Geometry-Complete Diffusion Model (GCDM)

• The paper introduces a geometry-complete diffusion model that jointly denoises molecular coordinates and features via SE(3)-equivariant networks.
• GCDM leverages explicit local geometric frames and vector message passing to encode orientation, chirality, and force-field properties in 3D molecule generation.
• Empirical results on datasets like QM9 and GEOM-Drugs demonstrate enhanced molecule stability and property prediction compared to traditional methods.

Geometry-Complete Diffusion Model (GCDM) denotes, in its canonical usage, a denoising diffusion probabilistic model for 3D molecule generation and optimization whose denoising network is both geometry-complete and SE(3)-equivariant. In that formulation, geometry-completeness means that denoising is performed not only over atom coordinates but over a richer local geometric representation built from non-degenerate orthonormal frames and vector message passing, so that local orientation, chirality, and force-field-like structure are explicitly represented during generation (Morehead et al., 2023). Subsequent literature has also used the term more broadly, or treated it as a conceptual template for geometry-anchored generative modeling beyond molecules, while the acronym itself is polysemous: in mammogram synthesis, for example, GCDM stands for “Gated Conditional Diffusion Model,” not “Geometry-Complete Diffusion Model” (Cao et al., 26 Jan 2026, Li et al., 25 Jul 2025).

1. Terminological scope and historical positioning

The term “Geometry-Complete Diffusion Model” is introduced explicitly in “Geometry-Complete Diffusion for 3D Molecule Generation and Optimization” (Morehead et al., 2023). In that work, the central claim is that prior 3D molecular diffusion models were “unable to learn important geometric and physical properties of 3D molecules” because they used molecule-agnostic and non-geometric denoisers, whereas geometry-complete denoising requires direction-robust local geometric reference frames and vector-valued message passing. The name therefore refers simultaneously to a modeling objective—complete geometric representation—and to a concrete architectural choice: a geometry-complete, SE(3)-equivariant denoising network.

Later work broadened the phrase from a molecule-specific model name into a more general design principle. FreeOrbit4D formulates a “geometry-complete 4D proxy” for monocular video camera redirection and explicitly presents a conceptual blueprint “Toward a ‘Geometry-Complete Diffusion Model (GCDM)’,” in which diffusion is structurally grounded by an explicit, complete 4D representation rather than left to hallucinate geometry from weak conditioning alone (Cao et al., 26 Jan 2026). In molecular generation, PropMolFlow positions itself as a geometry-complete SE(3)-equivariant flow-matching counterpart to diffusion-based approaches such as GCDM, while JODO and MUDiff instantiate closely related ideas under the headings of “complete molecule generation” and joint 2D–3D diffusion (Zeng et al., 27 May 2025, Huang et al., 2023, Hua et al., 2023).

A separate terminological complication is the acronym collision with the mammography paper “Joint Holistic and Lesion Controllable Mammogram Synthesis via Gated Conditional Diffusion Model,” where GCDM expands to “Gated Conditional Diffusion Model.” That work is geometry- and structure-aware through soft anatomical masks and geometric lesion descriptors, but it is not a Geometry-Complete Diffusion Model in the Morehead–Cheng sense (Li et al., 25 Jul 2025).

Usage context	Meaning of GCDM	Representative paper
3D molecules	Geometry-Complete Diffusion Model	(Morehead et al., 2023)
4D video redirection	Conceptual blueprint for geometry-complete diffusion	(Cao et al., 26 Jan 2026)
Mammogram synthesis	Gated Conditional Diffusion Model	(Li et al., 25 Jul 2025)

2. Geometry-completeness in the original molecular formulation

In the original formulation, a molecule is represented as a fully connected 3D graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$ with node positions $\mathbf{X} \in \mathbb{R}^{N \times 3}$, node scalar features $\mathbf{H} \in \mathbb{R}^{N \times h}$, node vector features $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$, edge scalar features $\mathbf{E} \in \mathbb{R}^{E \times e}$, and edge vector features $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$. The generated object is written as $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$, so diffusion acts jointly on geometry and molecular attributes rather than on coordinates alone (Morehead et al., 2023).

