Text-to-3D systems often produce "Janus" objects: a dog with a face on both sides of its head, a statue with two fronts. The cause is that the 2D image generator guiding the 3D optimization judges each rendered view on its own, and from almost any angle it prefers to see the canonical front. The difficulty is that a Janus object is usually locally coherent. Neighboring views agree with each other, and the finished asset is a perfectly consistent 3D shape that happens to have two faces. Checks that compare nearby views, or that reconstruct the views and re-render them, can therefore pass it. The error only shows up when a view is compared with one far around the orbit: the front and the back look like the same side of the object.
This note proposes catching that error the way ranking theory catches inconsistent preferences. We render $M$ views around the object and run a small learned network on many pairs of views. For each pair it predicts how far the camera rotated between them. Because the true camera poses are known, every pair gives a residual, the gap between predicted and true rotation. We then split the residuals over the whole comparison graph with combinatorial Hodge decomposition, a standard tool from statistical ranking that separates pairwise data into a part explained by one score per item and a cyclic part that no such scores can explain. The claim is that the cyclic part (the "curl") is a cheap, localizable certificate of duplicated fronts. A second claim comes from rank-breaking theory, the study of which pairs to extract from multi-item data. Which view pairs we compare decides whether Janus can be seen at all. Comparing only neighbors around the ring is provably blind to it. A sparse set of long-range pairs recovers most of the signal at a fraction of the cost of comparing every pair.
Core Claim
Write $\theta_i$ for the known azimuth of view $i$ and $\hat\phi_{ij}$ for the network's predicted rotation from view $i$ to view $j$. Define the wrapped residual $r_{ij} = \mathrm{wrap}(\hat\phi_{ij} - (\theta_j - \theta_i))$ on each compared pair $(i,j)$, with confidence weight $w_{ij}$ taken from the sharpness of the network's predicted distribution. Following Statistical ranking and combinatorial Hodge theory, we fit one offset $s_i$ per view and keep what is left over:
$$ s^\star = \arg\min_{s} \sum_{(i,j)} w_{ij}\,\big(r_{ij} - (s_j - s_i)\big)^2, \qquad J = \frac{\sum_{(i,j)} w_{ij}\,\big(r_{ij} - (s^\star_j - s^\star_i)\big)^2}{\sum_{(i,j)} w_{ij}}. $$
The offsets $s^\star$ form the "gradient" part. They describe views in which the object itself seems rotated relative to its neighbors, which is the frame-to-frame drift that multi-view diffusion models sometimes produce. The leftover energy $J$ is the curl. A Janus object produces large curl on triangles such as (front, side, back). The two short legs each report roughly 90°, but the long leg reports roughly 0° because both ends show a face, and no choice of per-view offsets can reconcile the three. The edges that carry the curl identify which pair of views is duplicated. That gives a per-view signal, not a single scalar.
We do not claim that pairwise pose estimation is new. We claim that the decomposition separates two different failures (drift versus duplication), that the comparison-graph design determines whether duplication is observable, and that the result is a reward and detector cheap enough to sit inside optimization loops.
How the Two Directions Fit
The rank-breaking literature supplies the method and the design rules, and the Janus problem is where they are applied. Three results matter.
The first is Optimal Sample Complexity of M-wise Data for Top-K Ranking. It breaks an $M$-wise ranking into $M$ cyclic adjacent pairs and shows that this is statistically efficient for recovering the top items. On an azimuth orbit, that circular breaking corresponds to comparing only neighboring views. Our first prediction is that this design, which is optimal for its own goal, is close to useless for ours. On a ring of neighbor edges the only cycle is the full orbit, and a two-faced object still winds once around it with locally correct steps.
The second is Minimax-optimal Inference from Partial Rankings. It shows that treating broken pairs as independent remains near-optimal, and that error is governed by the spectral gap $\lambda_2$ of the comparison graph's Laplacian, roughly a measure of how well connected the graph is. A ring has a gap that shrinks quadratically in $M$, while sparse random graphs with long chords keep it bounded. The variance of $s^\star$, and so the noise floor under $J$, should follow this gap. That gives a testable rule for spending a fixed budget of network calls.
The third, Generalized Rank-Breaking: Computational and Statistical Tradeoffs, warns that pairwise breaking is biased when the underlying score model is wrong. Here that warning becomes the detection principle. Janus is exactly the case where a one-score-per-view model is wrong, and the curl measures the misspecification instead of averaging it away.
