acceptodds
Under review as a conference paper at ICLR 2027

Trajectory Constraints for Imaging Inverse Problems

Abstract

Diffusion models have achieved strong performance in imaging inverse problems, but typically rely on domain-matched pretrained models and external training data. Moreover, errors introduced at intermediate reverse-sampling steps can accumulate and degrade the final reconstruction, with their propagation strongly influenced by how successive states are related. We propose TRACE, a training-free TRAjectory-Constrained rEconstruction framework for imaging inverse problems. TRACE replaces the pretrained diffusion model with an untrained overparameterized network optimized directly from the observed measurements and organizes reconstruction as a sequence of intermediate states. At each step, TRACE enforces stochastic proximal consistency (SPC) by inferring each reconstruction from a stochastically perturbed predecessor under measurement consistency, while proximally constraining the resulting estimate to its corresponding unperturbed state. This formulation admits a proximal interpretation, from which we derive stability bounds characterizing the variation between consecutive states under approximate network updates. A local Jacobian analysis further characterizes how SPC regularizes the network-parameterized reconstruction. Experiments on natural-image restoration and CT reconstruction demonstrate competitive performance against both untrained and pretrained neural-prior methods. We further show that the proposed trajectory constraint can be incorporated into existing diffusion-based inverse solvers to improve reconstruction performance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-supervised, diffusion model, training-free, inverse problems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_samjOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Tahis paper proposes TRACE, training-free framework for imaging inverse problems that combines deep image priors with a diffusion-inspired trajectory. The method optimizes a network for each measurement instance, using scheduled input perturbations and Stochastic Proximal Consistency (SPC) to couple successive reconstruction states. Experiments cover CT reconstruction, inpainting, super-resolution, and deblurring, while additional results examine SPC within existing diffusion-based solvers. The paper presents a clear motivation, achieves competitive results across multiple tasks, and includes extensive component ablations and hyperparameter analyses. However, the specific benefit of the diffusion-inspired schedule over general perturbation or annealing strategies remains insufficiently established. The distinction between SPC and existing proximal constraints, particularly those in DAPS, also requires clarification. Furthermore, TRACE’s practical advantages over available pretrained diffusion models are not fully demonstrated, especially given its higher reconstruction cost. Overall, the approach is promising, but its novelty and most suitable application settings need to be articulated more clearly.

Critical strengths

S1. The paper is well-structured and clearly presents the research motivation and method design.

S2. Combining DIP with diffusion-style trajectories has certain value, and the proposed method achieves good reconstruction performance on most evaluated tasks.

S3. The authors provide extensive ablation studies and hyperparameter analyses, which help clarify the roles of different components and the impact of parameter settings on performance.

Critical weaknesses

W1. Limited novelty.

The proposed method mainly combines per-instance network optimization, noise scheduling, and proximal constraints between consecutive states. Compared with existing methods such as SDI and aSeqDIP, the paper does not sufficiently clarify what substantial new contribution arises from this combination.

W2. Insufficient validation of the claimed benefits of diffusion-style scheduling.

The authors claim to incorporate the advantages of diffusion scheduling into DIP. However, the existing ablation studies mainly examine the effects of noise perturbation and proximal constraints, which are insufficient to establish the specific value of the proposed scheduling strategy. In particular, the method still optimizes a clean image estimate, and its connection to diffusion trajectories requires further explanation. Comparisons with fixed noise levels, alternative annealing schedules, and designs inspired by flow matching or diffusion bridges would help determine whether the improvements arise from the specific schedule or from generic perturbation and progressive optimization.

W3. Incorrect citation information.

SDI (Self-Diffusion for Solving Inverse Problems) is listed as a NeurIPS 2026 paper, whereas its official conference version was published at NeurIPS 2025. The authors should verify and correct the reference information to ensure consistency.

W4. Insufficient explanation of the definition, implementation, and distinctive contribution of SPC.

The method section is overly brief, making it difficult to understand how Stochastic Proximal Consistency (SPC) differs from existing techniques. Frameworks such as Half-Quadratic Splitting (HQS) already incorporate measurement consistency and proximal constraints, while the paper does not adequately explain how stochastic perturbations make the proposed mechanism fundamentally different.

Regarding Table 3, the authors should also provide a clearer explanation of how SPC is incorporated into each diffusion solver. In particular, DAPS already employs Gaussian-form quadratic constraints and noise-scale scheduling. Therefore, the distinction between the additional SPC term and its β\beta schedule and the corresponding mechanisms already present in DAPS remains unclear.

W5. Insufficient justification of the practical scenarios in which avoiding pretraining is advantageous.

For the natural-image and CT tasks considered in this paper, suitable datasets and pretrained diffusion models are already available. Therefore, reliance on training data does not necessarily constitute a practical disadvantage. Meanwhile, Table 8 shows that several diffusion-based methods have substantially shorter reconstruction times than TRACE.

The value of TRACE may be more pronounced in scenarios where suitable training data or pretrained priors are difficult to obtain, but the current experiments do not sufficiently validate this advantage. Furthermore, [1] shows that even mismatched or weak diffusion priors can achieve strong reconstruction performance when the observations provide sufficient information. Therefore, the authors should more clearly delineate the practical advantages of TRACE over existing diffusion-prior-based approaches.

Reference

[1] Weak Diffusion Priors Can Still Achieve Strong Inverse-Problem Performance.

Questions for authors

Q1. Benefits of diffusion-style scheduling (W2).

