Text-to-3D systems often build a 3D object by repeatedly rendering it from random camera angles and asking a 2D image generator whether each render looks right. These systems are called score distillation: a pretrained diffusion model, which generates images by gradually denoising random noise, supplies a gradient that nudges every render toward a plausible image. They regularly produce the "Janus" artifact, an object with a face on the front and another face on the back. We argue that this artifact is best described as a topological defect, and that it can be detected and prevented with a topological constraint. For a fixed camera prompt, the images the 2D model considers plausible fall into a few separate groups: face-forward, profile, and back-of-head. Score distillation lets each camera angle choose its group independently. Nothing forces the groups chosen around a full loop of the camera to match a single solid object, which shows exactly one front and one back.
The proposal has two parts. First, we read out from the diffusion model itself which group each render belongs to, and we count how many separate "front" stretches appear as the camera circles the object. A consistent object has exactly one; a Janus object has two or more. Second, during optimization we use that count as a hard constraint on which view description each camera receives. We also down-weight gradients from the transition views, where the 2D model is caught between groups. Both parts are cheap, use no 3D training data, and require no large multimodal model in the loop. Because the object's front direction is inferred rather than fixed to camera azimuth zero, the method also covers prompts whose front does not sit at the canonical camera pose.
Core Claim
Two existing results meet at this claim. Work on the union-of-manifolds hypothesis showed that image data lies on several disconnected pieces rather than one smooth surface, and that respecting the disconnection helps generation (Verifying the Union of Manifolds Hypothesis for Image Data). Separately, diffusion models were shown to hallucinate by interpolating between such pieces, and this interpolation leaves a measurable trace: high variance in the model's predicted clean image across denoising steps (Understanding Hallucinations in Diffusion Models through Mode Interpolation). That 2D paper never touches 3D lifting. The Janus literature, for its part, treats the artifact as dataset bias toward front views or as contradictions inside the text prompt (Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation), or as a side effect of discrete view encodings (LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D).
Our claim is more specific than "Janus is mode interpolation." Write $\ell(\phi) \in \{F, S, B\}$ for the group (front, side, back) that the 2D prior assigns to the render at camera azimuth $\phi$. As $\phi$ goes once around the circle, a single rigid object must give a cyclic label sequence equivalent to $F\,S\,B\,S$: exactly one front run and one back run. Equivalently, the folded orientation map from the camera circle to the front-to-back interval has degree one. We define the front-run count
$$N_F = \#\{\text{maximal arcs of the camera circle on which } \ell(\phi) = F\},$$
and the Janus condition is simply $N_F \ge 2$. Per-view score distillation optimizes each $\ell(\phi)$ locally, so it has no term that penalizes $N_F$. Conditioning cannot fix this on its own, which explains the tension in the literature. LCGen shows that discrete view prompts create several probability peaks over the sphere. OrientDream switches to continuous pose encodings and still reports a 7.8% multi-face rate (OrientDream: Streamlining Text-to-3D Generation with Explicit Orientation Control). Both approaches act on each view separately. The constraint we need is global around the loop.
Research Program
Reading out the labels. The first component is a label estimator built from the diffusion prior already in the loop. For a render $x$ at azimuth $\phi$, we compute the denoising error under the three view-augmented prompts ("front view of …", "side view of …", "back view of …"). We use shared noise samples and a few timesteps, and turn the errors into a soft posterior $q_\phi(\ell)$. This is the standard way of using a diffusion model as a zero-shot classifier. It reuses forward passes of the same network and adds no new model.
Alongside it we compute the Aithal-style trajectory variance $\mathrm{Hal}(x)$ of the predicted clean image over a short denoising window. The original authors tuned this window by hand per architecture. We will instead calibrate it once, on rendered views of a held-out set of Objaverse objects (Objaverse: A Universe of Annotated 3D Objects), whose true orientations are known.
Constrained label assignment. Every $M$ optimization steps we render $K$ evenly spaced azimuths and find the most probable label sequence that satisfies the one-front, one-back constraint. Under that constraint the labeling is fully described by a front direction $\phi_0$ and two arc widths. A grid search over these three parameters, maximizing $\sum_k \log q_{\phi_k}(\ell_k)$, is exact and inexpensive.
Between sweeps, each sampled camera at azimuth $\phi$ receives the view prompt given by the constrained labeling. It no longer receives the prompt fixed by absolute azimuth, as DreamFusion does (DreamFusion: Text-to-3D using 2D Diffusion). In addition, its score gradient is multiplied by a weight that decreases with $\mathrm{Hal}(x)$. This suppresses the in-between views where the prior is bridging two groups. Because $\phi_0$ is estimated, the object is free to settle into whatever front direction the early geometry suggests. This addresses the "non-canonical orientation" failure flagged as open in the Janus literature.
Monitoring. Throughout training, $N_F$ computed from the unconstrained per-view argmax serves as a diagnostic. It tells us when, and for which prompts, the per-view choices drift toward a Janus configuration.
Scope. The method plugs into score distillation as in DreamFusion or Score Jacobian Chaining (Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation) and into variational variants such as ProlificDreamer. It targets single objects with a meaningful front. Scenes with several instances, such as "three corgis," would require one constraint per instance. We treat that case as a stated limitation, not a claim.
Baselines
Naive reference. Vanilla DreamFusion-style score distillation with azimuth-fixed view prompts and the standard negative prompts. It shows how often Janus occurs when nothing is done. LCGen reports a Janus Rate of 56.67% for DreamFusion and 63.33% for ProlificDreamer on its prompt library (LCGen).
