Skip to content

SkelEM: Training-Signal Decoupling of Skeleton and Diffusion for Self-supervised Axial Super-Resolution in Volume Microscopy

Conference: ECCV 2026
arXiv: 2606.30012
Code: None
Area: Medical Imaging
Keywords: Axial Super-Resolution, Volume Electron Microscopy, Training-Signal Decoupling, Truncated Diffusion, Self-Supervised Learning

TL;DR

SkelEM proposes a self-supervised axial super-resolution framework that completely decouples the training signals (no shared gradients) between a topological skeleton network and a diffusion refiner. By utilizing a frozen, deterministic skeleton to anchor the structural topology and using truncated diffusion to recover high-frequency biological textures in 5 steps or fewer, this method achieves the best balance between fidelity, perceptual quality, and inference speed among self-supervised methods, while reaching SOTA performance on downstream membrane segmentation tasks.

Background & Motivation

Background: Volume Electron Microscopy (VEM) and Volume Light Microscopy (VLM) enable 3D imaging of organelles and tissue architectures. However, limited by mechanical sectioning or optical physics, the axial (Z-direction) resolution is substantially lower than the lateral (XY) resolution—for example, SBF-SEM has a lateral resolution of 4-10 nm but an axial resolution of only 25-40 nm. Since isotropic ground truth is extremely difficult to acquire, the field has shifted from supervised methods to self-supervised axial super-resolution (ASR). Three main paradigms exist: 2D super-resolution based on orthogonal planes, video frame interpolation, and 3D refinement using 2D diffusion models trained on the XY plane.

Limitations of Prior Work: Existing self-supervised ASR methods face a fundamental trilemma. Regression-based 3D convolutional networks (such as D2R-AENet) are constrained by pixel-level L1/L2 losses and inevitably regress to the mean, producing over-smoothed volumes. Pure diffusion models can synthesize high-frequency details but at the cost of structural hallucinations and extremely high inference latency (hundreds to thousands of denoising steps per single patch). Video frame interpolation methods rely on natural video pre-training, which struggles to adapt to the non-linear deformation of biological ultraststructures, thereby causing axial discontinuities.

Key Challenge: The authors attribute both failure modes—over-smoothing and hallucination—to a single root cause: the lack of a faithful, domain-adaptive structural skeleton to capture highly non-linear deformations of biological ultrastructures. Without such a skeleton, regression-based models can only smooth out high-frequency textures to minimize loss, while generative models fabricate structures in the absence of topological constraints. At the same time, truncated diffusion can theoretically accelerate inference by starting from an informative prior, but the quality of this prior directly determines the effectiveness of truncation—without a reliable structural anchor, truncation merely shifts the hallucination problem earlier rather than resolving it.

Goal: To design a self-supervised ASR framework that simultaneously satisfies three objectives: (1) preserving 3D structural continuity (avoiding axial artifacts), (2) recovering realistic high-frequency biological textures (avoiding over-smoothing), and (3) maintaining an inference speed fast enough for practical deployment on biological volumes on the scale of hundreds of gigabytes.

Key Insight: The authors observe that if the two sub-tasks of constructing low-frequency topology and generating high-frequency texture can be completely decoupled at the level of training signals—preventing gradients from the texture-generation stage from backpropagating to the topological network—each sub-network can focus on the problem it solves best. The topological network only needs to ensure spatial continuity, the diffusion refiner only needs to complete the texture, and a lightweight residual estimator builds a physically plausible bridge between them.

Core Idea: Replace end-to-end or cascaded training with training-signal decoupling: a frozen topological skeleton network + a diffusion refiner adapted through sparse real slices via bidirectional self-alignment + a truncated sampling scheme based on cycle-consistent residuals, enabling high-fidelity axial super-resolution reconstruction in no more than 5 diffusion steps.

Method

Overall Architecture

SkelEM is a two-stage, three-round training self-supervised ASR framework. The input is an anisotropic volume V_low (comprising d sparse axial slices), and the goal is to reconstruct an isotropic high-resolution volume V_high (comprising d_hat = (d-1)*r+1 slices, where r is the upsampling rate, r=8 by default). For any adjacent pair of real slices I_0 and I_1, the missing intermediate slice Î_τ at relative position τ needs to be synthesized.