On the 3D side, the motivating evidence is the reported persistence of the failure. MVDream adds cross-view attention and camera conditioning trained on Objaverse renders. Even so, DreamCS reports that 52% of MVDream assets show Janus-type flaws, and that 3D mesh supervision is needed to reduce this. DreamReward learns a multi-view reward from human comparisons of whole assets, but by the file's account its Reward3D does not model which views relate to which. Carve3D rewards agreement between generated views and a re-rendering of their feed-forward reconstruction. That checks photometric consistency, but a coherent two-faced shape can satisfy it. Hallo3D asks a large multimodal model to spot hallucinations, which is semantic but slow and gives no smooth loss. Our proposal sits between these. Like Hallo3D it is semantic, but it is fast. Like Carve3D it rests on consistency, but it applies to semantics rather than pixels. Unlike DreamReward and DreamCS, it is defined over view pairs within a single asset.
Research Program
The first component is the pairwise rotation network. We train it on Objaverse renders, pairing two views of the same object at a random azimuth gap with random elevation and lighting. The network outputs a distribution over signed rotation in 10° bins. We hold out object categories for evaluation. Symmetric objects such as vases and balls teach the network to output flat or multimodal distributions, and we turn that ambiguity into low edge weights $w_{ij}$ so that genuinely symmetric objects are not flagged. We want the network to be "fooled" by a duplicated face in the same way a human is, reporting about 0° between a front face and a back face. Training on real, non-Janus objects encourages exactly that behavior, because two face-bearing views of a real head are always nearby. Because residuals near 180° have ambiguous sign, we fit $s$ with a circular loss (one minus the cosine of the error) rather than plain least squares on wrapped values. We report both variants.
The second component is the comparison graph, which we vary. The designs are the neighbor ring (circular breaking), ring plus antipodal chords, ring plus $O(M\log M)$ random chords, and all $\binom{M}{2}$ pairs (Hüllermeier et al.). We measure $\lambda_2$ for each and relate it to the noise in $J$ on clean objects and to detection power on Janus objects.
The third component puts the signal to use in two ways. In score-distillation optimization of a radiance field, the curl carried by each view is used to down-weight or re-noise the views implicated in duplication. In reinforcement-learning fine-tuning of a multi-view diffusion model, we follow Carve3D's setup and substitute $-J$, alone or added to its reward, for the reconstruction-consistency reward.
Baselines
The naive reference is the unmodified generator: MVDream used directly, and MVDream-guided score distillation. DreamCS reports a 52% Janus rate for the latter. We will re-measure that rate under our own labeling protocol rather than assume it transfers.
The strongest detector-free competitor on Janus rate is DreamCS, which reports a drop from 52% to 30%, and to 28% when combined with 2D rewards (DreamCS). It is strongest because it is the only cited method with a reported Janus rate. We run it on the same MVDream backbone, prompts, seeds and per-asset GPU-hour budget. Beating it means a lower Janus rate on our human-labeled set at equal or lower wall-clock cost, given that DreamCS reports about 30% meshization overhead.
The strongest consistency-reward competitor is Carve3D, which reports a mean reconstruction-consistency score of 0.0606 against 0.0892 for its Instant3D base on the 415-prompt DreamFusion set (Carve3D). We fine-tune the same base with the same 30 training prompts, the same policy-gradient recipe and the same number of reward evaluations, swapping only the reward. Beating it means a lower human-labeled Janus rate while staying within a small margin of its consistency score, because our reward does not target that score directly.
The semantic critic baseline is a Hallo3D-style check (Hallo3D) that uses an open multimodal model run on the same views at the same cadence. For overall quality we report GPTEval3D Elo, where DreamReward reports 1480 against MVDream's 1224 (DreamReward), together with T3Bench quality scores, which the T3Bench paper reports as having Spearman correlation above 0.75 with human experts (T3Bench). That paper reports no per-model numbers, so we compute those ourselves.
Three ablations isolate the key ingredients. The first scores with the raw residual energy $\sum w_{ij} r_{ij}^2$ and no decomposition, which tests whether separating drift from duplication matters. The second uses an absolute per-view azimuth classifier, described below. The third is ring-only breaking, which tests the graph-design claim.
Why not just…?