Strongest same-regime competitor. LCGen adds a certainty loss that pushes the denoiser toward a fixed variance profile over azimuth. It is the strongest training-free method that works with an off-the-shelf 2D prior, and it reports Janus Rates of 16.67% (DreamFusion) and 20.00% (ProlificDreamer). Its profile is tied to absolute camera azimuth, whereas ours is inferred and constrained topologically. To beat it, we must match or improve these Janus Rates on its prompt library with no loss in CLIP score. We must also clearly outperform it on prompts with non-canonical orientations.
Debiasing competitor. D-SDS uses score clipping plus language-model pruning of contradictory prompt words (Debiasing Scores and Prompts). It reports an 81.4% success rate on prompts with countable faces, against 29.3% for Score Jacobian Chaining, on its 70-prompt set. We will run it on that same set and expect to exceed 81.4%. Because the two methods address different causes, we will also test the combination D-SDS plus our constraint.
Fine-tuned upper reference. OrientDream fine-tunes Stable Diffusion on posed real-world images and reports a 7.8% multi-face rate. That figure comes from its own evaluation, so it is not directly comparable. OrientDream needs 3D-posed training data, so we treat it as an upper reference rather than a like-for-like competitor. We also test whether our constraint further reduces its residual failures.
Fair comparison. All methods share the same 2D prior, renderer, iteration count, and seeds. We give every method the same wall-clock budget, counting our extra sweep forward passes against us. Each method receives the same number of hyperparameter trials on a disjoint tuning split of prompts.
Ablations. These isolate the key ingredient. The first uses the same label estimator with per-view argmax prompts and no loop constraint, which tests whether the topology matters or the estimator alone helps. The second fixes $\phi_0 = 0$ while keeping the constraint, which tests whether inferring the front direction matters. The third removes the Hal-based weighting, and the fourth uses Hal weighting alone.
Why not just…?
Why not just fine-tune on 3D data like Objaverse? Carve3D reports that supervised fine-tuning on 3D data shifts the model's distribution and erodes its 2D texture quality and open-vocabulary range (Carve3D). Objaverse itself documents noisy, low-poly assets and a roughly 20% drop in CLIP accuracy at non-canonical angles. Fine-tuned multi-view models are nonetheless a real threat on standard prompts, and OrientDream's 7.8% may simply win there. We do not argue otherwise. Our claim is narrower: with an unmodified 2D prior, and on non-canonical prompts, a global constraint closes most of that gap. The OrientDream-plus-constraint run will show whether the constraint is complementary or redundant.
Why not just sanitize prompts, as D-SDS does? Hong et al. themselves show that even a clean prompt leaves the unconditional score biased toward faces. A front-run constraint acts on the outcome rather than on the wording, so it should catch failures that prompt editing misses. The combination experiment measures exactly that residual.
Why not just use a multimodal language model to spot extra faces, as in Hallo3D? Hallo3D (Hallo3D) had to run its critic only every fourth iteration and switch it on late, because of cost. Hong et al. note that geometry mangled early is hard to repair later. Our detector uses the existing prior's forward passes and can act from the first sweeps. We will use a vision-language model only offline, as an independent judge of our $N_F$ detector.
Why not just apply Hal(x) filtering by itself? Variance flags uncertain views but says nothing about whether two confident front views coexist. A Janus object can be confidently wrong at every azimuth. The Hal-only ablation tests this directly. If it matches the full method, the topological part of our claim fails.
Proposed Experiments
The first experiment checks the premise before building the full method. For a few dozen prompts we sample many images from the 2D prior under each view prompt and estimate how much probability mass falls in each label group. This tests whether "back view" prompts really have a sizable, separate face-forward component. We then render 3D outputs from vanilla runs and test whether $N_F \ge 2$ predicts human Janus labels better than a per-view face detector and than LCGen-style certainty profiles, measured by area under the curve. If $N_F$ is not a better predictor, the topological framing is wrong and we stop there.
The main experiments then run on the LCGen prompt library (Janus Rate and CLIP score), on the 70-prompt set from Hong et al. (countable-face success rate and the user-preference protocol), and on the 415-prompt DreamFusion benchmark. For the 415-prompt benchmark we score reconstruction consistency with the round-trip LPIPS measure from Carve3D. Carve3D's 0.0606 was measured on multi-view diffusion outputs, so it serves only as a scale reference, not a direct comparison. We also report CLIP R-Precision, for which DreamFusion reports 88.4% on its MS-COCO subset, to show that text alignment is preserved.
Finally, we build a new set of non-canonical-orientation prompts, such as an upside-down lion or a dog seen from behind as its default pose. We found no reported result on Lacuna for any method on such a set. Here the decisive comparison is the full method against the fixed-$\phi_0$ ablation and against LCGen. The proposal succeeds if inferring the front direction under the loop constraint cuts the Janus Rate substantially where absolute-azimuth methods cannot. It fails if per-view assignment without the constraint does just as well.
References
- Verifying the Union of Manifolds Hypothesis for Image Data
- Understanding Hallucinations in Diffusion Models through Mode Interpolation
- Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation
- LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D
- OrientDream: Streamlining Text-to-3D Generation with Explicit Orientation Control
- Objaverse: A Universe of Annotated 3D Objects
- DreamFusion: Text-to-3D using 2D Diffusion
- Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation
- Carve3D: Improving Multi-view Reconstruction Consistency for Diffusion Models with RL Finetuning
- Hallo3D: Multi-Modal Hallucination Detection and Mitigation for Consistent 3D Content Generation