The core design of the framework is to decouple the ASR task at the level of training signals into two non-overlapping optimization objectives: the topological skeleton network f_flow is solely responsible for warping a deterministic low-frequency skeleton from boundary slices via optical flow and is frozen after training (gradients never enter subsequent stages); the diffusion refiner f_θ is responsible only for reclaiming high-frequency biological textures on top of the skeleton, optimized independently through two stages of synthetic manifold pre-training and real sparse-slice self-alignment; a lightweight residual estimator f_Δ serves as a bridge, extracting high-frequency residual prior via cycle-consistent reconstruction on real slices, which allows truncated diffusion to finish refinement in only S steps (S <= 5).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: I0, I1, τ"] --> B["Topological Skeleton Generation<br/>f_flow bidirectional optical flow warp + fusion"]
    B --> C["Residual Estimation<br/>f_Δ predicts high-frequency residual Δτ"]
    C --> D["Prior Truncated Initialization<br/>Skeleton + Residual -> Forward noise addition to t_S"]
    D --> E["Truncated Diffusion Denoising Step<br/>f_θ backward denoising S≤5 steps"]
    E --> F["Output: Îτ"]

Key Designs

1. Topological Skeleton and Synthetic Manifold Pre-training: Training the optical flow network with in-domain pseudo-HR volumes and discarding the detail refiner to enforce topology-texture separation

To address the non-linear deformation of biological ultrastructures across large axial gaps, which optical flow models pre-trained on natural videos (such as RIFE and InterpolAI) fail to capture, the authors use the pseudo-high-resolution volume V_hat_high generated by D2R-AENet as training data to train the intermediate flow network (IFNet) of RIFE as the topological network f_flow. Given adjacent real slices I_0, I_1, and the relative position τ, f_flow predicts bidirectional optical flow F_{τ→0}, F_{τ→1}, and a fusion mask M_τ, synthesizing the structural skeleton via backward warping:

\[ \hat{I}^{(skel)}_{\tau} = M_{\tau} \odot \mathcal{W}(I_0, F_{\tau \rightarrow 0}) + (1 - M_{\tau}) \odot \mathcal{W}(I_1, F_{\tau \rightarrow 1}) \]

The key design choice is to intentionally discard the detail refiner of RIFE and freeze f_flow after training. This is because V_hat_high inherently contains cross-planar artifacts; if the detail refiner were maintained, these synthetic artifacts would leak into the skeleton, violating the principle of topology-texture separation. Discarding the refiner forces f_flow to output only deterministic low-frequency topological anchors, providing a clean structural foundation free of synthetic bias for subsequent diffusion refinement.

2. Diffusion Refiner and Sparse Self-Alignment: Establishing a generative prior with synthetic manifold pre-training, followed by debiasing on real slices via a bidirectional cycle

The skeleton I_hat_τ^(skel) lacks the high-frequency biological textures lost during physical sectioning. To address this, a diffusion refiner f_θ is introduced, conditioned on the boundary slices I_0, I_1, and position τ, and pre-trained on V_hat_high via a standard diffusion objective (L_base, the standard noise-prediction MSE loss). However, a refiner trained solely on synthetic data inherits the cross-planar artifacts of V_hat_high.

To solve this problem, the authors propose bidirectional sparse self-alignment: fine-tuning f_θ directly on the real sparse slices of V_low. Given three consecutive real slices {I_{-1}, I_0, I_1}, the current refiner is used to synthesize two intermediate anchors Î_{τ-1} (between I_{-1} and I_0) and Î_τ (between I_0 and I_1). Then, through a bidirectional condition combination Φ = {(Î_{τ-1}, Î_τ, 1-τ), (Î_τ, Î_{τ-1}, τ)}, f_θ is required to consistently reconstruct the central observed slice I_0 from both directions:

\[ \mathcal{L}_{adapt} = \sum_{\phi \in \Phi} \mathbb{E}_{\epsilon, I_0, t} \left[ \| \epsilon - f_{\theta}(x_t, t, \phi) \|_2^2 \right] \]

The elegance of this design lies in the fact that gradients of the synthetic anchors Î_{τ-1} and Î_τ are detached; they do not serve as optimization variables and are used only as fixed conditional inputs. This forces f_θ to learn debiasing from real data rather than simply replicating the features of synthetic anchors. Intuitively, a direction-agnostic refiner should consistently reconstruct I_0 regardless of the direction from which it is synthesized—this symmetry constraint both eliminates synthetic bias and enhances topological consistency.