Can the authors explain what specific advantages diffusion-style scheduling brings to TRACE? How do these advantages differ from those of general noise perturbation or annealing strategies?

Q2. Definition and novelty of SPC (W4).

Can the authors more clearly explain the meaning of "Stochastic" in Stochastic Proximal Consistency (SPC) and how it differs from existing proximal constraints? In particular, what exactly does SPC add to DAPS?

Q3. Practical advantages and applicability of TRACE (W5).

When pretrained diffusion models are already available and offer faster reconstruction, what are the practical advantages of TRACE? In which scenarios would TRACE be a more suitable choice?

9 comments
AuthorsOpenReview ↗

Response to W1: Methodological novelty.

TRACE reconstructs an image by sequentially optimizing a pretraining-free network on noise-perturbed inputs while coupling each output to the preceding unperturbed reconstruction.

(1) Distinct formulation. Unlike aSeqDIP, which updates its input deterministically and repeatedly solves the same objective form with a fixed λ\lambda, TRACE constructs a stochastic optimization problem at each outer iteration using annealed perturbations σt\sigma_t and time-varying proximal coupling βt\beta_t. Unlike SDI, which optimizes measurement fidelity on perturbed inputs, TRACE constrains the reconstruction from each perturbed input to remain close to its preceding unperturbed state, introducing a distinct trajectory regularization mechanism.

(2) Distinct regularization mechanism. Proposition 4.3 shows that SDI-type stochastic perturbations induce measurement-space sensitivity regularization proportional to σt2∥AJD∥F2\sigma_t^2\|AJ_D\|_F^2, whereas SPC additionally induces βtσt2∥JD∥F2\beta_t\sigma_t^2\|J_D\|_F^2. Crucially, this term also regularizes measurement-nullspace sensitivity, which measurement fidelity alone cannot directly constrain. This stochasticity-induced image-space Jacobian regularization is absent from both aSeqDIP's deterministic objective and SDI's formulation.

(3) Empirical and theoretical significance. TRACE achieves 29.55 dB in sparse-view CT, compared with 24.36 dB for SDI (+5.19 dB) and 27.15 dB for aSeqDIP (+2.40 dB), while outperforming both across all six natural-image settings. Our controlled ablations and null-space experiments further demonstrate that appropriate SPC strength improves peak PSNR, post-peak stability, and reconstruction of measurement-nullspace components while reducing null-space Jacobian sensitivity, compared with stochastic perturbations alone.

We further report additional PSNR curves of sparse-view CT reconstruction that compare TRACE and SDI over 300 outer stages, with 150 inner iterations per stage (Fig. R1: panel (a), panel (b), panel (c), panel (d), panel (e)). Across all five displayed examples, TRACE has a smooth PSNR curve during extended optimization, whereas SDI exhibits larger fluctuations and subsequent degradation. These results establish SPC as a distinct, significant, and previously overlooked design principle for imaging inverse problems.

We thank the reviewers for raising this question. We will clarify these methodological distinctions in the manuscript.

Response to W3 We thank the reviewer for pointing this out. We will modify it in the revised manuscript.

AuthorsOpenReview ↗

Response to W2 and Q1: Role of diffusion-inspired scheduling.

In TRACE, noise is added to the network input, uk=xk−1+σkϵku_k=x_{k-1}+\sigma_k\epsilon_k, while the output remains a clean-image estimate fitted to the measurements under the SPC constraint. We borrow the noise-level schedule from diffusion, but do not perform reverse-diffusion sampling. The schedule lets us reduce input perturbation as reconstruction proceeds. Proposition 4.3 explains its local effect through a sensitivity penalty whose strength depends on the noise amplitude.

Our default schedule uses noise level decreases from approximately 0.428 to 0.010. A fresh Gaussian noise sample is drawn at each stage.

According to your suggestion, we added comparisons on 14 images per task, keeping T=40T=40 and K=150K=150.

For the fixed-perturbation comparison (Table R5), we tested seven constant amplitudes from 0.01 to 0.80 and reused one noise sample across stages. Table R5 shows gains of 1.65 and 1.74 dB over the best tested fixed perturbations. This comparison includes both scheduling and noise resampling.

For the fixed noise level comparison (Table R5'), we compare TRACE with seven constant noise levels, drawing fresh noise at every outer step in all settings. Results in Table R5' show that decreasing the noise amplitude is beneficial for super-resolution under the tested settings, but is not necessary for strong inpainting performance.

For the schedule comparison (Table R6), we kept the endpoints and noise samples unchanged and tested two alternatives: geometric decay, σk+12=qσk2\sigma_{k+1}^{2}=q\sigma_k^{2} with q≈0.825q\approx0.825, and a flow matching (FM)-inspired shift, σkshift=σmin⁡+(σmax⁡−σmin⁡)3rk1+2rk\sigma_k^{\mathrm{shift}}=\sigma_{\min}+(\sigma_{\max}-\sigma_{\min})\frac{3r_k}{1+2r_k}, where rk=T−kT−1r_k=\frac{T-k}{T-1}. Table R6 isolates the schedule shape: the default schedule exceeds geometric decay by 0.29 and 0.79 dB, while the shifted schedule gives slightly higher PSNR than the default.

We also tested FM-inspired input interpolation (Table R7). Setting τk=σk1+σk\tau_k=\frac{\sigma_k}{1+\sigma_k} gives u~k=(1−τk)xk−1+τkϵk=uk1+σk\widetilde u_k=(1-\tau_k)x_{k-1}+\tau_k\epsilon_k=\frac{u_k}{1+\sigma_k}. Thus, at the same reconstruction state, this changes only the overall input scale. Table R7 shows little change under input interpolation, consistent with the approximate scale invariance introduced by BatchNorm.

