PoseFM: Generative Pose Inference Methods

• PoseFM is a term used for various generative pose estimation methods that replace deterministic pose regression with conditional inference techniques.
• The visual odometry variant employs monocular flow matching and ODE integration to estimate relative camera motion while incorporating uncertainty through sampling.
• The Flose method and PFAvatar’s pose fusion illustrate a shift toward modeling pose as a conditional distribution, improving robustness against noise and ambiguities.

PoseFM is a label used in recent computer-vision literature for distinct pose-centric formulations rather than a single unified method. In one usage, it denotes a monocular visual odometry framework that reformulates relative camera pose estimation as conditional flow matching over pose parameters (Kuczkowski et al., 24 Apr 2026). In another, the term refers to Flose, a generative instance-level 6D pose estimator that frames rigid alignment as conditional flow matching in $\mathbb{R}^3$ and then recovers an $\mathrm{SE}(3)$ pose by robust registration (Hamza et al., 23 Feb 2026). In a third usage, introduced for clarity rather than as an original paper term, PoseFM denotes the pose-fusion mechanism in PFAvatar, where pose cues are fused with appearance in a two-stage pipeline for personalized avatar reconstruction (Xi et al., 17 Nov 2025). Taken together, these works place PoseFM at the intersection of generative modeling, geometric inference, and conditional pose guidance.

1. Terminological scope

The term PoseFM is used in three different senses across the cited works.

Usage	Paper	Core meaning
PoseFM	"PoseFM: Relative Camera Pose Estimation Through Flow Matching" (Kuczkowski et al., 24 Apr 2026)	Relative camera pose estimation through Flow Matching
PoseFM as Flose	"Generative 6D Pose Estimation via Conditional Flow Matching" (Hamza et al., 23 Feb 2026)	Generative 6D pose estimation by denoising point sets in $\mathbb{R}^3$
PoseFM as Pose-Fusion mechanism	"PFAvatar: Pose-Fusion 3D Personalized Avatar Reconstruction from Real-World Outfit-of-the-Day Photos" (Xi et al., 17 Nov 2025)	Coupled pose-aware diffusion and pose-conditioned NeRF distillation

The most explicit use is the visual-odometry paper whose title defines PoseFM as “Relative Camera Pose Estimation Through Flow Matching” (Kuczkowski et al., 24 Apr 2026). In the 6D object-pose paper, the detailed description states that “PoseFM in this work is Flose,” identifying the method as a conditional-flow-matching formulation for pose recovery in $\mathbb{R}^3$ rather than direct regression in $\mathrm{SE}(3)$ (Hamza et al., 23 Feb 2026). In PFAvatar, by contrast, the paper “does not explicitly name a component ‘PoseFM,’” and the designation is applied to the combined pose-fusion design spanning ControlBooth and BoothAvatar (Xi et al., 17 Nov 2025).

This suggests that PoseFM is best understood as a family resemblance across methods that replace or augment deterministic pose estimation with generative transport, denoising, or pose-conditioned guidance. A plausible implication is that the shared label reflects a broader shift from point estimation toward conditional generative inference in pose-related tasks.

2. PoseFM for relative camera pose estimation

In monocular frame-to-frame visual odometry, PoseFM addresses the problem: given two consecutive RGB frames $(I_t, I_{t+1})$, estimate the relative $6$-DoF camera motion $T \in \mathrm{SE}(3)$ between the views (Kuczkowski et al., 24 Apr 2026). The method conditions on dense optical-flow features extracted from the image pair and models a conditional distribution $p(T \mid I_t, I_{t+1})$ rather than a single deterministic estimate. The stated motivation is that deterministic regression lacks the uncertainty awareness required for robust applications, especially under weak texture, motion blur, lighting changes, and inherent motion ambiguity.

The pose is represented in a Euclidean parameterization to enable flow matching. Rotations are represented in the Lie algebra $\mathfrak{so}(3)$ by an axis-angle vector $\mathrm{SE}(3)$0, translations as $\mathrm{SE}(3)$1, and the exponential map from $\mathrm{SE}(3)$2 to $\mathrm{SE}(3)$3 reconstructs the rigid transform. The paper gives

$\mathrm{SE}(3)$4

and

$\mathrm{SE}(3)$5

When $\mathrm{SE}(3)$6, the usual series limits are applied for numerical stability (Kuczkowski et al., 24 Apr 2026).

