Text-to-3D systems often build a 3D object by repeatedly rendering it from random cameras and asking a 2D image generator whether each render looks right. When they fail, they fail in a characteristic way: the object grows a second face on its back. This is the "Janus problem," and the field mostly treats it as a statistical bias, since web images overwhelmingly show subjects from the front. We propose a different reading, borrowed from polarization optics. A polarizer's transmission follows $\cos^2\theta$, so it cannot distinguish an orientation $\theta$ from $\theta+\pi$. Nanophotonics engineers who need that sign build deliberately asymmetric detectors and read the difference between paired pixels. A 2D image prior judging one view at a time behaves like the polarizer. It confirms "this is a dog" equally well from the front and the back, because front and back silhouettes are mirror images. The only cue that tells it which side faces the camera is weak interior texture, and that cue is biased toward "front."
The proposal is to measure and enforce the sign directly. For every camera at azimuth $\theta$ we also render the opposite camera at $\theta+\pi$. We use the diffusion model itself to score each render for "front versus back." We then penalize the object whenever the two scores fail to be opposite. A consistent object looks frontal from one side and dorsal from the other. A Janus object looks frontal from both, and that shows up as a measurable even-symmetric component in the score around a ring of cameras. In two sentences: Janus artifacts are what you get when the 3D optimizer only ever sees a sign-blind signal, so we add a paired, sign-sensitive one. The same quantity doubles as a cheap, automatic Janus detector that can be checked against human judgments.
Core Claim
Let $s(\theta)\in\mathbb{R}$ be a front-versus-back logit for the render at azimuth $\theta$. It is obtained zero-shot from the diffusion model by comparing its denoising error under a "front view" prompt and a "back view" prompt on the same noised image, so it costs no extra network. Expand $s$ over a ring of $K$ cameras in azimuthal harmonics:
$$ s(\theta)=a_0+\sum_{m\ge 1}\big(a_m\cos m\theta+b_m\sin m\theta\big), $$
where $a_m,b_m$ are fitted Fourier coefficients. A 3D-consistent object with a front has a logit dominated by the first harmonic: positive on one side, negative on the other, so $s(\theta+\pi)=-s(\theta)$. An object with no front, such as a vase, has $s\approx 0$ everywhere. A Janus object has two frontal lobes, which puts energy into $a_0$ and the even harmonics ($m=2,4,\dots$). These are exactly the components that survive the substitution $\theta\to\theta+\pi$, which is the same symmetry that makes Malus' law sign-blind.
We define the Janus energy as the even-symmetric part of the logit and use it both as a metric and as a loss on paired renders:
$$ \mathcal{L}_{\text{anti}}=\mathbb{E}_{\theta}\Big[\big(s(\theta)+s(\theta+\pi)\big)^2\Big], $$
where the expectation is over sampled camera azimuths and gradients flow through both renders. This loss is added to standard score distillation. Score distillation sampling (SDS) is the procedure of nudging the 3D representation so each render becomes more probable under a 2D diffusion model. The claim is that $\mathcal{L}_{\text{anti}}$ removes most Janus artifacts while leaving prompt fidelity intact. Two predictions follow. First, the even-harmonic energy of $s$ should rank objects by human-judged Janus severity better than existing proxies. Second, because the constraint is relative between opposite views rather than tied to an absolute "front" direction, it should degrade less than existing methods on unusual poses.
Where the Idea Comes From and What It Is Not
The physics enters as a design lens rather than as a simulation, and it is worth being explicit about which papers play which role. The method template comes from asymmetric metasurface photodetectors. Liu et al. pair pixels with mirrored asymmetric angular responses and subtract their photocurrents. This differential readout cancels the symmetric background and leaves the sign of a phase gradient, which a symmetric detector cannot see. Our paired-render loss is the same move applied to a learned image scorer. Polarization-controlled holography with flat optics supplies the cleanest example of a response that is even in orientation: the $\cos^2\theta$ amplitude term, decoupled from the geometric phase. We use it only to name the failure: a sign-blind channel.
The polarization reconstruction papers constrain the design. PolarGuide-GSDR and GNeRP both face the $\pi$-ambiguity of polarization normals and resolve it with a coarse global geometry estimate rather than local evidence. PolarGuide uses early Gaussian-splat geometry to choose between candidate normals. That tells us the sign signal should not be trusted before a coarse shape exists, so we switch $\mathcal{L}_{\text{anti}}$ on after a warm-up phase. Like these works, we do not use real polarization sensors. Unlike them, we work in the generative setting, where no measurement exists, and only the symmetry argument transfers.
On the generative side, three papers motivate the gap. DreamFusion introduced view-dependent prompts and still reported multi-faced objects. DreamBooth3D traces the problem to canonical-view bias in the 2D prior. Pure CLIP guidance for voxel NeRFs shows that per-frame scoring trades global 3D integrity for local score. None of these compares opposite views.