The experiments show that the schedule affects performance, but the default form is not essential. We will clarify this in the manuscript.

AuthorsOpenReview ↗

Response to W4 & Q2: Contribution and implementation of SPC.

Meaning of SPC and distinction from existing proximal constraints. TRACE enforces SPC along the reconstruction trajectory through a proximal constraint that keeps the reconstruction generated from each stochastically perturbed intermediate trajectory estimate close to its preceding unperturbed state. Specifically, "stochastic" refers to the injection of random noise into the input, ut=xt+1+σtϵt\mathbf{u}_t=\mathbf{x}_{t+1}+\sigma_t \epsilon_t, where ϵt∼N(0,I)\epsilon_t\sim\mathcal{N}(0,I). The resulting reconstruction is proximally coupled to the preceding unperturbed state through βt2∥Dθt(ut)−xt+1∥22.\frac{\beta_t}{2}\Vert D_{\theta_t} (\mathbf{u}_t)-\mathbf{x}_{t+1}\Vert_2^2. Unlike conventional proximal regularization, SPC explicitly applies the proximal constraint to a reconstruction generated from a stochastically perturbed input.

Also, we would like to clarify that standard HQS is an optimization algorithm. HQS introduces an auxiliary variable and uses a quadratic penalty to enforce agreement between the split variables. SPC instead penalizes deviations of the network output under random input perturbations from the preceding reconstruction. Our contribution is this perturbation–consistency interaction and its analysis, rather than the quadratic penalty itself.

Explanation of Table 3. For all the diffusion solvers, we use the same implementation. For all the experiments, we only add the SPC term βt2∥x−xt+1∥22\frac{\beta_t}{2}\Vert \mathbf{x}-\mathbf{x}_{t+1}\Vert_2^2 between its current clean-image estimate xt\mathbf{x}_{t} and a stored reference xt+1\mathbf{x}_{t+1} to their objective function Φ(x)\Phi(\mathbf{x}). The augmented objective Φ(x)+βt2∥x−xt+1∥22\Phi(\mathbf{x})+\frac{\beta_t}{2}\Vert \mathbf{x}-\mathbf{x}_{t+1}\Vert_2^2 is then solved by their default updating optimizer.

DAPS’s existing quadratic keeps the sampled image close to the denoised estimate obtained at the current annealing step. SPC adds a second quadratic that keeps it close to the stored denoised estimate from the preceding step. More specific settings are as follows.

SolverVariableOriginal objective at each stepObjective with SPCSolved by (unchanged)
DPSxix_i∥y−A(x^0(xi))∥22\lVert y-\mathcal{A}(\hat{x}_0(x_i))\rVert_2^2∥y−A(x^0(xi))∥22+βi2∥x^0(xi)−x^0(xi+1)∥22\lVert y-\mathcal{A}(\hat{x}_0(x_i))\rVert_2^2+\frac{\beta_i}{2}\lVert\hat{x}_0(x_i)-\hat{x}_0(x_{i+1})\rVert_2^2One guidance step after the unconditional reverse step xi−1′x'_{i-1}: xi−1=xi−1′−ζi∇xi(⋅)x_{i-1}=x'_{i-1}-\zeta_i\nabla_{x_i}(\cdot), with ζi=ζ′∥y−A(x^0(xi))∥2\zeta_i=\frac{\zeta'}{\lVert y-\mathcal{A}(\hat{x}_0(x_i))\rVert_2}.
DAPSx0x_0∥y−A(x0)∥222βy2+∥x0−x^0(xt)∥222rt2\frac{\lVert y-\mathcal{A}(x_0)\rVert_2^2}{2\beta_y^2}+\frac{\lVert x_0-\hat{x}_0(x_t)\rVert_2^2}{2r_t^2}∥y−A(x0)∥222βy2+∥x0−x^0(xt)∥222rt2+βt∥x0−x^0(xt+1)∥222rt2\frac{\lVert y-\mathcal{A}(x_0)\rVert_2^2}{2\beta_y^2}+\frac{\lVert x_0-\hat{x}_0(x_t)\rVert_2^2}{2r_t^2}+\frac{\beta_t\lVert x_0-\hat{x}_0(x_{t+1})\rVert_2^2}{2r_t^2}Langevin dynamics on this energy: x0←x0−η∇x0(⋅)+2η,ϵ.x_0\leftarrow x_0-\eta\nabla_{x_0}(\cdot)+\sqrt{2\eta}\\,\epsilon. The result is re-noised to level σt−1\sigma_{t-1}.
RED-diffμ\mu∥y−A(μ)∥22+λt(sg[ϵθ(xt;t)−ϵ])⊤μ,\lVert y-\mathcal{A}(\mu)\rVert_2^2+\lambda_t(\mathrm{sg}[\epsilon_\theta(x_t;t)-\epsilon])^{\top}\mu, with xt=αtμ+σtϵx_t=\alpha_t\mu+\sigma_t\epsilon.∥y−A(μ)∥22+λt(sg[ϵθ(xt;t)−ϵ])⊤μ+βt2∥μ−μˉ∥22\lVert y-\mathcal{A}(\mu)\rVert_2^2+\lambda_t(\mathrm{sg}[\epsilon_\theta(x_t;t)-\epsilon])^{\top}\mu+\frac{\beta_t}{2}\lVert\mu-\bar{\mu}\rVert_2^2. The detached snapshot μˉ\bar{\mu} is refreshed every 20 iterations because successive RED-diff iterates are nearly identical.One Adam step on μ\mu per iteration.
PnP-DMzz∥y−A(z)∥222σy2+∥z−x(k)∥222ρk2\frac{\lVert y-\mathcal{A}(z)\rVert_2^2}{2\sigma_y^2}+\frac{\lVert z-x^{(k)}\rVert_2^2}{2\rho_k^2}∥y−A(z)∥222σy2+∥z−x(k)∥222ρk2+βk∥z−x(k−1)∥222ρk2\frac{\lVert y-\mathcal{A}(z)\rVert_2^2}{2\sigma_y^2}+\frac{\lVert z-x^{(k)}\rVert_2^2}{2\rho_k^2}+\frac{\beta_k\lVert z-x^{(k-1)}\rVert_2^2}{2\rho_k^2}Langevin dynamics on this energy gives z(k)z^{(k)} (likelihood step). The prior step then draws x(k+1)x^{(k+1)} by reverse diffusion from z(k)z^{(k)} at noise level ρk\rho_k.
SITCOMvv∥y−A(x^0(v;σt))∥22+λ∥v−xt∥22\lVert y-\mathcal{A}(\hat{x}_0(v;\sigma_t))\rVert_2^2+\lambda\lVert v-x_t\rVert_2^2∥y−A(x^0(v;σt))∥22+λ∥v−xt∥22+βt2∥x^0(v;σt)−x^0(v^t+1;σt+1)∥22\lVert y-\mathcal{A}(\hat{x}_0(v;\sigma_t))\rVert_2^2+\lambda\lVert v-x_t\rVert_2^2+\frac{\beta_t}{2}\lVert\hat{x}_0(v;\sigma_t)-\hat{x}_0(\hat{v}_{t+1};\sigma_{t+1})\rVert_2^2KK Adam steps on vv, starting from v=xtv=x_t, give v^t\hat{v}_t. Then x^0(v^t;σt)\hat{x}_0(\hat{v}_t;\sigma_t) is re-noised to level σt−1\sigma_{t-1}.

