Counted Fins, Locked Phases: Structural Priors Against the Janus Problem in Generated Swimming Animals

Current methods for generating 3D objects from text borrow knowledge from 2D image generators. These are usually diffusion models, generators that learn to turn random noise into images step by step. The 3D shape is optimized so that its renderings from every camera angle look plausible to the 2D model. Because that model judges each view separately and has mostly seen front-facing pictures, the result often shows the "Janus problem": extra faces, duplicated limbs, or front features painted onto the back. Fish, rays and other swimming animals are an especially bad case. They have many similar-looking control surfaces (paired pectoral fins, pelvic, dorsal, anal and tail fins), and their leading and trailing edges differ in subtle ways. When the object must also move, the same failure spreads through time: fins appear, vanish, swap sides or melt between frames.

We propose to build the animal explicitly out of a counted set of parts. There is a body that bends as a traveling wave, plus a fixed list of fins, each with a mirror-image partner where the anatomy calls for one. Every fin moves as a periodic oscillation with its own amplitude and a phase locked to a shared rhythm. This is the representation that bio-inspired robotics uses to coordinate multi-fin swimmers. The 2D diffusion model then chooses fin shapes, amplitudes and phase offsets, rather than a free-form deformation field. Our claim is that this rules out, by construction, the spatial and temporal Janus failures that per-view scoring cannot catch. It should also produce motion that is more coherent than what current text-to-4D methods produce.

Core Claim

Two failure modes need to be separated. The first is inconsistency: different views disagree about the same 3D point. The second is consistent anatomical error: a two-tailed fish or a four-pectoral-fin ray that is perfectly self-consistent from every angle. Existing fixes mostly target the first. Carve3D fine-tunes a multi-view diffusion model with reinforcement learning. Its reward is the Multi-view Reconstruction Consistency score (MRC): the generated views are reconstructed into a 3D field and re-rendered, and the reward measures how far the re-renders drift from the originals. MVD-Fusion makes views agree by generating depth and passing features between views along that depth. Neither method penalizes a duplicated fin that is rendered consistently, because consistency is all they measure. Neither handles motion either.

4D-fy does handle motion. It first fixes static Janus artifacts with a 3D-aware image prior, then adds motion through a video diffusion prior acting on an unconstrained deformation field. The file reports that its motion is limited to simple movements and flickers over time. Our claim is that for animals whose body plan is known, the cheapest and most reliable fix for both failure modes is structural. The representation should make it impossible to count fins wrongly, and it should make fin motion a small set of phase-locked oscillations rather than arbitrary per-frame deformation.

The motion parameterization is borrowed from multi-fin robotics. Design and Control of an Embedded Vision Guided Robotic Fish with Multiple Control Surfaces drives body, tail, pectoral and pelvic fins with central pattern generators. These are networks of coupled oscillators with explicit phase relationships between actuators. In steady swimming, fin $i$ then follows

$$\theta_i(t) = b_i + A_i \sin(\omega t + \phi_i),$$

where $\theta_i$ is the fin's stroke angle, $b_i$ a bias, $A_i$ an amplitude, $\omega$ a shared frequency and $\phi_i$ the phase offset relative to the body wave. Mirror pairs share $A_i$ and $\phi_i$ for straight swimming and are allowed to differ only for turning.

We should be honest about one point. Once the coupled oscillators have settled, they reduce exactly to this set of phase-locked sinusoids. For steady swimming, then, the benefit comes from the part structure and the phase locking, not from the oscillator dynamics as such. The coupled dynamics matter only when the gait changes over time, for example "a fish cruises, then turns." Yu et al. report that coordinated modulation gives smooth transitions between gaits and better pitch and roll stability than plain sine-wave body equations. We treat gait transitions as a separate, secondary claim with its own test.

Research Program

The representation is a dynamic 3D Gaussian scene. 3D Gaussian splatting represents a scene as many small colored, semi-transparent blobs. In our version, each blob is bound to exactly one part: the head-and-trunk, the body segments, or a named fin. The number of parts and their mirror pairings come from a body-plan template chosen per prompt from a small set: a ray-like plan, a generic ray-finned fish, an eel-like plan, and a robotic fish matching the Yu et al. layout. Choosing the plan is a single discrete decision, made by a text classifier, a scored comparison across plans, or the user. The fin shapes are free, which keeps morphological diversity.

The head is rigid and oriented along a fixed forward axis, so the facial Janus failure becomes a question of where one head is placed, not whether a second one grows. Fins deform according to their $\theta_i(t)$. A chordwise flexibility term lets trailing edges bend more than leading edges. This follows Effects of Bionic Bone Flexibility on the Hydrodynamics of Pectoral Fins, which found a stiff leading edge and flexible trailing edge beneficial. We use it only as a regularizer on shape, not as a hydrodynamic claim.

Optimization runs in three stages. Stage 1 optimizes static geometry and appearance with score distillation, the gradient that pushes renderings toward what the 2D model finds likely, introduced by DreamFusion. This stage uses a 3D-aware prior, as 4D-fy does, together with view-dependent prompting. In Stage 2, a video diffusion prior optimizes only the oscillator parameters ($\omega$, $A_i$, $\phi_i$, $b_i$), the body-wave parameters and a small per-fin residual deformation. Stage 3 is optional: a short Carve3D-style reinforcement learning pass using a reward that adds an anatomical term to MRC. The anatomical term is a fin-count check on segmented renders, which lets us test whether rewarding correct anatomy helps beyond enforcing it structurally.