The flow-matching formulation learns a conditional time-dependent vector field $\mathrm{SE}(3)$7:

$\mathrm{SE}(3)$8

where $\mathrm{SE}(3)$9 denotes visual features from a frozen optical-flow encoder (Kuczkowski et al., 24 Apr 2026). Training uses Conditional Flow Matching with a linear path,

$\mathbb{R}^3$0

and the objective

$\mathbb{R}^3$1

The base distribution factorizes as $\mathbb{R}^3$2 and $\mathbb{R}^3$3, with rotation samples mapped to $\mathbb{R}^3$4 by the logarithm map (Kuczkowski et al., 24 Apr 2026).

The architecture is a two-stage pipeline. A frozen optical-flow front-end provides dense motion cues, and the paper supports PWCNet and WAFT. The vector-field network extends the TartanVO architecture and includes a ResNet backbone, a time embedding for $\mathbb{R}^3$5, a pose-state embedding for $\mathbb{R}^3$6, fusion of these embeddings with visual features, and two prediction heads for translation-space and rotation-space velocity (Kuczkowski et al., 24 Apr 2026). An intrinsics layer, as in TartanVO/CUVO, is used for cross-dataset generalization, and RandomCropResize is used for augmentation.

Training is supervised on TartanAir with Adam on a single NVIDIA Tesla V100. For PoseFM+PWC, the reported schedule is 100 epochs on ground-truth optical flow with batch size 128 and learning rate $\mathbb{R}^3$7, followed by 50 epochs of fine-tuning on PWCNet-predicted flow with batch size 96 and learning rate $\mathbb{R}^3$8; the optical-flow encoder remains frozen. PoseFM+WAFT uses precomputed WAFT flow and trains the vector-field network for 100 epochs with the same optimizer and learning-rate schedule (Kuczkowski et al., 24 Apr 2026).

At inference, the probability-flow ODE is integrated from $\mathbb{R}^3$9 to $\mathbb{R}^3$0 with a fixed-step midpoint solver using 5 steps by default. The method samples $\mathbb{R}^3$1 initial conditions, with $\mathbb{R}^3$2 in experiments, averages the resulting poses for a point estimate, and summarizes uncertainty by sample variance or standard deviation across generated poses (Kuczkowski et al., 24 Apr 2026). This is the paper’s central distinction from deterministic frame-to-frame visual odometry.

3. Empirical behavior in monocular visual odometry

Evaluation is reported on TartanAir, KITTI odometry, and TUM-RGBD using Absolute Trajectory Error (ATE) with $\mathbb{R}^3$3-DoF $\mathbb{R}^3$4 alignment to address monocular scale ambiguity:

$\mathbb{R}^3$5

The paper also provides the Relative Pose Error definition for completeness, although it is not reported in the results tables (Kuczkowski et al., 24 Apr 2026).

On TartanAir, among frame-to-frame methods, the reported averages are TartanVO 3.34, DytanVO 3.90, CUVO 1.99, PoseFM+PWC 3.08, and PoseFM+WAFT 2.96 (Kuczkowski et al., 24 Apr 2026). The paper states that PoseFM+PWC is approximately $\mathbb{R}^3$6 better than TartanVO and that each PoseFM variant achieves the best ATE on at least one trajectory. On KITTI odometry without fine-tuning, PoseFM+PWC reports an average ATE of 38.74, compared with CUVO at 23.74, DytanVO at 26.66, and TartanVO at 33.94; PoseFM+PWC achieves best ATE on one sequence and second-best on two sequences among frame-to-frame VO methods (Kuczkowski et al., 24 Apr 2026). PoseFM+WAFT underperforms on KITTI, which the paper attributes to domain shift in intrinsics and motion statistics.

On TUM-RGBD without fine-tuning, PoseFM+PWC reports an average ATE of 0.203, nearly matching CUVO at 0.193 and improving over TartanVO at 0.32 (Kuczkowski et al., 24 Apr 2026). The method is best among frame-to-frame methods on 3 of 9 scenes—360, desk2, and floor—while multi-frame approaches such as DPVO and MambaVO remain stronger overall.