Notation: yy is the measurement, A\mathcal{A} is the forward operator, x^0(⋅)\hat{x}_0(\cdot) is the clean-image estimate of the pretrained diffusion model, ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I), and sg[⋅]\mathrm{sg}[\cdot] is the stop-gradient. Each solver keeps the symbols of its paper. βy\beta_y (DAPS) and σy\sigma_y (PnP-DM) are their likelihood parameters and are unrelated to the SPC weight βt\beta_t.

Reviewer_Pd2FOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a method for solving image inverse problems without a pretrained model. It uses an untrained neural network and updates the image through a series of steps. At each step, it adds noise and uses a constraint to keep the new image close to the previous one. The paper gives a theoretical analysis of this process and tests the method on image restoration and CT reconstruction. The results show good performance compared with several existing methods.

Critical strengths
  1. A useful method that does not need pretrained models. TRACE combines ideas from Deep Image Prior and diffusion models. It does not need external training data. This is useful for imaging problems where pretrained models may not be available.

  2. Good experimental results across different tasks. The paper tests the method on several image inverse problems. It also studies different parts of the method and shows that the trajectory constraint can improve existing diffusion-based methods.

Critical weaknesses

-interpretation of impact carried by weight inheritance Table 4 shows that weight inheritance is very important. Removing it reduces PSNR from 32.77 dB to 26.01 dB, a much larger drop than removing proximal coupling. However, the paper gives little explanation of why weight inheritance works or how it fits the theory.

Questions for authors
  1. Table 4 shows that weight inheritance has the largest effect on performance. Why is it so important, and how does it relate to the proximal analysis
  2. Can the authors test the stability bounds in the network-free experiments, where the regularizer is explicitly defined? Which of the tested regularizers satisfy the assumptions?
4 comments
AuthorsOpenReview ↗

Response to W1 & Q1: Interpretation of weight inheritance.

Thank you for pointing out that the role of weight inheritance was insufficiently explained. Weight inheritance is an initialization strategy for the network optimization at each stage: the network starts from the parameters optimized at the preceding stage, rather than from randomly initialized parameters. This provides an initialization already adapted to the current measurements. Although the preceding reconstruction is still available as the input and reference when weight inheritance is removed, the network must fit the stage objective again from a random initialization. A small number of inner iterations may therefore be sufficient with weight inheritance but insufficient without it.

We have added four complementary experiments to examine this interpretation. PSNR values are means over 14 evaluation images per task, reported in dB. Δ\DeltaPSNR denotes the paired difference with and without weight inheritance, using the comparison specified in each caption, and is reported as mean ± standard error. Parentheses give the number of images on which the compared variant with weight inheritance achieves higher PSNR, out of 14. Relative updates are calculated as TK/6000TK/6000, where 6,000 updates correspond to the default setting T=40,K=150T=40,K=150. Weight inheritance retains both network parameters and optimizer state between stages; disabling it reinitializes both. Bold indicates the highest mean PSNR for each task within each table.

Matched optimization budgets (Table R1). At T=40T=40 and K=150K=150, weight inheritance improves PSNR by 2.18 dB on ×4 super-resolution and 5.92 dB on 50% inpainting. Its advantage remains when both variants receive larger budgets: at K=1200K=1200, the gains are 0.91 and 1.26 dB, respectively. Thus, the benefit persists across the tested matched budgets.