Three analyses run alongside the method. The first is a measurement study of how often and how strongly existing text-to-3D and text-to-4D methods produce duplicated or missing fins on aquatic prompts. Objaverse reports roughly a 20% drop in zero-shot CLIP accuracy on non-canonical rotations, and we expect the pull toward canonical views to be worse for ventral and posterior views of fish. The second tests whether low MRC can coexist with wrong anatomy. The third checks whether the phase offsets recovered from video diffusion priors resemble the coordinated patterns in the robotics literature, or whether the priors prefer arbitrary phases.

Baselines

The naive reference point is DreamFusion-style score distillation with view-dependent prompting on a static representation, extended to motion with an unconstrained deformation field. It shows how bad the problem is when nothing is done. The DreamFusion authors themselves report that the model still sometimes produces multiple faces or redundant parts despite such prompting. The file gives DreamFusion's CLIP R-Precision of 88.4% with CLIP-L/14 on MS-COCO prompts. That number concerns text alignment, not Janus rate, so it serves only as a sanity check that our method does not lose alignment.

The strongest dynamic competitor is 4D-fy. It already removes static Janus artifacts with a 3D-aware prior and adds motion through video score distillation, so it is exactly the pipeline our method modifies. We will use the same 3D-aware image prior, the same video diffusion model, the same number of optimization steps and the same prompts for both. The only difference will be our part-structured, phase-locked motion space in place of its free deformation field. No reported result found on Lacuna for 4D-fy on any quantitative Janus or fin-count metric. Beating it means a lower fin-count error rate and fewer fin appearance and disappearance events over time, at equal or better CLIP alignment and human-rated motion quality.

The strongest static consistency competitor is Carve3D. On the 415 DreamFusion prompts it reports an average MRC of 0.0606, against 0.0892 for the Instant3D-10K base model, and 68.33% of users preferred its consistency. We will compute MRC per frame for all methods using the same reconstruction model. Beating Carve3D on our question does not require a lower MRC. It requires a lower anatomical error rate at comparable MRC, which is the gap we claim Carve3D's reward leaves open. MVD-Fusion is the image-conditioned static reference, with a reported Chamfer Distance of 0.031 on Google Scanned Objects. We include it for single-image-to-3D fish inputs.

Two ablations isolate the key ingredient. The first keeps part structure but replaces phase-locked oscillations with a free per-part deformation over time. This separates the contribution of counting fins from that of locking their phases. The second keeps phase locking but lets the number of fins be learned, with Gaussians free to migrate between fins. This tests whether the fixed count is doing the work.

Why not just…?

Why not just rely on stronger view-dependent prompting or negative prompts? Prompting shifts probabilities but cannot remove a mode the 2D model strongly prefers. The DreamFusion authors say as much, and the canonical-view accuracy drop measured in Objaverse suggests rear and ventral views of uncommon organisms are poorly modeled. The naive baseline is exactly this approach with the best prompts we can write, including negative prompts such as "two tails." If it matches our fin-count error, the structural prior is unnecessary for static shapes.

Why not just add Carve3D's reconstruction-consistency reward? This is the most serious alternative for static shapes. A consistent two-tailed fish reconstructs and re-renders well, though, so MRC should not penalize it. We will test this directly by measuring the correlation between MRC and human-counted anatomical errors across all methods. We will also run the Stage-3 variant that adds a fin-count reward. If rewarded anatomy alone matches enforced anatomy, our contribution shrinks to the motion parameterization. We would report it that way.

Why not just rig an existing fish mesh and animate it with a sine wave? Template fitting would also fix the part count. It would, however, tie shape to whichever meshes exist, and specialized forms such as robotic fish or rays are sparse in web-scraped 3D collections like Objaverse. We will include a template baseline: a rigged mesh retrieved from Objaverse and textured by score distillation. We expect it to score well on anatomy and poorly on prompt fidelity for unusual morphologies. If it scores well on both, it is the better practical answer.

Why not use independent sinusoids instead of coupled oscillators? For steady swimming there is no difference, as noted above, and we use that equivalence. The oscillator coupling is tested only in the gait-transition experiment. There, a prompt asks for a speed or direction change, and we compare coupled-oscillator transitions against switching abruptly between two sets of sinusoid parameters, measuring temporal flicker and human-rated smoothness.

Proposed Experiments

We will build an aquatic prompt set of about one hundred text prompts covering fish, rays, eels, cephalopod-free controls and robotic fish. We will add whichever aquatic prompts exist in the DreamFusion 415-prompt set, so that we can report comparable MRC. For image-conditioned tests we will render reference views from Objaverse-LVIS aquatic categories, to the extent those categories exist. For each asset we render eight azimuths plus top and bottom views over a two-second motion clip.

The primary metrics come from blinded human annotation. The first is fin-count error rate: the fraction of assets whose count of each fin type differs from the body plan implied by the prompt. The second is Janus rate: the fraction of assets with any duplicated head or tail. The third is temporal part-identity events: fin appearances, disappearances or side swaps per clip, counted with segmentation tracking and verified by annotators. Secondary metrics are per-frame MRC as defined by Carve3D, CLIP R-Precision, and pairwise human preference for motion realism.

The decisive check has two parts. The method must cut fin-count error and temporal part events relative to 4D-fy under matched compute. And the MRC-versus-anatomy correlation must be weak enough to show that consistency rewards miss the failure we target. If either fails, the central claim fails.

A secondary check fits the phase parameterization to real footage of fish or of a multi-fin robotic fish. This tests whether the representation can express real gaits at all. We will also compare the fin phase offsets that the video prior settles on against the ranges reported in the robotics work. We treat that comparison as a descriptive finding, not a hydrodynamic validation.