Text-to-3D systems that sculpt a 3D object by repeatedly asking a 2D image generator "does this rendered view look right?" often produce the Janus artifact: a dog with a face on the back of its head, or a cat with four faces. The cause is well understood. The 2D model judges each camera view on its own, and because web photos of animals are mostly frontal, it pushes every view toward a face. This note proposes to detect and suppress this failure with a statistic borrowed from quantum optics. Physicists recognize a single-photon source by checking that two photons are almost never found at the same time. We would recognize a single-faced object by checking that two widely separated viewpoints almost never both look like "the front." Concretely, we measure the pairwise correlation of a per-view "frontness" score across the sphere of camera directions and penalize it during optimization. The method needs no assumption about which direction is the front, and it does not push views toward uniform appearance.
The case for this is narrow but real. Current fixes each make an assumption that breaks in known cases. Some fix the canonical azimuth in advance, which fails for an "upside-down lion." Some force a uniform distribution of apparent poses, which corrects global head duplication but leaves local flicker. Some require multi-view training data, which loses the diversity of 2D models and collapses on multi-object scenes. A pairwise, orientation-free exclusion test addresses the property that actually defines the artifact: one canonical part appearing twice.
Core Claim
Score distillation sampling (SDS), introduced by DreamFusion, renders a 3D representation (a neural radiance field or a set of 3D Gaussians) from random cameras $\theta$, adds noise, and uses a frozen 2D diffusion model's denoising direction as a gradient. Let $c(\theta)\in[0,1]$ be a per-view "canonical-ness" score of the current render. Our primary choice is the diffusion model's own denoising certainty, since LCGen shows that 2D diffusion has low denoising variance on canonical views. Define the view-coincidence statistic
$$G_2(\alpha)=\frac{\mathbb{E}_{\theta,\theta':\,d(\theta,\theta')>\alpha}\,[c(\theta)\,c(\theta')]}{\left(\mathbb{E}_\theta[c(\theta)]\right)^2},$$
where $d$ is the angle between two camera directions and $\alpha$ is a separation threshold (for example 90°). The statistic $G_2$ is the analogue of the second-order coherence $g^{(2)}(0)$ used to certify single-photon emission.
A single-faced object concentrates $c$ in one cap of the sphere, so at large separations $G_2$ is near zero ("antibunched"). A two-faced object has two caps, which pushes $G_2(\alpha)$ toward or above one ("bunched"). $G_2$ does not depend on where the face points, so a rotated or upside-down object is scored correctly. A uniformly bland object has small $\mathbb{E}[c]$ but is not penalized for having one clear front. Our claims are, first, that $G_2$ is a reliable, label-free Janus detector, and second, that adding the penalty $\lambda\max(0,G_2-\tau)$ to SDS or variational score distillation suppresses duplicated canonical parts. Here $\tau$ is a tolerance and $\lambda$ a weight. We expect it to do this as well as the best current fixes on standard prompts, and clearly better on prompts with non-standard orientation.
How the physics enters, and where it stops
The quantum-optics papers supply the measurement idea and one design choice. They do not supply a physical model of diffusion. Probing the ladder of dressed states shows a quantum dot in a photonic-crystal cavity switching between photon blockade ($g^{(2)}(0)<1$, a second excitation is energetically forbidden) and photon-induced tunneling ($g^{(2)}(0)>1$), diagnosed by intensity correlations alone. That is the template for $G_2$: we diagnose a second "excitation" of the canonical mode by correlation, without a model of what the face looks like.
Strong photon antibunching in weakly nonlinear two-dimensional exciton-polaritons shows that strong antibunching does not need an enormous repulsive nonlinearity. Destructive interference between two excitation pathways can produce it. That result motivates a second variant of our method. Instead of adding a large penalty, which behaves like brute-force score clipping in D-SDS and slows down coarse shape formation, we remove from the SDS gradient only its component along $\nabla G_2$. The variant projects out the pathway that creates a second face and leaves the rest of the generative push intact.