More inner iterations without weight inheritance (Table R2). Keeping the inherited reference at T=40,K=150T=40,K=150, increasing the inner iterations of the reinitialized variant from K=150K=150 to K=2400K=2400 improves its PSNR. The gaps to the inherited reference shrink from 2.18 to 0.90 dB and from 5.92 to 0.73 dB in SR and inpainting tasks, respectively. This directly qualifies the original ablation: the large degradation was measured under a fixed, limited inner-iteration budget. Random initialization can yield substantially better results with additional inner optimization, although it remains below the inherited reference within the tested budgets, even with 16 times as many updates.

Fewer inner iterations with weight inheritance (Table R3). Conversely, weight inheritance allows useful reconstructions with fewer updates. With only K=50K=50, it achieves 27.41 and 33.31 dB using 2,000 updates, exceeding the reinitialized variant with K=2400K=2400, which achieves 26.69 and 32.89 dB using 96,000 updates. These results illustrate how initialization affects the number of updates needed to obtain a given reconstruction quality.

More outer stages without weight inheritance (Table R4). Increasing the number of outer stages provides much smaller improvements when each stage still starts from a random initialization and receives only K=150K=150 updates. Increasing TT from 40 to 320 changes PSNR from 25.41 to 25.50 dB on super-resolution and from 27.70 to 28.14 dB on inpainting. At the same total budget of 48,000 updates, allocating more updates within each stage (T=40,K=1200T=40,K=1200) instead achieves 26.55 and 32.62 dB. This supports the interpretation that sufficient optimization within each reinitialized stage matters more than simply adding further stages.

Relation to the proximal analysis. The proximal objective specifies the image-space update, while weight inheritance specifies the initialization used to obtain its neural approximation. In the analysis, the discrepancy between the realized update and the ideal image-space subproblem is represented by ϵt\epsilon_t. The bound does not require weight inheritance, but the initialization can affect the approximate update obtained within a finite number of optimization steps. Our experiments explain why this implementation choice has a large practical effect under the original fixed budget; they assess reconstruction quality rather than directly measuring ϵt\epsilon_t.

We will clarify this distinction and add the budget-dependent results to the revised manuscript.

AuthorsOpenReview ↗

Numerical results of Response to W1 & Q1

Table R1. Effect of weight inheritance at matched optimization budgets. Both variants use T=40T=40 and the same number of inner iterations KK. PSNR values are reported with / without weight inheritance. Δ\DeltaPSNR is the paired difference between these two variants at each budget.

Inner iterations KKRelative updatesSR ×4 PSNR (with / without)SR ×4 Δ\DeltaPSNRInpainting 50% PSNR (with / without)Inpainting 50% Δ\DeltaPSNR
150 (default)1×27.59 / 25.41+2.18 ± 0.27 (14/14)33.62 / 27.70+5.92 ± 0.31 (14/14)
3002×27.43 / 26.02+1.41 ± 0.20 (14/14)33.88 / 30.11+3.78 ± 0.27 (14/14)
6004×27.50 / 26.21+1.28 ± 0.25 (14/14)33.77 / 31.83+1.93 ± 0.14 (14/14)
12008×27.46 / 26.55+0.91 ± 0.17 (14/14)33.89 / 32.62+1.26 ± 0.13 (14/14)

Table R2. Effect of additional inner iterations without weight inheritance. All settings use T=40T=40. The reference with weight inheritance is fixed at K=150K=150, while the variants without weight inheritance use increasing KK. Δ\DeltaPSNR is the reference PSNR minus the PSNR of each variant without weight inheritance.

Weight inheritanceInner iterations KKRelative updatesSR ×4 PSNRSR ×4 Δ\DeltaPSNRInpainting 50% PSNRInpainting 50% Δ\DeltaPSNR
Yes150 (default)1×27.59–33.62–
No1501×25.41+2.18 ± 0.27 (14/14)27.70+5.92 ± 0.31 (14/14)
No2004/3×25.72+1.88 ± 0.26 (14/14)28.85+4.77 ± 0.33 (14/14)
No2505/3×25.95+1.64 ± 0.23 (14/14)29.57+4.05 ± 0.33 (14/14)
No3002×26.02+1.57 ± 0.25 (14/14)30.11+3.51 ± 0.33 (14/14)
No3507/3×26.11+1.48 ± 0.26 (14/14)30.55+3.07 ± 0.31 (14/14)
No4008/3×26.23+1.36 ± 0.20 (14/14)30.87+2.75 ± 0.31 (14/14)
No4503×26.31+1.29 ± 0.20 (14/14)31.19+2.43 ± 0.31 (14/14)
No50010/3×26.37+1.22 ± 0.17 (14/14)31.44+2.18 ± 0.30 (14/14)
No55011/3×26.30+1.29 ± 0.22 (14/14)31.63+1.99 ± 0.29 (14/14)
No6004×26.21+1.38 ± 0.27 (14/14)31.83+1.78 ± 0.28 (14/14)
No9006×26.43+1.16 ± 0.20 (14/14)32.37+1.25 ± 0.28 (13/14)
No12008×26.55+1.04 ± 0.20 (14/14)32.62+1.00 ± 0.25 (12/14)
No240016×26.69+0.90 ± 0.15 (14/14)32.89+0.73 ± 0.25 (12/14)

Table R3. Reconstruction with weight inheritance under reduced inner-iteration budgets. All settings use T=40T=40. Variants with weight inheritance and K∈{25,50,100,150}K\in\{25,50,100,150\} are compared with a reference without weight inheritance using K=2400K=2400. Δ\DeltaPSNR is the PSNR of each variant with weight inheritance minus the reference PSNR.

