Quantum implicit neural representations have a hidden flaw. These are small parameterized quantum circuits that map a pixel or 3D coordinate to a color, and they are read out by measuring the average value of an observable. That average is exactly the operation that ruins automatic colorization. When the training targets for a point disagree, say red in one view and blue in another, an averaging readout returns muddy gray. Classical colorization fixed this a decade ago by turning color prediction into classification over a few hundred discrete color bins. A quantum circuit already produces such a categorical distribution for free: measuring nine qubits yields one of 512 outcomes, with probabilities set by the circuit (the Born rule). We propose to read quantum colorization fields out that way. Each measurement outcome is assigned to one quantized color bin, so the circuit's outcome distribution is the color distribution, with no classical output layer.
The point is not that quantum hardware colorizes better than GPUs. It almost certainly does not at current scales. The point is to settle two narrower questions. Does the readout choice, rather than circuit expressivity, decide whether quantum visual models desaturate on ambiguous color targets? And does a readout that needs no parameters give any real parameter saving over a classical softmax head of matched size? We state in advance a result that would make the quantum part pointless, and we will publish that result if it occurs.
Core Claim
Two lines of quantum vision models predict continuous color: Quantum Implicit Neural Representations (QIREN) and QRF: Implicit Neural Representations with Quantum Radiance Fields. QIREN interleaves classical layers with data re-uploading circuits and shows they act as Fourier series, giving 35% lower fitting error than SIREN. QRF replaces NeRF's network with circuits and reports 23.98 PSNR on the Drums scene. Both read color as an expectation value. For a circuit $U_\theta(x)$ acting on input $x$ and an observable $O$, that readout is
$$\hat c(x) = \langle 0|U_\theta(x)^\dagger\, O\, U_\theta(x)|0\rangle = \sum_k p_\theta(k\mid x)\,\lambda_k ,$$
where $p_\theta(k\mid x)=|\langle k|U_\theta(x)|0\rangle|^2$ is the probability of measuring basis state $k$ and $\lambda_k$ is the eigenvalue of $O$ for that state. The readout is therefore a probability-weighted mean by construction.
Under a squared-error loss, training then drives $\hat c(x)$ toward the conditional mean of the target colors. This is the mechanism that Colorful Image Colorization and Learning Representations for Automatic Colorization identified as the cause of desaturation. QIREN's extra frequency capacity cannot help here, because the failure comes from the loss and the readout, not from limited capacity.
Our claim is that dropping the eigenvalue sum is enough to remove this failure. We assign the 313 in-gamut $ab$ bins of Zhang et al. to basis states, and train $p_\theta(k\mid x)$ directly with class-rebalanced cross-entropy. The model then keeps both modes instead of collapsing to gray, while leaving the circuit itself unchanged. A secondary, riskier claim is that this categorical head costs zero output parameters, whereas a classical 313-way softmax costs $313\times d$ weights on top of a $d$-dimensional hidden layer. The file lists this output-layer overhead as an open scaling problem for 3D and video fields.
Research Program
The testbed is per-scene colorization of monochromatic multi-view data, following the setting of Colorizing Monochromatic Radiance Fields (ColorNeRF). ColorNeRF lifts 2D colorizations from CT2: Colorization Transformer via Color Tokens into a NeRF through a classifier over 313 bins. It also shows that continuous regression on these pseudo-labels causes desaturation and ghosting. That setting supplies genuine multimodality: CT2 colorizes each view independently, so the same 3D point often receives conflicting colors across views.
We build three things. The first is a 2D colorization field that maps (x, y, L) to a color distribution, where L is luminance. It is supervised by several stochastic colorizations of one grayscale image. The second is the same field lifted to radiance-field rendering at reduced resolution on the synthetic monochromatic LLFF scenes. The third is a controlled synthetic task in which the true conditional color distribution is known exactly. In that task, objects have their hue flipped between two saturated modes across views with a set probability.
The quantum model is a QIREN-style hybrid with nine qubits and data re-uploading, simulated in PennyLane or TorchQuantum. It uses the parameterized two-qubit gates from QMLP to keep entanglement under control. The 199 basis states that do not correspond to bins are handled in one of two ways. The default is to renormalize their probability mass away. The alternative is a luminance-dependent mask that zeroes out bins implausible for the given lightness, borrowed from CT2's luminance-selecting module. At decode time we apply the annealed-mean operator of Zhang et al. to $p_\theta$, which trades vividness against spatial jitter.
We also measure the cost of estimating $p_\theta$ from a finite number of measurement shots rather than from exact simulation. A 313-way distribution needs many shots, and this cost decides whether the method could ever run on real hardware.
Baselines
The naive reference point is continuous regression in $ab$ space with squared-error loss, run under two architectures. One is a classical SIREN. The other is the identical quantum circuit with expectation readout, which is the QIREN/QRF design. These baselines show the mean trap we expect to see. The expectation-readout circuit doubles as the main ablation. It uses the same circuit, initialization, data and optimizer, and differs only in the readout and loss, so any change in saturation comes from the readout alone.