We deliberately do not pursue two other routes the file suggests. One is using view-consistent generators for 3D nanophotonic cavity inverse design. The file gives no evidence that diffusion-based inverse design such as Diffusion Generative Inverse Design suffers from Janus artifacts, so that bridge would rest on an unsupported premise. The other is anything involving physical "Janus" monolayers such as MoSSe, which share only the name.
Relation to existing Janus fixes
Three lines of work suppress Janus artifacts inside score distillation, and each makes an assumption ours removes. D-SDS clips large scores early and strips view-conflicting prompt words. It also narrows the "front" azimuth, which fixes the front direction in camera coordinates. LCGen permits high certainty only at prescribed canonical angles. We reuse its certainty signal but replace its angle-indexed constraint with a pairwise, rotation-invariant one, because LCGen itself notes failure on non-standard poses. RecDreamer reweights the score so that apparent pose is uniformly distributed, using a DINOv2 pose classifier tracked with an exponential moving average. We borrow that running-buffer device to estimate $G_2$ cheaply. Our objective differs in kind: it forbids coincidence of canonical views rather than demanding a uniform distribution, and RecDreamer reports that its global correction leaves local flicker.
A different strategy changes the generator. Multi-view diffusion models such as MVDream are trained on rendered 3D data. Carve3D fine-tunes such a model with reinforcement learning against a reconstruction-consistency reward. Our method is complementary: it adds a test-time term and needs no fine-tuning, so it can be layered on either kind of generator.
Research Program
The work has three stages, and each must pass before the next is worth running.
The first stage validates the diagnostic. We generate assets with vanilla SDS and VSD on the 70-prompt set from D-SDS and the single-object tier of T3Bench. We then compute $G_2$ from 161 icosahedral views (T3Bench's view set) under three choices of $c$. The first is diffusion certainty. The second is DINOv2 similarity to the object's own most-canonical view, which is self-referenced and needs no category prior. The third is an off-the-shelf face detector. We measure how well each separates Janus from non-Janus assets, using human Janus labels as ground truth and LCGen's Janus Rate definition.
The second stage adds the penalty and the projection variant to threestudio-style SDS and VSD pipelines. $G_2$ is estimated from a rolling buffer of recent renders, so the extra cost is a handful of additional certainty evaluations per step. We report wall-clock time against the 180-minute VSD and 297-minute RecDreamer figures from RecDreamer.
The third stage stresses the two places the idea could break. The first is non-canonical orientation, using a new split of 50 prompts such as "an upside-down lion" or "a dog lying on its back." The second is multi-object scenes. Three corgis legitimately have three faces, and a naive $G_2$ penalty would wrongly suppress them. Following the physics, where blockade holds per emitter and $N$ emitters can hold $N$ excitations, we compute $G_2$ per instance after clustering the 3D representation by density and features. We also test a count-conditioned variant that penalizes only the $(N{+}1)$-th canonical peak when the prompt states $N$. These are evaluated on T3Bench's multi-object tier, where the benchmark reports a performance cliff for all methods.
Baselines
The naive reference is vanilla SDS on DreamFusion-style pipelines and vanilla VSD on ProlificDreamer-style pipelines, with no Janus mitigation. LCGen reports Janus Rates of 56.67% for DreamFusion and 63.33% for ProlificDreamer. We expect our reruns to land in that range and will report our own measured values rather than assume theirs.
The strongest in-family competitor is LCGen. It uses the same certainty signal, targets the same failure, and is plug-in, which makes it the cleanest comparison. It reports Janus Rates of 16.67% on DreamFusion and 20.00% on ProlificDreamer. We will rerun it with its released settings on our backbones, prompts, seeds and iteration budget. Beating it means a lower Janus Rate on the standard prompts at matched CLIP or ImageReward quality, or parity there plus a clear win on the non-canonical split.
RecDreamer is the strongest distribution-level competitor. It reports uFID 165.25 and pose entropy 1.2488. We compare on those metrics and on Janus Rate under the same VSD pipeline. We give it equal hyperparameter-search budget and report runtime alongside.