The closest competitors are LCGen and RecDreamer. LCGen models discrete view encodings as a Gaussian mixture and lowers the denoising certainty of facial features at non-canonical views. That is a per-view, absolute constraint keyed to which view is called "front." RecDreamer uses a DINOv2 pose classifier with a moving average to push the marginal distribution of rendered poses toward uniform. That is a population-level constraint. A Janus object violates RecDreamer's target, but so do many other objects, and the uniform marginal can be met without any guarantee that each pair of opposite views disagrees. Our constraint differs from both in that it is paired and per-sample. It also needs no template sketches and no external pose classifier.
Carve3D fine-tunes a multi-view diffusion model by reinforcement learning against a reconstruct-and-rerender consistency reward. It is complementary and much heavier. Perp-Neg, which appears in the file only as a baseline row in RecDreamer, uses other views as negative prompts. It is the nearest relative of our idea, and we isolate the difference by ablation below.
Research Program
The work has three parts. The first is diagnostic and settles whether the even-symmetry reading is true before any optimization is built on it. We generate a pool of objects with plain SDS on the DreamFusion prompt set and render each on a ring of $K=16$ azimuths. We compute $s(\theta)$ with the diffusion-classifier construction, averaging the front-minus-back denoising error over a few noise levels and seeds. We fit harmonics and compare the even-harmonic energy ratio against independent human labels of "Janus / not Janus." We repeat the computation with DINOv2 features, the classifier type RecDreamer uses, to check that the signal is not an artifact of using the generator to grade itself. If even-harmonic energy does not separate Janus from non-Janus objects, the core claim is wrong and we stop there.
The second part is the loss. We add $\mathcal{L}_{\text{anti}}$ to SDS in a standard NeRF text-to-3D pipeline and, separately, in a 3D Gaussian Splatting variant. 3D Gaussian Splatting represents the scene as many small colored blobs that are rasterized to images. Each training step samples a camera, renders it and its opposite, and backpropagates the SDS term for the first camera plus the antisymmetry term through both renders. Following the PolarGuide lesson, the loss weight ramps up only after the coarse density stabilizes.
We also handle objects without a front. When $|s|$ stays small on both views, the loss is near zero by construction, and we verify this does not push symmetric objects into fake asymmetry. A ring-level variant penalizes $a_0$ and the even harmonics fitted over all $K$ cameras every few steps. It tests whether whole-ring enforcement adds anything over random pairs.
The third part tests the regime where existing methods are reported to break. LCGen notes failure on multi-subject prompts and non-standard orientations such as upside-down subjects. Our loss refers only to opposite views, not to an absolute canonical direction, so we predict it is less sensitive to orientation. We construct a small prompt subset of upside-down, lying-down, and two-animal prompts to check this. For two subjects, $s$ is a mixture, and we expect our method to help less. We will report that honestly rather than hide it.
Baselines
The naive reference point is DreamFusion-style SDS with view-dependent prompt suffixes, the most common default. On the file's evidence, LCGen reports a Janus rate of 56.67% for the DreamFusion baseline and 63.33% for ProlificDreamer (LCGen). We rerun it in our codebase with the same diffusion backbone, resolution, step count, and seeds as every other method, so that differences reflect the guidance and not the implementation.
The strongest existing method for Janus rate in the file is LCGen. Its reported number is a Janus rate of 16.67% when applied to DreamFusion, and 20.00% when applied to ProlificDreamer. We will run LCGen and our method on the same prompts and backbone, giving each an equal hyperparameter search budget over loss weight and schedule, tuned on a held-out prompt split. Beating it means a lower Janus rate on the test prompts at equal or better CLIP R-precision, with the difference outside seed-to-seed variation. A tie on Janus rate together with a clear win on the non-standard-pose subset would also count as a meaningful result, because that is LCGen's documented weakness.
RecDreamer is the strongest competitor on pose diversity and image quality. It reports an unbiased FID of 165.25 and a pose entropy of 1.2488, at about 297 minutes on an RTX 4090 (RecDreamer). We compare on uFID, pose entropy, and wall-clock time at matched hardware. Our method doubles rendering per step and adds a few classifier evaluations, so we must show it does not exceed RecDreamer's time budget for comparable quality.
For multi-view reconstruction consistency we include Carve3D's reported MRC of 0.0606 on the 415-prompt DreamFusion set, against 0.0892 for its Instant3D-10K starting point (Carve3D). We cite this number as a reference rather than as a matched rerun, because Carve3D fine-tunes a different, multi-view generator and an exact comparison would require its training loop.