3. Truncated Prior Sampling: Injecting real-domain high-frequency cues into the diffusion initialization through cycle-consistent residual extraction, enabling convergence in only S refinement steps

Standard diffusion requires starting from pure Gaussian noise and undergoing hundreds of denoising steps, which is highly impractical for biological volumes on the scale of hundreds of gigabytes. SkelEM's solution is to perform cycle-consistent reconstruction on real slice sequences using the frozen f_flow to extract physically plausible high-frequency residual priors, which are injected directly into the initial state of diffusion to drastically truncate the denoising trajectory.

Specifically, on {I_{-1}, I_0, I_1}, two virtual anchors I_{-τ} and I_{1-τ} are synthesized at symmetric topological locations using f_flow, and then I_0 is reconstructed through cyclic warp—the difference between the reconstructed Î_0 and the real I_0, i.e., Δ_gt = I_0 - Î_0, precisely isolates the high-frequency details lost during the non-linear warp. A lightweight residual estimator f_Δ (ResNet-9) learns to predict this residual from the skeleton using a frequency-domain L1 loss:

\[ \mathcal{L}_{freq} = \| \mathcal{F}_{fft}(f_{\Delta}(\hat{I}^{(skel)}_{\tau})) - \mathcal{F}_{fft}(\Delta_{gt}) \|_1 \]

During inference, for a target slice, the deterministic skeleton I_hat_τ^(skel) and predicted residual Δ_τ are first obtained and combined into a pixel-space prior I_hat_τ^(prior) = I_hat_τ^(skel) + Δ_τ. Then, the diffusion trajectory is truncated to an intermediate time step t_S (S << T), initialized via the forward noise addition formula:

\[ x_{t_S} = \sqrt{\bar{\alpha}_{t_S}} (\hat{I}^{(skel)}_{\tau} + \Delta_{\tau}) + \sqrt{1 - \bar{\alpha}_{t_S}} \epsilon \]

A DDPM backward sampling of only S steps (the paper defaults to S=3) is subsequently executed on f_θ. The value of this step is that the skeleton provides a deterministic structural anchor, and the residual injects high-frequency cues extracted from the real domain, allowing diffusion to generate realistic textures through only localized stochastic refinement rather than "hallucinating" the entire slice from scratch.

A Complete Example

Take inference with r=8 on the EPFL dataset as an example. The input consists of three consecutive slices {I_{-1}, I_0, I_1} in V_low, and the goal is to synthesize the intermediate slice at τ=0.5 between I_0 and I_1.

  1. Skeleton Generation: f_flow takes (I_0, I_1, τ=0.5) as input to predict the bidirectional optical flow and fusion mask, warping out the skeleton I_hat_0.5^(skel). At this stage, the slice possesses correct low-frequency topology (membrane structural positions, organelle contours) but has smooth textures and lacks the granularity of real electron microscopy images.
  2. Residual Estimation: f_Δ takes I_hat_0.5^(skel) as input and predicts the pixel-space residual Δ_0.5, supplementing the high-frequency details (such as sharp textures at membrane boundaries) lost during the warp process.
  3. Prior Initialization: The prior is merged as I_hat_0.5^(prior) = skeleton + residual (with λ=1.0). From a 1000-step DDPM schedule, t_S=300 is selected (corresponding to S=3 backward steps), and forward noise is added according to Eq. (6) to obtain the initial latent variable x_300. At this juncture, x_300 contains approximately 70% of the prior signal and 30% noise.
  4. Three-step Refinement: Conditioned on (I_0, I_1, τ=0.5), f_θ executes 3 steps of DDPM backward denoising from x_300 → x_200 → x_100 → x_0. Each step predicts a clean state x̂_0 and samples the next latent variable according to standard DDPM posterior sampling. The final output is x_0 = Î_0.5, taking approximately 151 ms (on a single RTX 4090 for a 256x256 patch).
  5. Results: Compared to the pure skeleton (high PSNR but blurry), the 3-step refinement reduces the LPIPS in the XZ/YZ views from 0.52 to 0.32, yielding sharp membrane boundary textures. Compared to pure diffusion generation from scratch (which requires hundreds of steps and has unreliable structures), the skeleton anchoring ensures the topological integrity of the membrane structures, preventing breakages or false connections.