The strongest competitor on this specific question is a classical coordinate network with a 313-way softmax head and class rebalancing. This is ColorNeRF's injection module and Zhang et al.'s loss, placed on a SIREN trunk. We run it at two sizes. The first is unconstrained. The second matches the hybrid model's total trainable parameters, which forces a narrow or low-rank output layer.
A second classical competitor is a two- or three-component Gaussian mixture head in $ab$ space. This is the cheapest classical way to represent bimodal colors.
A third baseline isolates entanglement. It is a classical model that outputs nine independent Bernoulli probabilities, whose product forms a 512-outcome distribution. This is exactly the Born distribution of an unentangled circuit. If it matches the entangled circuit, then the correlations between bits that entanglement provides are doing no work.
All models use the same scenes, the same pseudo-labels, the same number of optimizer steps, and an equal hyperparameter search of twenty trials each, with learning rate and width chosen by validation-view colorfulness error. As external reference points we cite reported full-scale numbers. On monochromatic LLFF, ColorNeRF reports PSNR 20.76 and colorful score 55.40, ahead of the runner-up's 18.90 PSNR. On ImageNet, CT2 reports colorfulness 38.48 and FID 5.51, and DDColor reports FID 3.92. These are ceilings at much larger scale, not matched comparisons.
Beating a baseline is defined as follows. Against the naive and expectation-readout baselines, the Born-readout field must close at least half of their gap to the target's mean chroma, without lowering PSNR. Against the parameter-matched classical softmax head, it must achieve lower colorfulness error ($\Delta$CF) and higher mode coverage, with confidence intervals across scenes that do not overlap.
Why not just…?
Why not just put a softmax head on a classical SIREN? This is the real threat. It will very likely remove desaturation as well as the Born readout does, since the fix comes from classification, not from anything quantum. We do not argue around this. The parameter-matched softmax baseline exists to settle it. If that baseline matches the hybrid model in colorfulness error and mode coverage at equal parameters, the quantum contribution is null. The paper then becomes a diagnostic result for the quantum implicit representation literature: expectation readout causes the mean trap, and a categorical readout is the fix. That remains useful, because QIREN and QRF currently claim visual quality gains while using a readout that guarantees averaging on ambiguous targets.
Why not just add a colorfulness loss to the regression output? DDColor adds a loss that penalizes low color variance. We will apply it to the expectation-readout circuit. Our prediction is that it raises average saturation without choosing a mode. On the synthetic two-mode task, the output should then drift toward a single hue that may be wrong, or toward an oversaturated color between the two modes. DDColor's own reported failures include garish oversaturation and color bleeding. Mode coverage on the synthetic task will show which of these happens.
Why not use a generative 2D colorizer per view, or ColorNeRF itself? For final image quality, use them. BigColor and diffusion colorizers such as Control Color produce vivid single images. ColorNeRF documents that applying 2D colorizers view by view flickers across views, and its own field is a classical classifier, which is exactly our strongest baseline. Our question is about how quantum fields should be read out, and these systems serve as label sources and ceilings, not as competitors on that question.
Why expect a variational circuit to train at all, given the benchmarking evidence? Benchmarking Variational Quantum Algorithms for Combinatorial Optimization in Practice found that variational solvers barely beat greedy search, and that success probability decayed beyond about 31 variables. We use nine qubits, a supervised loss, and no combinatorial search, so we make no claim of an optimization advantage. Trainability remains an empirical question here. We report gradient variance across random initializations to catch barren plateaus, which are regions where gradients vanish and training stalls.
Proposed Experiments
The first decisive check is the synthetic two-mode task, swept over the flip probability. Expectation readout should place predicted chroma near zero when the flip probability is 0.5. Born readout should place probability on both modes, with mass on each close to the true flip probability. We measure total-variation distance between the predicted and true bin distributions, mean chroma, and mode coverage. Mode coverage is the fraction of pixels whose most likely bin lies within one grid step of a true mode.
The second check runs the 2D field on ImageNet validation images at 64×64. Supervision comes from multiple CT2 samples, and we report colorfulness, $\Delta$CF and PSNR against the ground-truth color image.
The third check runs the reduced-resolution radiance field on monochromatic LLFF. We report PSNR and colorful score using ColorNeRF's protocol, plus cross-view color variance as a measure of consistency.
Across all three, we plot quality against trainable-parameter count for the hybrid model, the classical softmax heads and the product-Bernoulli ablation. We also plot quality against measurement shots per pixel for the hybrid model. The quantum hypothesis survives only if the hybrid curve sits clearly above the matched classical softmax curve and above the unentangled ablation, at a shot count that remains feasible. If either comparison fails, we report the readout diagnosis as the main result.
References
- Quantum Implicit Neural Representations
- QRF: Implicit Neural Representations with Quantum Radiance Fields
- Colorful Image Colorization
- Learning Representations for Automatic Colorization
- Colorizing Monochromatic Radiance Fields
- CT2: Colorization Transformer via Color Tokens
- QMLP: An Error-Tolerant Nonlinear Quantum MLP Architecture using Parameterized Two-Qubit Gates
- DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders
- BigColor: Colorization Using a Generative Color Prior for Natural Images
- Control Color: Multimodal Diffusion-Based Interactive Image Colorization
- Benchmarking variational quantum algorithms for combinatorial optimization in practice