(x)=\mathrm{top}\text{-}k\left(\mathrm{Softmax}(Wx+\boldsymbol{\xi})\right),$$

$x$0

The first equation indicates that expert routing is stochastic during training through Gaussian noise and sparse through top-$x$1 selection; the second shows that expert output is multiplicatively gated by a shared global branch (Cheng et al., 31 Jul 2025).

DAEM is trained in Stage 2: Detail Refinement of UniLDiff, after the degradation-modeling stage. The paper states that Stage 2 fine-tunes the VAE decoder together with DAEM using the composite objective

$x$2

The typesetting is described as malformed in the source, but the paper immediately identifies the intended terms: $x$3 is pixel-wise reconstruction loss, $x$4 is structural similarity loss, and $x$5 is an auxiliary load-balancing loss that “prevents expert collapse” by encouraging more uniform expert utilization across batches (Cheng et al., 31 Jul 2025). The exact closed form of $x$6 is not present in the supplied material.

Several implementation facts delimit the module’s actual scope. The full framework uses Stable Diffusion XL (SDXL) as the latent diffusion base, leveraging pre-trained encoder, decoder, and UNet components (Cheng et al., 31 Jul 2025). DAEM itself is attached to the VAE decoder. The paper specifies top-$x$7 sparse routing with $x$8 in experiments, lightweight NAFBlock experts with varied receptive fields, and a shared branch based on transposed self-attention (Cheng et al., 31 Jul 2025). It does not specify the number of experts, exact branch widths, or DAEM-specific FLOPs.

These details matter because they distinguish DAEM from several neighboring expert paradigms. DMEP, for example, also involves experts and routing, but it operates at the level of LoRA-MoE fine-tuning and adapts expert count per Transformer submodule through dynamic pruning, not through decoder-side detail refinement (Li et al., 29 Apr 2026). DAMEX uses dataset-aware supervised routing in detection, again with explicit experts but for dataset specialization rather than local texture recovery (Jain et al., 2023). UniLDiff’s DAEM is therefore best understood as a restoration-specific, decoder-stage, sparse routing module for fine-structure reconstruction.

5. Empirical role and observed effects

The clearest quantitative evidence for DAEM comes from UniLDiff’s component ablation. With DAFF + Task Prompt but without DAEM, the reported performance is 27.36 PSNR and 62.77 MUSIQ. Adding DAEM yields 30.27 PSNR and 63.06 MUSIQ, corresponding to +2.91 dB PSNR and +0.29 MUSIQ (Cheng et al., 31 Jul 2025). The paper explicitly interprets this as a significant boost, and the qualitative analysis states that DAEM “visibly reduces residual artifacts and sharpens fine structures.”

The textual evidence also identifies the types of detail improved by DAEM. The module is said to target “textures,” “edges,” “small structures,” “text and facial contours,” while reducing “residual artifacts” and sharpening “fine structures” (Cheng et al., 31 Jul 2025). These claims are consistent with the module’s placement in the decoder, where latent features are mapped back toward image space and where high-resolution reconstruction errors are most exposed.

Broader benchmark results in UniLDiff are reported for the full model rather than DAEM alone, but the paper repeatedly attributes perceptual quality gains to the combination of DAFF and DAEM. In the three-task setting, UniLDiff reports best LPIPS 0.0651, DISTS 0.0639, CLIPIQA 0.6653, MUSIQ 68.89, and MANIQA 0.7038 (Cheng et al., 31 Jul 2025). In the five-task setting it is best on several MUSIQ/MANIQA metrics, including deraining, denoising, deblurring, and low-light restoration. These are not isolated DAEM numbers, but they support the claim that decoder-side detail refinement is important to the final perceptual profile.