Loss & Training

The three training stages are optimized independently, with no shared gradients:

  • Stage 1 (Topological Skeleton): Train f_flow on V_hat_high using L1 loss for 300 epochs, discard the detail refiner, and freeze the network. This takes about 12 GPU hours.
  • Stage 2 (Residual Estimator): Train f_Δ on V_low real slices using the frequency-domain L1 loss L_freq for 5 epochs. Crucially, as Δ_gt is derived from cycle-consistent reconstruction, it ensures that the residuals are physically high-frequency components of the real domain rather than synthetic artifacts.
  • Stage 3 (Diffusion Refiner): Two-phase training. (a) Pre-train on V_hat_high using standard diffusion noise prediction MSE (L_base) for 880 epochs to establish a generative prior; (b) Fine-tune on real slices of V_low using the bidirectional self-alignment loss L_adapt for 20 epochs to eliminate synthetic bias. The gradients of the synthetic anchors Î_{τ-1} and Î_τ are detached during the optimization of L_adapt to prevent erroneous anchor signals from misleading the refiner.

All networks are trained on 256x256 patches using the Adam optimizer (β1=0.9, β2=0.99), taking approximately 72 hours on two RTX 4090 GPUs to complete the whole pipeline. During inference, the default configuration of λ=1.0 and S=3 is adopted, requiring 151 ms per single patch.

Key Experimental Results

Main Results

The main results of r=8 axial super-resolution on the FIB-25 and EPFL datasets are compared below. SkelEM achieves the best perceptual quality (obtaining the optimal LPIPS in both XZ and YZ views) and highly competitive fidelity among self-supervised methods. It leads the runner-up D2R-AENet (s̄=0.76) by 0.16 with an overall ranking score of s̄=0.92.

Dataset Method 3D-PSNR ↑ SSIM XY ↑ LPIPS XZ ↓ LPIPS YZ ↓ Type
FIB-25 Bicubic 25.33 0.6528 0.5881 0.5971 Self-supervised
FIB-25 D2R-AENet 27.62 0.7373 0.4136 0.4246 Self-supervised
FIB-25 Lee et al. 26.18 0.6607 0.3862 0.3760 Self-supervised
FIB-25 vEMDiffuse-a 23.51 0.5699 0.4260 0.4466 Self-supervised
FIB-25 InterpolAI 24.27 0.6140 0.5535 0.5819 Self-supervised
FIB-25 SkelEM 26.24 0.6765 0.3658 0.3687 Self-supervised
EPFL Bicubic 23.09 0.4938 0.6794 0.6674 Self-supervised
EPFL D2R-AENet 26.35 0.6387 0.4788 0.4795 Self-supervised
EPFL Lee et al. 24.99 0.5593 0.3327 0.3431 Self-supervised
EPFL vEMDiffuse-a 22.98 0.4879 0.4269 0.4060 Self-supervised
EPFL InterpolAI 23.67 0.5116 0.5727 0.5861 Self-supervised
EPFL SkelEM 25.56 0.6192 0.3213 0.3298 Self-supervised

The BRAVE-ASR zero-shot cross-instrument transfer experiment (trained on EPFL, directly tested on BRAVE-ASR) further validates generalization capability: SkelEM again achieves the best LPIPS in the XZ/YZ views (0.3203 / 0.3093) and is the only method that does not suffer from catastrophic failures across all 25 evaluation metrics.

In downstream membrane segmentation tasks, SkelEM achieves state-of-the-art (SOTA) performance across all four metrics: F1 Score (0.8643), IoU (0.7611), ARE (0.2129), and VoI (0.4811), even surpassing two supervised baselines (SRUNet and vEMDiffuse-i). Notably, although SkelEM is not ranked first in PSNR/SSIM (where D2R-AENet shows the highest fidelity), it comprehensively leads in segmentation accuracy. This demonstrates that pixel-level fidelity metrics are insufficient to evaluate the quality of downstream structural analysis.

Ablation Study

Ablations are performed along three axes on the EPFL dataset (see Table): the necessity of the two-stage design, the quality of the skeleton prior, and the effect of sparse data self-alignment.

Ablation Dimension Configuration 3D-PSNR ↑ LPIPS XZ ↓ Key Findings
Skeleton only Skeleton only 25.96 0.5171 Highest fidelity but severely over-smoothed
Refiner only Refiner only (No adaptation) 22.43 0.4390 Fidelity collapses due to lack of structural guidance
Refiner only Refiner only (With adaptation) 20.26 0.5952 Adaptation cannot replace structural anchors; erroneous anchors cause degradation
Skeleton source Bilinear interpolation 22.55 0.4071 No structural prior, comprehensive degradation
Skeleton source RIFE pre-trained on natural videos 25.19 0.3542 Natural video prior is partially effective but inferior to in-domain retraining
Skeleton source InterpolAI pre-trained 24.78 0.3406 Ditto
Self-alignment method No adaptation 25.94 0.3976 High fidelity but poor perceptual quality (replicating smooth textures)
Self-alignment method Sparse3Diff adaptation 25.59 0.3640 Recovers some perceptual quality but inferior to the proposed L_adapt
Self-alignment method SkelEM L_adapt 25.56 0.3213 Minimum fidelity cost for maximum perceptual gains