The ablations reported for PoseFM are tied to the behavior of the conditional flow formulation. Fine-tuning on predicted flow is described as essential for the PWC variant, since without it average ATE degrades by approximately $\mathbb{R}^3$7 (Kuczkowski et al., 24 Apr 2026). For ODE discretization depth, 5 steps performs best, while 2 or 10 steps remain close, with example average ATE values 3.03, 2.90, and 2.98 respectively (Kuczkowski et al., 24 Apr 2026). The comparison between PWC and WAFT under GT-flow training and predicted-flow testing yields similar average ATE, 3.71 versus 3.59, which the paper interprets as robustness to front-end choice.

These results position PoseFM as uncertainty-aware and competitive with established frame-to-frame regressors, but not as a replacement for multi-frame or globally optimized systems. The paper explicitly notes that the method does not perform temporal smoothing or loop closure (Kuczkowski et al., 24 Apr 2026).

4. PoseFM as Flose in instance-level 6D object pose estimation

In "Generative 6D Pose Estimation via Conditional Flow Matching," the detailed description identifies PoseFM with Flose, a generative instance-level 6D pose estimator (Hamza et al., 23 Feb 2026). The task is instance-level 6D pose estimation: given a known 3D object model and an RGBD image with a target instance visible, estimate the rigid transformation relative to the camera. If the query model point cloud is $\mathbb{R}^3$8 and the target observation is $\mathbb{R}^3$9, the goal is to recover a pose $\mathrm{SE}(3)$0 that aligns $\mathrm{SE}(3)$1 to the observed target (Hamza et al., 23 Feb 2026).

The paper’s motivation is explicit. Direct $\mathrm{SE}(3)$2 regression struggles on symmetric objects because multiple poses can be equally valid while training typically uses a single ground-truth pose; indirect methods based on local correspondences are vulnerable when local geometric or texture cues are weak and fail under occlusion or clutter (Hamza et al., 23 Feb 2026). Flose therefore formulates pose estimation as conditional flow matching in $\mathrm{SE}(3)$3, where the model denoises point positions rather than regressing pose variables on the manifold.

The state is built as

$\mathrm{SE}(3)$4

with a linear interpolation toward a Gaussian-noise endpoint $\mathrm{SE}(3)$5:

$\mathrm{SE}(3)$6

The conditional signal concatenates learned point-wise features $\mathrm{SE}(3)$7 and positional encodings $\mathrm{SE}(3)$8:

$\mathrm{SE}(3)$9

A neural flow model $(I_t, I_{t+1})$0 predicts a velocity field

$(I_t, I_{t+1})$1

and is trained with

$(I_t, I_{t+1})$2

At test time, deterministic Euler integration is used:

$(I_t, I_{t+1})$3

After $(I_t, I_{t+1})$4 steps, the model produces $(I_t, I_{t+1})$5, where $(I_t, I_{t+1})$6 approximates the rigidly aligned target $(I_t, I_{t+1})$7 (Hamza et al., 23 Feb 2026).

A crucial design choice is that the flow is learned in $(I_t, I_{t+1})$8 rather than on $(I_t, I_{t+1})$9. During inference, velocity components for the query $6$0 are zeroed so that only the target subset moves:

$6$1

The final rigid pose is then recovered by robust 3D-3D registration between $6$2 and $6$3, rather than by direct regression of $6$4 and $6$5 (Hamza et al., 23 Feb 2026). This is the mechanism by which Flose sidesteps symmetry-induced ambiguity in $6$6-space losses.

The feature pipeline combines overlap-aware geometry with appearance-aware semantics. Overlap-aware features come from $6$7 (PointTransformerV3), trained to classify overlapping points using nearest-neighbor labels with $6$8 of object diameter. Appearance-aware features come from a frozen DINOv2-Giant, lifted from RGB to points for the target and fused from multi-view renders for the query; both are PCA-projected to $6$9. Fusion is

$T \in \mathrm{SE}(3)$0

The positional encoder uses sinusoidal embeddings over coordinates, normals, noisy-initialization coordinates, and a binary $T \in \mathrm{SE}(3)$1 flag, linearly projected to $T \in \mathrm{SE}(3)$2 with $T \in \mathrm{SE}(3)$3 (Hamza et al., 23 Feb 2026).