A broader pattern emerges when comparing DAEM-like modules across tasks. DR.Experts shows that adding detail-/distortion-aware refinement and dynamic weighting improves BIQA substantially over unified image encoding, especially on smaller datasets where structured prior modeling matters more (Fu et al., 10 Feb 2026). DeRainMamba shows that adding MDPConv alone to a baseline improves PSNR from 41.13 to 41.49, while combining it with the frequency-aware branch yields 41.71 (Zhu et al., 8 Oct 2025). PT-DETR reports that replacing a backbone block with its Partially-Aware Detail Focus (PADF) module improves VisDrone performance while reducing parameters relative to baseline (Huo et al., 30 Oct 2025). Taken together, these results suggest a stable empirical theme: expert-like or selectively modulated detail modules tend to matter most where subtle local evidence is fragile, heterogeneous, or easily suppressed by shared global processing.

6. Conceptual neighbors, extensions, and limitations of the DAEM idea

The exact name Detail-Aware Expert Module appears in UniLDiff, but several later or parallel works can be organized around closely related design logics. The following table summarizes the most direct relationships already described in the source material.

Work	Relation to DAEM	Key mechanism
UniLDiff (Cheng et al., 31 Jul 2025)	Exact named DAEM	Decoder-side sparse MoE with varied receptive fields and global modulation
DR.Experts (Fu et al., 10 Feb 2026)	Strong conceptual neighbor	Differential distortion refinement plus dynamic distortion weighting
TDATR (Qin et al., 24 Mar 2026)	DAEM-like “perceive-then-fuse” design	Detail-aware multitask pretraining plus structure-guided cell localization
DeRainMamba (Zhu et al., 8 Oct 2025)	Spatial detail expert analogue	Multi-directional detail enhancement with directional differential branches
PT-DETR (Huo et al., 30 Oct 2025)	Lightweight DAEM-like front-end	Partial convolution plus dual partial attention for small-object detail
Remote sensing LVLM (Park et al., 27 Jun 2025)	Semantic-granularity-aware expert variant	Level-specific semantic experts for coarse-to-fine semantics

Several distinctions are important. DR.Experts should not be treated as an exact DAEM because its closest analogue is the combination of DSDM + DDWM, with DDWM acting as the expert-weighting mechanism and DSDM supplying the detail-/distortion-aware refinement (Fu et al., 10 Feb 2026). TDATR is also not a classical expert-routing module; its DAEM-like quality lies in the “perceive-then-fuse” decomposition, where fine-grained table details are learned through auxiliary tasks before being fused into final HTML parsing (Qin et al., 24 Mar 2026). The remote sensing LVLM uses explicit experts, but they are organized by semantic granularity rather than by restoration detail (Park et al., 27 Jun 2025).

These comparisons clarify what DAEM is and is not. A DAEM need not be a textbook sparse MoE with independent full subnetworks and token dispatch. It may instead appear as a feature-level expert weighting mechanism, a dynamic modulation block, or a granularity-specific expert stack, provided it satisfies the broader functional criterion of preserving or emphasizing subtle, specialized information that a unified path under-models. This suggests that DAEM is best regarded as an architectural pattern centered on three recurring operations: detail-aware representation construction, specialized processing, and adaptive integration.

At the same time, the supplied literature identifies several limitations. UniLDiff does not specify the number of experts or many branch-level implementation details for DAEM (Cheng et al., 31 Jul 2025). DR.Experts shows that expert-style weighting alone is insufficient unless the detail-sensitive signal is first purified (Fu et al., 10 Feb 2026). PT-DETR’s PADF is DAEM-like “in spirit” but lacks explicit expert routing (Huo et al., 30 Oct 2025). MADE-IT, though outside restoration, warns that expert systems can drift into redundancy unless expert evolution is guided by principled similarity criteria rather than indiscriminate expansion (Qiu et al., 24 Apr 2026). A plausible implication is that future DAEM designs may need not only better detail-sensitive routing but also better mechanisms for expert compactness, diversity, and interpretability.

In its most precise current sense, then, DAEM denotes UniLDiff’s decoder refinement module for adaptive, expert-based high-frequency recovery (Cheng et al., 31 Jul 2025). In the wider expert-modeling literature, it also names a broader design problem: how to ensure that local structures, subtle artifacts, semantic granularity, or fine visual evidence are not erased by uniform shared processing, but instead receive specialized computation whose influence is conditioned on the input.