The model is constrained to be SE(3)-equivariant. Scalar quantities remain invariant under rigid motions, whereas coordinates and vector features transform equivariantly. This is not merely a symmetry nicety: the paper’s “Geometry-Complete Denoising” hypothesis states that geometric neural networks that achieve geometry-completeness are more robust in denoising 3D molecular inputs because they “unambiguously define direction-robust local geometric reference frames.” The intended contrast is with prior equivariant molecular DDPMs that pass essentially scalar messages derived from distances and update coordinates only at the output, without explicitly modeling local frames, chirality, or vector-valued directional structure (Morehead et al., 2023).

Geometry-completeness is formalized through a local geometric representation $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$ derived from a pair of node positions $(x_i^t, x_j^t)$. The representation is geometry-complete if it is non-degenerate and forms a local orthonormal basis located at the tangent space of $x_i^t$. In GCDM these basis vectors are constructed as

$\mathbf{X} \in \mathbb{R}^{N \times 3}$0

Because the frame includes cross products, it encodes orientation and chirality rather than only pairwise distance. This is the precise sense in which the denoiser is “geometry-complete”: it carries enough local structure to represent arbitrary directions in 3D, rather than compressing geometry into invariant scalars.

3. Diffusion process and denoising architecture

GCDM adopts a DDPM formulation over joint coordinate–feature latents. With $\mathbf{X} \in \mathbb{R}^{N \times 3}$1, the forward process is

$\mathbf{X} \in \mathbb{R}^{N \times 3}$2

and each transition factorizes as

$\mathbf{X} \in \mathbb{R}^{N \times 3}$3

or, in concatenated form,

$\mathbf{X} \in \mathbb{R}^{N \times 3}$4

The reverse chain is parameterized by replacing the unknown clean sample $\mathbf{X} \in \mathbb{R}^{N \times 3}$5 with a neural estimate $\mathbf{X} \in \mathbb{R}^{N \times 3}$6, obtained from predicted noise $\mathbf{X} \in \mathbb{R}^{N \times 3}$7. Coordinates are generated in the zero-center-of-gravity subspace, following the “zero center of gravity trick” adopted from GeoDiff, so that the marginal over coordinates is invariant to global translations (Morehead et al., 2023).

The denoiser is GCPNet++, an enhanced Geometry-Complete Perceptron Network. Its layerwise operation is described as a Geometry-Complete Graph Convolution acting on scalar and vector channels together: $\mathbf{X} \in \mathbb{R}^{N \times 3}$8 Message passing uses the geometry-complete local frames $\mathbf{X} \in \mathbb{R}^{N \times 3}$9, and coordinate updates are produced through vector outputs: $\mathbf{H} \in \mathbb{R}^{N \times h}$0 A scalar message attention mechanism further gates edge messages and is reported to be crucial for scaling to large molecules (Morehead et al., 2023).

Training uses the standard predict-noise objective

$\mathbf{H} \in \mathbb{R}^{N \times h}$1

with the empirical choice $\mathbf{H} \in \mathbb{R}^{N \times h}$2. In practice, the model uses fully connected graph attention and $\mathbf{H} \in \mathbb{R}^{N \times h}$3 diffusion time steps. The paper also reports that training and sampling with scaled node features $\mathbf{H} \in \mathbb{R}^{N \times h}$4 improves log-likelihood (Morehead et al., 2023).