Two ablations isolate the key ingredient. The first, unpaired odd guidance, amplifies the front-versus-back logit on each single view toward the label implied by the camera, much like a stronger view-prompt guidance or Perp-Neg. It differs from our method only in removing the pairing. If it matches the paired loss, the pairing is not doing the work. The second swaps the diffusion-classifier logit for a DINOv2-based one, which tests whether self-grading is a weakness.
Why not just…?
Why not just strengthen the view prompts or use negative view prompts? This is the most direct threat, because our logit is built from the same "front view"/"back view" prompts. The file's evidence is that prompt-level view conditioning alone leaves Janus rates above 50% (LCGen). RecDreamer also reports that a prompt like "back view of a dog" is often ignored because of the prior's pose bias (RecDreamer). Our argument is that any per-view signal inherits that bias, while a paired comparison cancels the bias common to both views, just as differential pixels cancel background intensity. The unpaired-odd-guidance ablation settles this directly. If it matches our loss, we will report that the gain came from guidance strength, not pairing, and the contribution reduces to the diagnostic.
Why not just use LCGen or RecDreamer, which already cut Janus rates substantially? They do, and LCGen's 16.67% is a hard number to beat. What they do not provide is a per-object, pose-relative check. LCGen's certainty constraint is defined relative to a canonical front and is reported to fail on non-standard poses. RecDreamer constrains only the marginal pose distribution, at substantial runtime cost. Our loss can also be stacked on either method. If it only ties them on standard prompts, the decisive evidence is the non-standard-pose subset and the combined LCGen-plus-antisymmetry run.
Why not just use a feed-forward multi-view generator such as CAT3D and skip score distillation? Multi-view generators reduce the Janus problem but do not eliminate inconsistency. CAT3D reports views that are not perfectly 3D-consistent and weakness on human faces, and MVD-Fusion describes cross-view attention as an inductive bias rather than a guarantee. Subject-driven pipelines like DreamBooth3D still rely on 2D priors. We concede that for generic text-to-3D the loss may matter less than a better generator. For that reason the Janus-energy diagnostic is designed to apply unchanged to multi-view generator outputs, and we evaluate it there too.
Why not just train against reconstruction consistency with reinforcement learning, as Carve3D does? Carve3D's reward catches any inconsistency, but it requires a reconstruction model in the loop, and its authors flag the reinforcement-learning overhead and the need for careful regularization against reward hacking (Carve3D). Our loss needs no fine-tuning and targets one specific, common failure. The two are complementary rather than competing: our logit could serve as a cheap auxiliary reward. We do not claim to replace Carve3D's general consistency.
Proposed Experiments
The decisive first check is the diagnostic. On SDS outputs for the DreamFusion prompt set, does even-harmonic energy of the front-versus-back logit predict human Janus labels? We measure this as area under the ROC curve, and compare it against the pose-entropy proxy RecDreamer reports and against the DINOv2 version of the logit. A weak correlation falsifies the premise.
The main comparison runs the naive baseline, Perp-Neg, LCGen, RecDreamer, our loss, and LCGen plus our loss. All methods share a backbone, step budget, and three seeds on the DreamFusion prompts. We report Janus rate judged by raters blind to method, CLIP R-precision with the CLIP-L/14 protocol DreamFusion used (88.4% reported for DreamFusion), uFID, pose entropy, and wall-clock time.
The stress test repeats this on the non-standard-pose and multi-subject subset. Our specific prediction is that the paired loss loses less ground than LCGen on rotated subjects and offers little on two-subject scenes.
Finally, we compute Janus energy on MVD-Fusion outputs for Google Scanned Objects views and on CAT3D-style generated rings, where available. This tests whether residual front/back confusion persists in multi-view generators, and whether the diagnostic tracks the Chamfer and PSNR degradations those papers report: MVD-Fusion reports Chamfer distance 0.031 on GSO and PSNR 21.19 on Objaverse.
A positive result would be the diagnostic separating Janus from clean objects, together with the paired loss beating the unpaired ablation. That combination would establish that the Janus problem is usefully understood as a sign-blind, even-symmetry error, and that the fix is a differential measurement rather than a stronger prior.
References
- Asymmetric metasurface photodetectors for single-shot quantitative phase imaging.
- Polarization-controlled optical holography using flat optics
- PolarGuide-GSDR: 3D Gaussian Splatting Driven by Polarization Priors and Deferred Reflection for Real-World Reflective Scenes
- GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors
- DreamFusion: Text-to-3D using 2D Diffusion
- DreamBooth3D: Subject-Driven Text-to-3D Generation
- UNDERSTANDING PURE CLIP GUIDANCE FOR VOXEL GRID NERF MODELS
- LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D
- RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation
- Carve3D: Improving Multi-view Reconstruction Consistency for Diffusion Models with RL Finetuning
- CAT3D: Create Anything in 3D with Multi-View Diffusion Models
- MVD-Fusion: Single-view 3D via Depth-consistent Multi-view Generation