Images from the real world are often damaged in several ways at once, such as rain on top of haze, or sensor noise on top of blur and JPEG compression. The current fix is an "agentic" restoration system: a planner inspects the image and chains together specialist restoration models drawn from a registry of tools. Once those tools are third-party models, each sold per call by a vendor who alone knows what it costs to serve, picking a chain stops being only a vision problem. It becomes a procurement auction. Every vendor has a reason to overstate its price, and the platform has to decide which chain to buy, and how much to pay, without knowing true costs.
This note claims that the auction side of the problem is solvable exactly, and that compound degradation is the source of all its difficulty. Myerson's virtual values, applied to costs rather than bids, are the tool that solves it. A virtual value is a bid corrected for the private-information advantage its owner holds. Each vendor's private information is a single number, its cost, so Myerson's theory of single-parameter mechanisms still applies. The awkward interactions between tools (substitutes, complements, order effects) appear only in the platform's own value for a chain. Revenue-optimal combinatorial design is otherwise out of reach. What remains are empirical questions: whether a learned model of chain value is accurate enough, whether cost distributions can be learned from bids, and how much surplus the optimal mechanism gains over simpler rules.
Core Claim
Let $x$ be a degraded image. Each vendor $j$ offers one tool in one stage. Following the physical ordering in Multi-Agent Image Restoration, which undoes degradation by removing compression artifacts first, then imaging artifacts, then scene effects such as rain or haze, the stages are compression, imaging, and scene. A feasible chain $S$ picks at most one tool per stage, and picking none is allowed. Let $V(S,x)$ be the platform's value for running chain $S$ on $x$, predicted by a model the platform owns. Vendor $j$ has a private per-call cost $c_j$ drawn from a distribution with CDF $G_j$ and density $g_j$. The procurement form of Myerson's result gives vendor $j$ the virtual cost
$$\psi_j(c_j) = c_j + \frac{G_j(c_j)}{g_j(c_j)},$$
which is the true cost plus the information rent a truthful mechanism must concede to that vendor. The platform's expected surplus (value minus payments) is maximized by choosing
$$S^*(x, c) = \arg\max_{S \text{ feasible}} \; V(S,x) - \sum_{j \in S} \psi_j(c_j),$$
and paying each selected vendor its threshold cost, meaning the highest cost at which it would still have been selected. Where $\psi_j$ is not increasing, it is replaced by its "ironed" version, the standard flattening that restores the required monotonicity.
The argument rests on three steps. Each vendor's private type is a scalar. Selection is monotone in each $c_j$ whenever each $\psi_j$ is increasing, however non-additive $V$ is. So the characterization that Approximating Optimal Combinatorial Auctions for Complements Using Restricted Welfare Maximization states for truthful monotone auctions carries over: the platform's surplus equals expected virtual surplus.
The rule also produces a "reserve" that has direct meaning in restoration. The platform should skip a stage entirely when the predicted marginal gain of every tool in that stage falls below its virtual cost, even when the true cost is lower.
Where the Two Directions Meet, and What Is New
The mechanism-design half of this proposal is borrowed, not invented. Virtual-value reasoning was carried into multi-item settings by Approximating Revenue-Maximizing Combinatorial Auctions and Automated Design of Revenue-Maximizing Combinatorial Auctions. Both learn affine boosts on bidder values because exact multi-dimensional characterizations are open. Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition learns data-driven price floors instead.
All of this work treats the combinatorics as living on the bidders' side, and it evaluates on synthetic valuations. Here the combinatorics sit on the buyer's side. That is exactly why the exact single-parameter solution survives, and why these learned-boost methods become baselines rather than necessities. The file notes that Optimal Dynamic Auctions Are Virtual Welfare Maximizers assumes known distributions, and that the $f$-VCG guarantees assume independence. Both limitations bind here as well. We confront the first directly by learning $G_j$ from observed bids.