4. Reported empirical performance and application domains

The original GCDM paper evaluates unconditional 3D molecule generation on QM9 and GEOM-Drugs, property-conditional generation on QM9, protein-pocket–conditional generation, and direct molecular optimization. On QM9 unconditional generation, GCDM reports NLL $\mathbf{H} \in \mathbb{R}^{N \times h}$5, atom stability $\mathbf{H} \in \mathbb{R}^{N \times h}$6, molecule stability $\mathbf{H} \in \mathbb{R}^{N \times h}$7, validity $\mathbf{H} \in \mathbb{R}^{N \times h}$8, and Val & Uniq $\mathbf{H} \in \mathbb{R}^{N \times h}$9. On GEOM-Drugs unconditional generation, it reports NLL $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$0, atom stability $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$1, and molecule stability $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$2, while the baselines summarized in the paper have GEOM-Drugs molecule stability approximately $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$3. For QM9 property-conditional generation, the reported MAEs include $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$4 for $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$5, $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$6 for $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$7, and $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$8 for $\bm{\chi} \in \mathbb{R}^{N \times (m \times 3)}$9, and the paper states that GCDM improves upon GeoLDM for four of the six molecular properties—$\mathbf{E} \in \mathbb{R}^{E \times e}$0, lumo, $\mathbf{E} \in \mathbb{R}^{E \times e}$1, and $\mathbf{E} \in \mathbb{R}^{E \times e}$2—by $\mathbf{E} \in \mathbb{R}^{E \times e}$3, $\mathbf{E} \in \mathbb{R}^{E \times e}$4, $\mathbf{E} \in \mathbb{R}^{E \times e}$5, and $\mathbf{E} \in \mathbb{R}^{E \times e}$6, respectively. In protein-pocket–conditioned generation, GCDM-SBDD-cond on Binding MOAD reports Vina $\mathbf{E} \in \mathbb{R}^{E \times e}$7, SA $\mathbf{E} \in \mathbb{R}^{E \times e}$8, and Diversity $\mathbf{E} \in \mathbb{R}^{E \times e}$9. For property-guided optimization, an example given for $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$0 shows initial EDM samples at MAE $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$1 and molecular stability $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$2, improving to MAE $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$3 and molecular stability $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$4 after GCDM-Opt (Morehead et al., 2023).

Setting	Reported outcome	Significance stated in the paper
QM9 unconditional	NLL $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$5, AS $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$6, MS $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$7	Highest percentage of probable, valid, and unique molecules
GEOM-Drugs unconditional	NLL $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$8, AS $\bm{\xi} \in \mathbb{R}^{E \times (x \times 3)}$9, MS $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$0	First deep learning model that can generate any stable large molecules
QM9 conditional	$\mathcal{M} = [\mathbf{X}, \mathbf{H}]$1, $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$2, $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$3	Improves over GeoLDM for four of six properties
Protein-conditioned SBDD	Vina $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$4, SA $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$5, Diversity $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$6	More synthesizable and diverse ligands with comparable docking scores
Optimization	$\mathcal{M} = [\mathbf{X}, \mathbf{H}]$7: MAE $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$8, MS $\mathcal{M} = [\mathbf{X}, \mathbf{H}]$9	Direct diffusion-based optimization of stability and property

The ablation studies tie these outcomes directly to geometry-completeness. Removing local frames (“GCDM w/o Frames”) degrades QM9 NLL from $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$0 to $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$1 and molecule stability from $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$2 to $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$3, while removing scalar message attention (“GCDM w/o SMA”) reduces QM9 molecule stability to $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$4. On GEOM-Drugs, “w/o Frames” yields molecule stability $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$5, whereas “w/o SMA” collapses to $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$6. The interpretation advanced in the paper is that both geometry-complete frames and attention-mediated scaling are structurally necessary for valid large-molecule generation, not merely incremental refinements (Morehead et al., 2023).

5. Generalizations beyond the original molecular model

A substantial later literature extends the underlying idea of geometry-complete generation, even when the exact name GCDM is not retained. In molecular modeling, MUDiff defines a molecule as $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$7 and learns a joint diffusion process over atom features, 2D bond graph, and 3D coordinates, thereby making “complete molecule generation” mean coherent generation of topology and geometry together rather than either alone (Hua et al., 2023). JODO similarly defines a continuous-time joint diffusion SDE over $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$8, with a Diffusion Graph Transformer that updates node scalars, edge scalars, and coordinates together, and explicitly presents this as complete molecule generation with atom types, formal charges, bond information, and 3D coordinates (Huang et al., 2023). PropMolFlow takes the same completeness criterion into the flow-matching regime: it represents each molecule as $$\bm{\mathcal{F}}_{ij}^{t} = (a_{ij}^{t}, b_{ij}^{t}, c_{ij}^{t}) \in \mathbb{R}^{3 \times 3}$$9, combines continuous flow matching for coordinates with CTMC-based discrete flow matching for atom types, charges, and bond orders, and reports stronger structural metrics than GCDM on QM9 while using $(x_i^t, x_j^t)$0 NFEs rather than $(x_i^t, x_j^t)$1 and reducing time for $(x_i^t, x_j^t)$2k molecules from $(x_i^t, x_j^t)$3 minutes for GCDM to $(x_i^t, x_j^t)$4 minutes (Zeng et al., 27 May 2025).