Weight inheritanceInner iterations KKRelative updatesSR ×4 PSNRSR ×4 Δ\DeltaPSNRInpainting 50% PSNRInpainting 50% Δ\DeltaPSNR
Yes251/6×27.30+0.61 ± 0.09 (14/14)32.58−0.31 ± 0.17 (5/14)
Yes501/3×27.41+0.71 ± 0.10 (14/14)33.31+0.42 ± 0.14 (13/14)
Yes1002/3×27.55+0.85 ± 0.14 (14/14)33.52+0.63 ± 0.13 (12/14)
Yes150 (default)1×27.59+0.90 ± 0.15 (14/14)33.62+0.73 ± 0.25 (12/14)
No240016×26.69–32.89–

Table R4. Effect of additional outer stages without weight inheritance. All settings use K=150K=150. The reference with weight inheritance is fixed at T=40T=40, while the variants without weight inheritance use increasing TT. Δ\DeltaPSNR is the reference PSNR minus the PSNR of each variant without weight inheritance.

Weight inheritanceOuter stages TTRelative updatesSR ×4 PSNRSR ×4 Δ\DeltaPSNRInpainting 50% PSNRInpainting 50% Δ\DeltaPSNR
Yes40 (default)1×27.59–33.62–
No401×25.41+2.18 ± 0.27 (14/14)27.70+5.92 ± 0.31 (14/14)
No802×25.62+1.97 ± 0.21 (14/14)27.80+5.82 ± 0.32 (14/14)
No1604×25.69+1.91 ± 0.21 (14/14)27.99+5.63 ± 0.32 (14/14)
No3208×25.50+2.09 ± 0.23 (14/14)28.14+5.48 ± 0.33 (14/14)
AuthorsOpenReview ↗

Response to Q2: Stability bounds in the network-free experiments.

We empirically evaluate the trajectory stability bound from Proposition 4.1 using the network-free experiments with Tikhonov, Total Variation (TV), and ℓ0.8\ell_{0.8} regularization.

Recall that Proposition 4.1 establishes βt2∥xt−xt+1∥22≤Ft(xt+1)−Ft(xt)+εt,\frac{\beta_t}{2}\Vert \mathbf{x}_t-\mathbf{x}_{t+1}\Vert_2^2 \leq F_t(\mathbf{x}_{t+1})-F_t(\mathbf{x}_t)+\varepsilon_t, where εt≥0\varepsilon_t\geq 0 denotes the suboptimality gap of the local proximal objective.

Rather than explicitly estimating εt\varepsilon_t, we examine whether the realized reconstruction states satisfy the stronger inequality βt2∥xt−xt+1∥22≤Ft(xt+1)−Ft(xt).\frac{\beta_t}{2}\Vert \mathbf{x}_t-\mathbf{x}_{t+1}\Vert_2^2 \leq F_t(\mathbf{x}_{t+1})-F_t(\mathbf{x}_t).

Since εt≥0\varepsilon_t\geq 0, satisfaction of this stronger inequality directly implies the original stability bound. We evaluate its left-hand side (LHS), representing the proximal coupling penalty, and right-hand side (RHS), representing the decrease in FtF_t, using the reconstruction states obtained after finite inner optimization (K=300K=300 steps per outer iteration).

bound_eval1

Figure R2: Empirical verification of the proximal-descent inequality for Tikhonov, TV, and ℓ0.8\ell_{0.8} regularization. Blue denotes the RHS, and green denotes the LHS.

bound_eval2

Figure R3: Zoomed-in view of iterations 10–20. The RHS (blue) remains above the LHS (green) for all three regularizers.

As shown in Figs. R2 and R3, the stronger inequality holds throughout the evaluated TV and ℓ0.8\ell_{0.8} trajectories and from the second iteration onward for Tikhonov. Fig. R3 provides a zoomed-in view of iterations 10–20.

Regarding the assumptions, all three regularizers satisfy the conditions of Proposition 4.1. Specifically, the data-fidelity term and regularization penalties are nonnegative, ensuring that FtF_t is bounded below. The objectives are finite at the evaluated reconstruction states and βt>0\beta_t>0 throughout the coupled experiments. Importantly, Proposition 4.1 does not require convexity or smoothness and therefore applies to the nonconvex ℓ0.8\ell_{0.8} regularizer as well.

We appreciate the reviewer's detailed feedback and would like to briefly summarize the clarifications and additional experimental evidence already presented in our previous rebuttal to address the concerns raised in the original review.

(1) Role of weight inheritance: We conducted four sets of additional experiments (Tables R1-R4) to systematically investigate the role of weight inheritance, as requested by the reviewer. These experiments demonstrate that weight inheritance substantially improves optimization efficiency, enabling reconstructions with only 2,000 updates to outperform those without inheritance using 96,000 updates. We also clarified its relationship to the proximal analysis, where weight inheritance affects the finite step approximation of the ideal image space update.

(2) Empirical validation of stability bounds: As requested by the reviewer, we conducted additional network-free experiments (Figs. R2-R3) to evaluate the trajectory stability bound in Proposition 4.1 using Tikhonov, Total Variation (TV), and nonconvex ℓ₀.₈ regularization. Our results verify a stronger proximal descent inequality throughout the evaluated TV and ℓ₀.₈ trajectories and from the second iteration onward for Tikhonov. We also clarified that all three regularizers satisfy the assumptions of Proposition 4.1, which does not require convexity or smoothness.