D-SDS is included as the clipping-and-prompt-debiasing reference. Its reported result is a 90.42% human view-consistency preference over SJC; its countable-faces success number is truncated in our source, so we will measure it ourselves.
As an upper reference we include an MVDream-based SDS run with and without our term. Carve3D's reconstruction-consistency score (0.0606 versus 0.0892 for its base model) serves as a secondary consistency metric.
Two ablations isolate the key ingredient. The first replaces the pairwise $G_2$ with an LCGen-style term at fixed angles, which tests whether orientation-freeness matters. The second compares additive penalty against gradient projection, which tests whether the interference-inspired design matters.
Why not just…?
Why not just use a multi-view diffusion model like MVDream? This is a serious alternative for single canonical objects, and we treat it as an upper reference rather than a straw man. The evidence against relying on it alone has three parts. Carve3D reports that supervised fine-tuning on 3D data costs texture fidelity and diversity. Objaverse shows that even 3D data carries canonical bias, with about a 20% CLIP accuracy drop under random rotations. T3Bench finds MVDream falls off a cliff on multi-object prompts. Because our term is generator-agnostic, the experiment that settles this question is MVDream with our term versus MVDream without it, on the non-canonical and multi-object splits. If MVDream alone already shows near-zero Janus Rate there, our contribution narrows to 2D-prior pipelines, and we will say so.
Why not just count faces with a detector and penalize extra ones? This is the most direct threat, because $G_2$ with a face detector as $c$ is close to exactly that. We include it as a variant. We expect certainty-based or self-referenced DINOv2 scores to outperform it on non-facial canonical parts, such as duplicated car fronts or limbs, and on odd poses where detectors misfire. We also expect the correlation form to be smoother to optimize than a discrete count. If the detector-based penalty matches the certainty-based one across all splits, the honest conclusion is that a simple counting penalty suffices, and the paper becomes an evaluation of that.
Why not just use LCGen or D-SDS clipping, which already work? On standard frontal prompts they may well match us. Both fix the front direction in camera coordinates, however, and LCGen reports failure on non-standard poses. The non-canonical split is designed to expose exactly this difference. If LCGen holds up there, the orientation-free claim fails.
Why not just enforce a uniform pose distribution as RecDreamer does? Uniformity forbids an object from having a dominant front, whereas exclusion only forbids having two. The fixed-angle versus pairwise ablation, together with RecDreamer's reported runtime overhead, will show whether the weaker constraint gives equal Janus suppression at lower cost.
Proposed Experiments
The first decisive check concerns the diagnostic. If certainty-based $G_2$ does not separate human-labeled Janus assets from clean ones substantially better than chance on the D-SDS and T3Bench single-object prompts, the method stops there. The second check is the head-to-head against LCGen and RecDreamer at matched seeds, steps and tuning budget. The outcomes are Janus Rate, T3Bench quality and alignment scores, uFID, pose entropy and runtime. The third check is the non-canonical split, where the orientation-free claim is either confirmed or falsified. The fourth is the T3Bench multi-object tier, where the per-instance and count-conditioned variants must not lower legitimate entity counts. We will count entities in the 12 BLIP-captioned views. A single-instance $G_2$ penalty is expected to fail this test; that is the intended control showing the per-instance correction is necessary. Finally, the projection-versus-penalty ablation tests whether removing only the second-face gradient preserves early shape formation better than clipping. We measure this by the iteration at which coarse geometry stabilizes and by final quality.
References
- DreamFusion: Text-to-3D using 2D Diffusion
- LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D
- Probing the ladder of dressed states and nonclassical light generation in quantum-dot–cavity QED
- Strong photon antibunching in weakly nonlinear two-dimensional exciton-polaritons
- Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation
- Diffusion Generative Inverse Design
- Enhancement of van der Waals Interlayer Coupling through Polar Janus MoSSe
- RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation
- Carve3D: Improving Multi-view Reconstruction Consistency for Diffusion Models with RL Finetuning
- T3Bench: Benchmarking Current Progress in Text-to-3D Generation