A joint ablation of the residual injection strength λ and the refinement step count S (Tab. 5) reveals an interesting non-monotonic trend: when S=1, λ=1.0 yields the best performance (residuals provide key structural information), but regardless of the value of λ, 1 step is insufficient to eliminate the regression bias, leading to poor perceptual quality; when S=3, the sensitivity to λ is greatly reduced, and λ=1.0 achieves the optimal balance between fidelity and perception; when S=5, fidelity slightly degrades, indicating that excessive diffusion dilutes the structural prior. The default configuration of λ=1.0 and S=3 yields a single-patch inference time of 151 ms, which is more than an order of magnitude faster than other diffusion-based methods (vEMDiffuse-a ~2071 ms, TPDM ~402,788 ms).

Key Findings

  • The skeleton is an enabling condition, not just icing on the cake: The "Refiner only" configurations collapse both with and without adaptation, proving that without a reliable structural anchor, diffusion models cannot learn topologically consistent reconstruction in a self-supervised manner—this is consistent with the theory in truncated diffusion literature that "a good prior is essential for effectiveness."
  • In-domain skeleton retraining is necessary: RIFE and InterpolAI pre-trained on natural videos perform significantly worse than the skeleton retrained on pseudo-HR volumes, as the non-linear deformations of biological ultrastructures (membrane fusion, disappearance, bifurcation) differ fundamentally from the rigid motions in natural videos.
  • Bidirectional self-alignment outperforms unidirectional: Compared to the unidirectional constraint of Sparse3Diff, the bidirectional cycle self-alignment of SkelEM achieves significant improvements in LPIPS across all views, proving that direction symmetry is an effective inductive bias.
  • S=3 is the sweet spot: 1 step is insufficient for de-biasing, 5 steps over-diffuses and dilutes the prior, while 3 steps precisely strike a balance between de-biasing and preserving the structure.

Highlights & Insights

  • Training-signal decoupling is a transferable design paradigm: Allocating low-frequency topology and high-frequency textures to two "unaware" networks (with no shared gradients and a frozen skeleton) allows the regression network to focus on spatial continuity and the diffusion model to focus on texture generation, avoiding objective conflicts inherent in end-to-end training. This approach is applicable not only to ASR but also to any image restoration task requiring simultaneous structural consistency and texture realism (e.g., video super-resolution, medical image reconstruction).
  • Discarding the detail refiner is counter-intuitive yet critical: In typical VFI pipelines, the detail refiner is a core component for improving PSNR. However, SkelEM intentionally discards it because the synthetic bias of the pseudo-HR volume leaks into the skeleton through the refiner. This design decision of "discarding a useful module for the sake of separation" reflects a deep understanding of the problem's essence.
  • Cycle-consistent reconstruction serves two purposes simultaneously: The same cycle-consistent mechanism (warping on {I_{-1}, I_0, I_1} to reconstruct I_0) provides both the real-domain high-frequency ground truth (Δ_gt) for the residual estimator and the symmetrically constrained conditional pairs for the diffusion refiner's bidirectional self-alignment—killing two birds with one stone is an extraordinarily elegant design.
  • Targeted choice of frequency-domain loss: The residual estimator uses an L1 loss in the FFT domain rather than the pixel-space loss, directly supervising the quality of high-frequency reconstruction. This fits perfectly with the division of labor with the diffusion refiner (where low-frequencies are guaranteed by the skeleton and high-frequencies are supplemented by the residuals).