The restoration half supplies both the setting and the hard part. Multi-Agent Image Restoration and Crafting a Toolchain for Image Restoration by Deep Reinforcement Learning already select tool chains per image. Both evaluate chains on quality alone and treat tools as free, cooperative components. RL-Restore also documents "middle-state" artifacts, where an early tool's output confuses later tools. That is precisely why $V$ cannot be written as a sum of per-tool gains.
RESIDE: A Benchmark for Single Image Dehazing shows that pixel fidelity and downstream detection accuracy can disagree. We therefore define $V$ in two ways, as fidelity gain and as detection gain. We then ask whether the mechanism's choices, and the platform's surplus, change with the definition.
The file's search found no existing work joining virtual values with image restoration. It suggested framing expert routing inside a network as an auction, and we deliberately do not do that, for reasons given below.
Research Program
The first component is a model of chain value, $\hat V(S,x)$. It takes the degraded image, reuses a degradation encoder of the kind in All-In-One Image Restoration for Unknown Corruption, and outputs a predicted gain with an uncertainty estimate for every feasible chain. With three stages and a handful of vendors per stage, the number of chains is small enough to score each one. We train it on offline audit runs, in which every chain is executed on training images and the realized gain is recorded. The key design choice is whether $\hat V$ represents interaction between tools. We compare a model that adds up per-tool gains against one that scores whole chains.
The second component is cost-prior learning. Vendors' bid histories are simulated from cost distributions anchored to each tool's measured latency, multiplied by a random vendor markup. From these bids we estimate $\psi_j$ with smoothing and ironing. We then measure how platform surplus degrades with the number of observed bids, and what happens when costs within a stage are correlated, for example because vendors share hardware.
The third component is analysis. When $\hat V$ is wrong, the mechanism remains truthful, because payments depend only on the platform's own model. Its surplus loss, however, is bounded by the model's error on the chains it compares. We will state and check this bound empirically. The main hypothesis to test is that the optimal mechanism's advantage over simple posted prices grows with how much the value of a chain varies across images. Compound, spatially mixed degradations should produce large variation, and single, mild degradations should produce small variation.
Baselines
The naive reference points are "do nothing," which returns the degraded image at zero cost, and a fixed full chain that runs one default tool per stage under MAIR's ordering and pays each vendor its report. The latter is a pay-as-bid scheme and invites inflated bids. We simulate vendors best-responding to it so the comparison is not flattered.
The cost-blind quality maximizer picks the chain with the highest $\hat V$, which is what current agentic schedulers such as Multi-Agent Image Restoration optimize, and pays posted prices. It bounds how much quality is available. Our mechanism should give up some quality in exchange for lower payments, and we report both quantities.
VCG procurement chooses the chain that maximizes $\hat V$ minus reported cost and pays each vendor its externality. It is truthful and efficient. It is the natural strongest "no priors" competitor, and it is known to overpay when few vendors compete within a stage.
The strongest learned competitors are two. One is an affine-maximizer, or VVCA-style, mechanism with per-vendor weights and per-chain boosts tuned by the local search of Automated Design of Revenue-Maximizing Combinatorial Auctions. The other is $f$-VCG with learned artificial competitors, following Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition. Alongside these we run learned per-tool posted prices. All learned mechanisms receive the same bid samples, the same $\hat V$, and the same tuning budget as ours.
For reference, the only relevant reported number in the file comes from Approximating Revenue-Maximizing Combinatorial Auctions. On a synthetic example, an optimized VVCA reached revenue 2.79 against VCG's 2.45. That is a different, forward-auction setting. No reported result was found on Lacuna for any mechanism that procures image-restoration tools.
Beating the strongest baseline means higher mean platform surplus than the best of VVCA, $f$-VCG, and posted prices, with non-overlapping bootstrap confidence intervals, at equal bid samples and on held-out images. The full-information oracle, which knows true costs and pays exactly those costs, serves as the upper bound.