FreeOrbit4D extends the concept from molecules to dynamic scenes. Its core object is a “geometry-complete 4D proxy”

$(x_i^t, x_j^t)$5

where a static background point cloud is combined with per-frame geometry-complete foreground point clouds aligned into a unified global scene. The paper’s conceptual blueprint for a Geometry-Complete Diffusion Model emphasizes explicit geometry-complete reconstruction, decoupled background/foreground modeling, canonical object space, dense pixel-synchronized 3D–3D alignment, and geometry-guided diffusion via rendered depth scaffolds. In that setting, the defining idea is that diffusion “is not left to guess geometry; it is anchored to an explicit, complete 4D structure” (Cao et al., 26 Jan 2026).

In language modeling, the term is used more analogically than nominally. DiHAL argues that diffusion should enter a pretrained transformer at a hidden layer selected by geometry-based proxies such as local curvature, global monotonicity, and effective rank; diffusion then reconstructs the chosen hidden state rather than tokens directly (Kong et al., 14 May 2026). RDLM, by contrast, locates continuous diffusion on the statistical manifold of categorical distributions, using the Fisher–Rao geometry of the simplex via a sphere diffeomorphism and a bridge-based diffusion mixture formulation (Jo et al., 17 Feb 2025). Neither paper uses “Geometry-Complete Diffusion Model” as an official name, but both treat geometry not as incidental regularization but as the representation space in which diffusion dynamics should be defined.

6. Limitations, controversies, and future directions

The original GCDM paper is explicit that geometry-complete denoising is computationally expensive. Because the model uses fully connected graph attention and $(x_i^t, x_j^t)$6 time steps, generating several thousand large molecules “can take a notable amount of time,” with the paper giving the example of about $(x_i^t, x_j^t)$7 minutes to generate $(x_i^t, x_j^t)$8 new large molecules. Reported hardware requirements are $(x_i^t, x_j^t)$9GB for QM9, $x_i^t$0GB for GEOM-Drugs, and roughly $x_i^t$1 days of training on a single A6000 for a $x_i^t$2-layer GEOM-Drugs model (Morehead et al., 2023). This has made sampling efficiency a central point of comparison, particularly in later flow-matching work.

A second limitation concerns what counts as “complete.” In the original molecular usage, geometry-completeness refers to local frame completeness and joint denoising of coordinates and features, but not to explicit bond-order generation; bonds are inferred post hoc from distances and atom types. Later models such as JODO, MUDiff, and PropMolFlow treat bond information and, in PropMolFlow, formal charges as part of the generative state itself, implying a stricter notion of completeness in which chemically relevant discrete structure is generated rather than reconstructed heuristically (Huang et al., 2023, Hua et al., 2023, Zeng et al., 27 May 2025). A plausible implication is that the meaning of “geometry-complete” has shifted from local geometric expressivity toward globally complete state specification.

Evaluation methodology has also become a point of contention. PropMolFlow argues that existing structural metrics can “mistakenly validate open-shell molecules or molecules with invalid valence-charge configurations,” and introduces closed-shell checks and DFT-based property validation to expose weaknesses that earlier geometry-complete generative models may not reveal (Zeng et al., 27 May 2025). In video generation, FreeOrbit4D identifies failure modes that remain even with geometry-complete scaffolds—heavily occluded parts, very fast motion, thin structures, and inconsistent multi-view generations—showing that explicit geometry does not eliminate ambiguity when the auxiliary geometry-estimation modules themselves fail (Cao et al., 26 Jan 2026). In language, DiHAL notes that geometry-aware hidden-state replacement remains a hybrid dependent on a large pretrained transformer suffix, and that deeper layers are substantially less “bridgeable” than shallow ones (Kong et al., 14 May 2026).

At a more abstract level, recent theory places diffusion and flow matching on the same Wasserstein manifold $x_i^t$3: diffusion follows a free-energy gradient flow and Flow Matching follows a Wasserstein geodesic (Yao et al., 23 Jun 2026). This suggests that future geometry-complete generative models may be defined less by a specific denoiser architecture than by whether the chosen representation space, conditioning signals, and transport dynamics are jointly consistent with an explicit underlying geometry. Under that broader interpretation, GCDM names not only a particular molecular DDPM, but also a research program: generative modeling in which geometry is represented completely enough that denoising or transport can proceed under hard structural constraints rather than weak post hoc correction.