Limitations & Future Work

  • Blurring of nanoscale double-membrane structures: The authors admit in the appendix that SkelEM tends to blur the nanoscale gaps between double-membrane structures under large upsampling factors such as r=8. The root cause is that when trained on biological data, the skeleton network learns a prior that "tissues may merge or disappear," leading it to predict probabilistic connections when facing high uncertainty in large gaps. Conversely, InterpolAI retains clearer geometric separation due to its rigid motion prior from natural videos. A concrete future direction is to initialize the skeleton network with a natural video VFI model and then fine-tune it in-domain, thereby inheriting its strong geometric boundary-preserving prior while adapting to biological non-linear deformations.
  • Dependence on the quality of pseudo-HR volumes: The pre-training of both the skeleton network and the diffusion refiner depends on the V_hat_high generated by D2R-AENet. If the pseudo-HR volume itself contains systematic bias (such as cross-planar artifacts), the skeleton quality is also compromised. Though discarding the detail refiner and the self-alignment mechanism partially mitigates this, completely eliminating this dependency may require training on more physically accurate simulated data or synthetic data.
  • Insufficient quantitative evaluation on VLM: Although qualitative results on zebrafish retina and quantitative results for r=8 on CSBDeep mouse liver are presented, the modal diversity of VLM (confocal, two-photon, light-sheet, etc.) far exceeds that of VEM. SkelEM's performance across more VLM modalities remains to be validated.
  • Inability to process in-slice anisotropy: The current method only addresses interpolation between axial slices, assuming that the lateral resolution is already high enough. For scenarios where lateral resolution is also insufficient (e.g., low-magnification light microscopy), joint super-resolution across both the XY and Z directions needs to be considered.
  • vs D2R-AENet: As self-supervised ASR methods, D2R-AENet is the strongest in fidelity metrics due to pixel-level constraints, but it is essentially a regression network, and its outputs are inevitably over-smoothed. SkelEM retains the pseudo-HR volumes generated by D2R-AENet as training data but separates "structure preservation" and "texture generation" into two networks, breaking the deterministic smoothness of regression via stochastic diffusion sampling. The relationship between the two is complementary rather than adversarial—D2R-AENet proves the utility of pseudo-HR volumes, while SkelEM addresses the deficiency in perceptual quality on top of it.
  • vs Diffusion-based ASR Methods (Lee et al. / vEMDiffuse / DiffuseIR / TPDM): These methods either rely entirely on diffusion starting from scratch (leading to structural hallucinations and several-second delays) or couple the supervision signals of regression and diffusion within a cascaded framework (leading to synthetic bias leakage). The core distinction of SkelEM lies in the thorough decoupling of training signals—where the skeleton never receives gradients from the diffusion phase—and the resulting truncated diffusion (≤5 steps vs. hundreds to thousands of steps). This speedup in inference (151 ms vs. 2,000–400,000 ms) transitions the method from an academic demo to a practically viable tool.
  • vs InterpolAI: As a video frame interpolation-based ASR method, InterpolAI demonstrates the potential of natural video pre-training for ASR. However, its underlying assumptions (scene continuity, object rigidity) do not hold for biological ultrastructures. SkelEM borrows the skeleton idea based on optical flow warp but achieves superior biological domain adaptation and topology-texture separation through in-domain retraining and explicitly discarding the detail refiner.
  • Inspiring Application of Truncated Diffusion (SDEdit / Truncated Diffusion): The original concept of truncated diffusion is to accelerate generation by initializing the denoising process with a partially noised guidance image, the efficacy of which heavily depends on the quality of initialization. SkelEM's contribution lies in providing a concrete scheme for obtaining high-quality initialization—extracting high-frequency cues from real data using cycle-consistent residuals instead of relying on synthetic data or external pre-trained models. This provides a reusable pipeline for implementing truncated diffusion in scientific imaging fields.

Rating

  • Novelty: ⭐⭐⭐⭐ The concept of training-signal decoupling is novel in the ASR domain. Discarding the detail refiner, bidirectional self-alignment, and cycle-consistent residual extraction collectively formulate an elegant overall system, although individual components (RIFE, DDPM, truncated diffusion) do not represent disruptive innovations on their own.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Three public datasets + self-constructed BRAVE-ASR, zero-shot cross-instrument transfer, downstream membrane segmentation, VLM cross-modality validation, computational efficiency comparison, joint ablation of residual injection intensity and step count, and a 25-metric normalized ranking heatmap—the experimental design meticulously addresses almost every possible angle of skepticism.
  • Writing Quality: ⭐⭐⭐⭐ The motivation chain is clear (trilemma → root cause → decoupling solution), and the methods section tightly couples formulas with explanations. The framework overview in Fig. 2 and the ranking heatmap in Fig. 9 are highly intuitive and powerful. However, the density of equations in the method section is relatively high, lacking some intuitive explanations.
  • Value: ⭐⭐⭐⭐ Advances self-supervised ASR from a "trilemma loss" to "trilemma balance," compressing the inference speed from seconds to milliseconds, which renders the method practically useful. The BRAVE-ASR dataset establishes a standardized benchmark for cross-instrument generalization, which is beneficial for the field. However, the method heavily relies on the quality of pseudo-HR volumes, imposing an implicit requirement on the training data.