Two ablations isolate the key ingredients. The first swaps the chain-level $\hat V$ for the additive version while keeping everything else fixed. The second replaces the virtual costs with raw reported costs, which amounts to VCG-style selection without the information-rent correction.
Why not just…?
Why not just put the auction inside the network, routing mixture-of-experts layers by virtual value? This is the obvious combination, and we reject it. Experts inside M2Restore or the prompt components of PromptIR have no private information and no incentive to misreport. Without strategic agents, the virtual-cost correction adds nothing, and budgeted routing reduces to a cost-penalized objective that ordinary gating already handles. Virtual values earn their place only where someone really does hold private costs.
Why not just use a cost-aware router that maximizes predicted gain minus quoted price? If vendors quote honestly, this is the full-information optimum and it beats us. That is precisely the point: it is a pay-as-bid rule, so honest quoting is not a stable behavior. We simulate vendors best-responding to it and report the surplus the platform loses. We also report the gap between our mechanism and the honest-quote ideal. That gap is the unavoidable cost of not knowing vendors' costs.
Why not just post a fixed price per tool? This is the real threat. Posted prices are simple and truthful, and a single-stage, single-vendor market reduces exactly to Myerson's optimal posted price. Our mechanism can only win through competition between vendors within a stage and through variation in value across images. We measure the gap between the two as a function of how much chain values vary. If posted prices come within a few percent on the compound-degradation sets, we will report that as the main finding.
Why not just use VCG? VCG is truthful and needs no priors. Its payments, however, are not tuned to the cost distributions and can be large when stages are thin. The VCG baseline and the raw-cost ablation together measure how much of our gain comes specifically from the information-rent term.
Proposed Experiments
The images come from four sources. We use MAIR's synthetic groups and its 100 paired real-world images, drawn from I-Haze, NH-Haze, DRealSR, RealSR, T-OLED, SIDD, and LHP-Rain. We use the All-weather dataset (Snow100K, Raindrop, and Outdoor-Rain) and DIV2K with mixed blur, noise, and JPEG corruption. We use RTTS for detection-based value.
Vendors are publicly available restoration models, several per stage, whose latencies we measure ourselves. The file reports 0.17 seconds for M2Restore on an RTX 3090 and roughly tens of seconds per image for MAIR, which indicates the spread of costs to expect. The value model is fit on training splits, and all mechanism evaluation uses held-out images and fresh cost draws.
We report platform surplus, realized quality gain, total payments, vendor utility, the rate at which stages are skipped, and truthfulness regret for the learned mechanisms that are not truthful by construction. Value is defined in two ways: PSNR gain on paired sets, and gain in Faster R-CNN mAP on RTTS.
Four checks decide the proposal. The first is whether the chain-level $\hat V$ beats the additive ablation in surplus on compound sets but not on single-degradation sets. That result would confirm that tool interactions are what make the problem combinatorial. The second is whether the optimal mechanism beats learned posted prices, and by how much, as value variation increases. The third is how many bid samples are needed before learned virtual costs come within a small margin of the known-prior mechanism. The fourth is whether switching the value definition from PSNR to detection mAP changes which chains are bought. RESIDE's observation that fidelity and detection can disagree predicts that it will.
References
- Multi-Agent Image Restoration
- Approximating Optimal Combinatorial Auctions for Complements Using Restricted Welfare Maximization
- Approximating Revenue-Maximizing Combinatorial Auctions
- Automated Design of Revenue-Maximizing Combinatorial Auctions
- Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition
- Optimal Dynamic Auctions Are Virtual Welfare Maximizers
- Crafting a Toolchain for Image Restoration by Deep Reinforcement Learning
- RESIDE: A Benchmark for Single Image Dehazing
- All-In-One Image Restoration for Unknown Corruption
- M2Restore: Mixture-of-Experts-Based Mamba-CNN Fusion Framework for All-in-One Image Restoration
- PromptIR: Prompting for All-in-One Blind Image Restoration