Furthermore, in response to concerns raised during the discussion, we conducted extensive additional comparisons against weak diffusion priors and introduced a new evaluation on real-patient dynamic cardiac MRI, demonstrating TRACE's practical advantages in a data constrained setting. Reviewer R3 subsequently acknowledged that their remaining concern regarding practical applicability had been addressed and increased their rating.

Given the extensive theoretical clarifications, controlled experiments and additional practical evaluations conducted in direct response to the reviewers concerns, we believe we have substantially strengthened the evidence supporting TRACE's methodological contribution and practical value. As the discussion period closes on November 18, we would greatly appreciate the reviewer's feedback on whether any concerns remain unresolved. If our responses have addressed the reviewer's concerns, we would be grateful if the reviewer could reconsider and increase their rating in light of the complete evidence presented.

Reviewer_YTmTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a reconstruction framework for imaging inverse problems that requires neither external training data nor pretrained reconstruction models. TRACE iteratively perturbs the current reconstruction and optimizes an untrained network using measurement fidelity and a quadratic consistency penalty anchored to the preceding unperturbed estimate. The authors provide a proximal interpretation, conditional bounds on successive-state variation and approximation-error propagation, and a local Jacobian analysis. Experiments cover natural-image restoration and CT reconstruction, with additional evaluations incorporating the consistency constraint into five pretrained diffusion inverse solvers.

Critical strengths

1.A practically relevant approach without external training data. TRACE addresses reconstruction when a suitable pretrained prior is unavailable. It achieves competitive results across natural-image restoration and CT reconstruction. In particular, it obtains the highest PSNR among the compared methods on limited-angle CT and the highest PSNR among methods without external training in five of the six natural-image settings. These results substantiate its practical usefulness within the evaluated settings.

2.Useful component analyses and extensions to existing solvers. The paper includes controlled ablations of stochastic perturbation and proximal coupling, measurements of null-space Jacobian sensitivity, spectral analyses, and computational comparisons. Applying the consistency constraint to five pretrained diffusion solvers also produces positive average PSNR changes across the evaluated tasks. These experiments provide evidence that the constraint has utility beyond the particular untrained-network implementation.

Critical weaknesses

1.The methodological novelty is limited and primarily combinatorial. SDI already uses scheduled stochastic perturbations and progressive reconstruction through continued optimization of an untrained network. Meanwhile, aSeqDIP already incorporates a quadratic penalty encouraging the network output to remain close to the preceding reconstruction. TRACE combines these existing mechanisms with a coupling schedule and a proximal interpretation. Although this combination is useful, the paper does not sufficiently establish a distinct algorithmic principle or a substantial conceptual advance over these closely related approaches.

2.The inference-time cost of per-instance optimization is not acknowledged in the main text. TRACE shares an important limitation with DIP, SDI, and aSeqDIP: network optimization is performed separately for each test image, imposing substantial reconstruction-time computation. Table 8 reports approximately 76 seconds per image for TRACE on ×4 super-resolution, compared with 0.3–0.5 seconds for pretrained CNN-denoiser-based PnP methods and 19.7 seconds for SITCOM. Although Appendix G.3 acknowledges this trade-off, the main text should explicitly discuss it and include the per-image runtimes from Table 8. Avoiding external pretraining comes with substantial per-instance optimization costs; it does not establish an overall computational advantage.

Questions for authors

1.Could you explicitly compare the objectives and update rules of TRACE, SDI, and aSeqDIP, and identify the technical contribution beyond combining scheduled input perturbation with the existing state-consistency penalty?

2.The term “training-free” is misleading. Every test image requires optimizing all network weights (6,000 Adam updates), so TRACE is not free of training. In addition, the diffusion inverse-problem literature usually uses “training-free” to mean using a pretrained model without task-specific training, which is close to the opposite of this setting. More precise terms would be “training-data-free”, “pretraining-free”, or “self-supervised per-instance optimization”.

3.Sec. 3 states that TRACE is summarized by “three design principles”, but only two are listed (P1 and P2). What is the third principle? If there are only two, please correct the count.

2 comments

Response to W1 & Q1: Methodological novelty relative to SDI and aSeqDIP. TRACE reconstructs an image by sequentially optimizing a pretraining-free network on noise-perturbed inputs while coupling each output to the preceding unperturbed reconstruction.

(1) Distinct formulation. Unlike aSeqDIP, which updates its input deterministically and repeatedly solves the same objective form with a fixed λ\lambda, TRACE constructs a stochastic optimization problem at each outer iteration using annealed perturbations σt\sigma_t and time-varying proximal coupling βt\beta_t. Unlike SDI, which optimizes measurement fidelity on perturbed inputs, TRACE constrains the reconstruction from each perturbed input to remain close to its preceding unperturbed state, introducing a distinct trajectory regularization mechanism.

(2) Distinct regularization mechanism. Proposition 4.3 shows that SDI-type stochastic perturbations induce measurement-space sensitivity regularization proportional to σt2∥AJD∥F2\sigma_t^2\|AJ_D\|_F^2, whereas SPC additionally induces βtσt2∥JD∥F2\beta_t\sigma_t^2\|J_D\|_F^2. Crucially, this term also regularizes measurement-nullspace sensitivity, which measurement fidelity alone cannot directly constrain. This stochasticity-induced image-space Jacobian regularization is absent from both aSeqDIP's deterministic objective and SDI's formulation.