Why not just use Carve3D's reconstruction reward? It is the most mature alternative, and on multi-view diffusion outputs whose views contradict each other it may already catch much of the Janus failure. Our argument is narrower. When the duplicated face is rendered consistently, reconstruction succeeds and re-rendering agrees, so the reward is satisfied. We will test this directly. We label Janus assets in which reconstruction is consistent and compare how well the reconstruction-consistency score and $J$ separate them from clean assets. If reconstruction consistency matches $J$ there, our advantage reduces to cost alone, and we will report it that way.
Why not just ask a multimodal model whether the object has two faces? That is semantically the right question, and it is the closest real threat on detection accuracy. Its known costs are latency, which leads Hallo3D to check only every fourth iteration, and the absence of a per-view gradient-like signal. The detection benchmark compares AUROC and per-asset seconds head to head. If the multimodal model wins clearly on accuracy, $J$ remains useful only as a dense, cheap reward, and the reinforcement-learning results decide whether that is enough.
Why not just classify each view's azimuth relative to the object's front? This is simpler than any graph, and it detects a back view that "looks like a front." It needs a canonical front, however, which many objects lack. It also inherits the canonical-view bias documented in Objaverse, where CLIP top-1 accuracy drops about 20% under non-standard rotations. Relative rotation between two views needs no canonical frame. The absolute classifier is formally the special case of a star-shaped comparison graph with a virtual "canonical" node, so it runs as an ablation on the same data. If it matches $J$ on categories that do have a front, we will say so and restrict the claim to categories without one.
Why not just compare every pair and skip the theory? All-pairs comparison is included and is probably the accuracy ceiling. The question is cost inside an optimization loop, where $\binom{M}{2}$ forward passes per step add up. The rank-breaking prediction is that a sparse chord graph with a large spectral gap comes within a few points of all-pairs AUROC. If it does not, the graph-design contribution fails.
Proposed Experiments
The first experiment is a detection benchmark. We generate assets with MVDream score distillation and plain Stable-Diffusion score distillation on T3Bench and DreamFusion prompts, and annotators label each one for Janus duplication and for which views are duplicated. We report AUROC, accuracy in localizing the duplicated pair, and seconds per asset for $J$, the raw residual, the absolute classifier, reconstruction consistency and the multimodal critic.
The second experiment, the decisive check on graph design, plots detection AUROC against the number of compared pairs for the ring, ring plus antipodes, random chords and all pairs, alongside each graph's $\lambda_2$. The prediction is near-chance performance for the ring and a monotone relation with $\lambda_2$ for the others. If the ring alone detects Janus, the theoretical framing is wrong.
The third experiment tests whether decomposition helps. We plant the two failures separately: synthetic drift, by rotating the object in some rendered frames, and synthetic duplication, by mirroring the front geometry onto the back of Objaverse objects. We then check that drift loads onto $s^\star$ while duplication loads onto $J$. If the raw residual separates the two equally well, the Hodge step is unnecessary.
The fourth experiment is downstream use. We run score distillation with curl-based view reweighting against DreamCS on MVDream, and reinforcement-learning fine-tuning with $-J$ against Carve3D's reward on its base model. Each comparison uses matched budgets and reports Janus rate, reconstruction consistency, GPTEval3D Elo, T3Bench quality and wall-clock time. The headline success criterion is a Janus rate below the 30% DreamCS reports, reproduced under our protocol, with no meshization and with reward evaluation cheaper than a reconstruction round trip.
References
- Statistical ranking and combinatorial Hodge theory
- Optimal Sample Complexity of M-wise Data for Top-K Ranking
- Minimax-optimal Inference from Partial Rankings
- Generalized Rank-Breaking: Computational and Statistical Tradeoffs
- MVDream: Multi-view Diffusion for 3D Generation
- Objaverse: A Universe of Annotated 3D Objects
- DreamCS: Geometry-Aware Text-to-3D Generation with Unpaired 3D Reward Supervision
- DreamReward: Text-to-3D Generation with Human Preference
- Carve3D: Improving Multi-view Reconstruction Consistency for Diffusion Models with RL Finetuning
- Hallo3D: Multi-Modal Hallucination Detection and Mitigation for Consistent 3D Content Generation
- Learning from Label Preferences
- T3Bench: Benchmarking Current Progress in Text-to-3D Generation