Pose recovery uses RANSAC plus Kabsch to solve the orthogonal Procrustes problem,

$T \in \mathrm{SE}(3)$4

with the candidate chosen by inlier count. ICP refinement is then run for 3000 iterations with a 1 cm threshold (Hamza et al., 23 Feb 2026).

5. Empirical profile of Flose and its relation to the PoseFM paradigm

Flose is evaluated on five BOP datasets—LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V—using BOP Average Recall (AR) (Hamza et al., 23 Feb 2026). The reported AR values are LM-O 86.1, T-LESS 86.9, TUD-L 98.8, IC-BIN 74.8, YCB-V 92.8, and average 87.9. Against the strongest per-dataset competitor PFA, the average gain is +4.5 AR overall, with per-dataset gains LM-O +6.4, T-LESS +1.9, TUD-L +2.8, IC-BIN +7.2, and YCB-V +4.0 (Hamza et al., 23 Feb 2026). The paper also reports that, against the per-object regressor GDRNPP (BOP’23), Flose attains +1.2 Avg AR higher while requiring approximately $T \in \mathrm{SE}(3)$5 fewer models, 54 versus 5, because Flose uses a single model per dataset (Hamza et al., 23 Feb 2026).

The ablation evidence in LM-O is directed at the paper’s central claims about conditioning. Fusing appearance and overlap features yields +15.0 AR over appearance-only and +2.6 AR over overlap-only, with strict-threshold Inlier Ratio gains such as +30% IR at $T \in \mathrm{SE}(3)$6 diameter (Hamza et al., 23 Feb 2026). For symmetric objects, gains are reported as +3.95 AR on LM-O symmetric items such as Eggbox and Glue bottle, compared with +2.12 AR on other objects (Hamza et al., 23 Feb 2026). The paper interprets this as evidence that semantic features help disambiguate object symmetries.

Outlier handling is another explicit component of the method. Because flow fields are predicted per point without enforcing rigidity, the deformed target $T \in \mathrm{SE}(3)$7 can contain substantial outliers: at $T \in \mathrm{SE}(3)$8 diameter, only 18% IR is reported, and even at $T \in \mathrm{SE}(3)$9, approximately 35% of points are outliers (Hamza et al., 23 Feb 2026). The paper compares SVD, SVD+ICP, RANSAC, and RANSAC+ICP, reporting 81.0 AR, 84.2 AR, 81.8 AR, and 86.1 AR respectively (Hamza et al., 23 Feb 2026). This directly motivates the use of RANSAC-based Kabsch alignment rather than global SVD on all correspondences.

The reported runtime scales linearly with the number of Euler steps: 1 step gives 0.09 s and 73.6 AR, 50 steps gives 0.86 s and 83.5 AR, and 100 steps gives 1.72 s and 84.0 AR (Hamza et al., 23 Feb 2026). Accuracy saturation at higher $p(T \mid I_t, I_{t+1})$0 yields an explicit accuracy-efficiency trade-off. This resonates with the visual-odometry PoseFM formulation, where few ODE steps are also sufficient for practical sampling (Kuczkowski et al., 24 Apr 2026). The shared methodological theme is not identical, but both methods use learned continuous-time transport with controllable sampling depth.

6. PoseFM as pose fusion in PFAvatar and broader methodological themes

In PFAvatar, PoseFM is a clarifying term for the paper’s “Pose-Fusion mechanism,” realized across two coupled stages: ControlBooth and BoothAvatar (Xi et al., 17 Nov 2025). In the first stage, pose priors are fused with text and appearance by conditioning a diffusion UNet on pose features alongside text embeddings. In the second, pose-conditioned 3D-SDS, canonical SMPL-X space sampling, and a local geometric constraint guide NeRF optimization toward a canonical, animatable avatar (Xi et al., 17 Nov 2025).

The diffusion stage takes OOTD images $p(T \mid I_t, I_{t+1})$1, pose maps $p(T \mid I_t, I_{t+1})$2, and text captions $p(T \mid I_t, I_{t+1})$3. Pose maps are obtained using a pre-trained ControlNet pipeline and converted by a neural block $p(T \mid I_t, I_{t+1})$4 into a pose-conditioning vector $p(T \mid I_t, I_{t+1})$5, while text is encoded as $p(T \mid I_t, I_{t+1})$6 by a text encoder $p(T \mid I_t, I_{t+1})$7. The personalized diffusion model is trained with

$p(T \mid I_t, I_{t+1})$8

and the Condition Prior Preservation Loss

$p(T \mid I_t, I_{t+1})$9

The total objective is

$\mathfrak{so}(3)$0

According to the paper, CPPL mitigates both language drift and control drift in few-shot tuning by regularizing the personalized UNet on prior samples generated by the frozen base model (Xi et al., 17 Nov 2025).