(3) Empirical and theoretical significance. TRACE achieves 29.55 dB in sparse-view CT, compared with 24.36 dB for SDI (+5.19 dB) and 27.15 dB for aSeqDIP (+2.40 dB), while outperforming both across all six natural-image settings. Our controlled ablations and null-space experiments further demonstrate that appropriate SPC strength improves peak PSNR, post-peak stability, and reconstruction of measurement-nullspace components while reducing null-space Jacobian sensitivity, compared with stochastic perturbations alone.

We further report additional PSNR curves of sparse-view CT reconstruction that compare TRACE and SDI over 300 outer stages, with 150 inner iterations per stage (Fig. R1: panel (a), panel (b), panel (c), panel (d), panel (e)). Across all five displayed examples, TRACE has a smooth PSNR curve during extended optimization, whereas SDI exhibits larger fluctuations and subsequent degradation. These results establish SPC as a distinct, significant, and previously overlooked design principle for imaging inverse problems.

We thank the reviewers for raising this question. We will clarify these methodological distinctions in the manuscript.

Response to W2 Thank you for pointing this out. We agree that TRACE incurs substantial reconstruction-time computation because its network weights are optimized separately for each test image. TRACE requires neither external training data nor a pretrained model, but it achieves this through per-instance optimization at reconstruction time. This makes it useful when suitable training data and pretrained priors are unavailable. We agree that avoiding pretraining does not by itself establish an overall computational advantage, and we will clarify this trade-off in the main text.

We additionally evaluate whether fewer inner updates can reduce this cost. Keeping 40 outer steps, we reduce the number of inner updates per step from 150 to 100, 50, and 25. All settings use the same timing protocol as the baselines: one job per GPU, five BSDS68 images, and three seeds per image.

Table R0. Reconstruction time and quality with fewer inner updates. PSNR values are averaged over the five images and three seeds.

Inner updates KKTime per imageSpeed-upSR×4 PSNR (Δ\Delta)Motion deblur PSNR (Δ\Delta)
150 (paper default)76 s1×30.3631.47
10053 s1.5×30.33 (−0.03)30.60 (−0.88)
5026 s2.9×30.32 (−0.04)29.47 (−2.00)
2513 s5.8×30.10 (−0.26)27.83 (−3.64)

Reducing the inner updates from 150 to 25 lowers runtime from 76 to 13 s per image. On this evaluation subset, the corresponding PSNR loss is only 0.26 dB for ×4 super-resolution, but 3.64 dB for motion deblurring. These results show that TRACE permits a practical trade-off between reconstruction time and quality, which can be optimized per task. They do not remove the computational limitation of per-instance optimization.

Response to Q2 & Q3 Thank you for pointing out these ambiguities. We agree that “training-free” can be misleading, since TRACE optimizes its network weights separately for each test image. We will replace “training-free” with “pretraining-free” when describing TRACE throughout the manuscript.

There are two design principles, P1 and P2. We will correct “three design principles” to “two design principles” in Sec. 3.

We appreciate the reviewer's detailed feedback and would like to briefly summarize the clarifications and additional experimental evidence already presented in our previous rebuttal to address the concerns raised in the original review.

(1) Methodological novelty: Unlike aSeqDIP's deterministic sequential updates and SDI's stochastic measurement fidelity optimization, TRACE introduces stochastic proximal consistency (SPC), explicitly coupling reconstructions from perturbed trajectory states to their preceding unperturbed states. Proposition 4.3 further establishes that SPC induces an additional image space Jacobian regularization term under a local approximation, including measurement nullspace sensitivity not directly constrained by measurement fidelity alone. Our controlled ablations, null space experiments, and extended optimization studies substantiate its distinct benefits for reconstruction accuracy and stability. Moreover, incorporating the proposed proximal state consistency constraint improves five existing pretrained diffusion inverse solvers, demonstrating its broader applicability.

We believe an important form of methodological novelty lies in identifying previously overlooked interactions between relevant mathematical principles, characterizing their distinct theoretical consequences, and demonstrating their broader applicability. TRACE makes precisely this contribution through SPC, establishing a distinct trajectory regularization principle whose benefits extend beyond training-data-free reconstruction.

(2) Computational cost: We explicitly acknowledged the cost of per instance optimization and conducted additional runtime quality experiments, demonstrating that TRACE achieves a 5.8× speedup with only a 0.26 dB PSNR reduction for ×4 super-resolution on the evaluated subset, while clarifying the task-dependent nature of this trade-off.

(3) Presentation and terminology: We clarified the distinction between pretraining free reconstruction and per instance optimization and committed to correcting the identified presentation inconsistencies.

Furthermore, in response to concerns raised during the discussion, we conducted extensive additional comparisons against weak diffusion priors and introduced a new evaluation on real-patient dynamic cardiac MRI, demonstrating TRACE's practical advantages in a data-constrained setting. Reviewer samj subsequently acknowledged that their remaining concern regarding practical applicability had been addressed and increased their rating.

Given the extensive theoretical clarifications, controlled experiments and additional practical evaluations conducted in direct response to the reviewer's concerns, we believe we have substantially strengthened the evidence supporting TRACE's methodological novelty and practical value. As the discussion period closes on November 18, we would greatly appreciate the reviewer's feedback on whether any concerns remain unresolved. If our responses have addressed the reviewer's concerns, we would be grateful if the reviewer could reconsider and increase their rating in light of the complete evidence presented.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.