The NeRF stage uses an Instant-NGP backbone and volume rendering

$\mathfrak{so}(3)$1

PoseFM here consists in using skeleton-conditioned images and canonical SMPL-X sampling for 3D-aware guidance. The 3D-SDS gradient is given as

$\mathfrak{so}(3)$2

The total BoothAvatar objective is

$\mathfrak{so}(3)$3

A local geometry loss stabilizes hands and faces by constraining NeRF densities near predefined meshes (Xi et al., 17 Nov 2025).

The reported outcomes are not pose-estimation metrics but personalization and reconstruction characteristics. The method completes personalization in approximately 5 minutes and reports a $\mathfrak{so}(3)$4 speed-up over PuzzleAvatar’s approximately 4-hour multi-asset pipeline (Xi et al., 17 Nov 2025). The paper states that PFAvatar outperforms InstantID, FreeCustom, DisenBooth, and PuzzleAvatar on identity-oriented diffusion metrics, and outperforms PuzzleAvatar, TeCH, and AvatarBooth on PuzzleIOI for NeRF reconstruction quality (Xi et al., 17 Nov 2025). Here PoseFM does not denote flow matching; rather, it denotes direct fusion of pose priors with full-body appearance and pose-conditioned NeRF guidance.

Across the three works, a common structural pattern emerges. Pose is not treated merely as a final regressed variable; it becomes a conditioning signal, a transport target, or a latent distribution. In (Kuczkowski et al., 24 Apr 2026), uncertainty is expressed through sampling in pose space. In (Hamza et al., 23 Feb 2026), symmetry and low texture are addressed by denoising points in $\mathfrak{so}(3)$5 before rigid registration. In (Xi et al., 17 Nov 2025), pose control is embedded into diffusion and 3D distillation so that avatar generation remains controllable and identity-preserving. This suggests that “PoseFM,” across its different meanings, indexes a broader methodological movement from deterministic pose prediction toward generative, conditional, and distributional formulations.

7. Limitations, misconceptions, and future directions

A common misconception would be to treat PoseFM as a single standardized framework. The cited literature does not support that interpretation. One paper uses PoseFM as the title of a visual-odometry method (Kuczkowski et al., 24 Apr 2026), one equates the term with Flose for generative 6D object pose estimation (Hamza et al., 23 Feb 2026), and one introduces PoseFM only as a clarifying label for PFAvatar’s pose-fusion design (Xi et al., 17 Nov 2025). The term therefore has contextual meaning.

Each formulation also carries distinct limitations. The visual-odometry PoseFM retains monocular scale ambiguity, relies on frame-to-frame estimation without temporal smoothing or loop closure, and can be affected by domain shift, especially for the WAFT variant on KITTI (Kuczkowski et al., 24 Apr 2026). The Flose formulation incurs latency from iterative denoising, depends on the quality of target segmentation, and currently assumes object-level masks; the authors identify single-step denoising and scene-level operation without provided masks as future directions (Hamza et al., 23 Feb 2026). PFAvatar’s authors identify two-stage training as a complexity that could be simplified with off-the-shelf descriptors, note that iterative flow matching increases inference time, and also point to scene-level operation without pre-provided masks as an important next step (Xi et al., 17 Nov 2025).

The most stable cross-paper conclusion is therefore narrower than the term’s surface similarity may suggest. PoseFM denotes several related but non-identical attempts to reframe pose-sensitive inference through generative mechanisms: conditional flow matching for relative camera motion, conditional flow matching in $\mathfrak{so}(3)$6 for rigid object alignment, and pose-fusion conditioning for personalized avatar generation. What unifies them is the replacement of brittle point estimates or isolated pose heads by learned transport, denoising, or conditioning processes that incorporate geometry, appearance, or pose priors more explicitly than conventional deterministic pipelines.