Any2Full: Universal Partial-to-Full Mapping

• Any2Full is a framework that maps arbitrary, sparse inputs (e.g., depth, medical imaging, and feature maps) into fully dense, complete representations.
• It employs a prompt-based, one-stage inference mechanism where pretrained models and scale-aware prompting convert partial evidence to a full output.
• The paradigm unifies diverse modalities—from depth completion to multimodal imaging and collaborative perception—via standardized interface engineering.

Searching arXiv for papers on "Any2Full" and closely related formulations. Any2Full denotes a problem formulation in which incomplete, sparse, heterogeneous, or low-resolution observations are mapped to a fuller target representation. In explicit usage, "Any to Full: Prompting Depth Anything for Depth Completion in One Stage" defines Any2Full as a one-stage, domain-general, pattern-agnostic depth completion framework that converts sparse depth of varying patterns into dense metric depth by prompting a pretrained monocular depth estimation model (Zhou et al., 5 Mar 2026). Closely aligned formulations extend the same partial-to-full logic to complete multimodal medical image stacks through "Any2all" (Gan et al., 9 Feb 2026), to zero-shot feature-modality translation in heterogeneous collaborative perception through UniTrans (Li et al., 18 May 2026), and to encoder-agnostic dense feature upsampling through AnyUp (Wimmer et al., 14 Oct 2025). Taken together, these works suggest that Any2Full is best understood as a broader design paradigm: a single reusable model, prior, or translator is conditioned at inference time to recover a more complete representation from arbitrary partial inputs.

1. Definition and scope

In its strictest sense, Any2Full refers to the depth completion setting formalized as

$$\hat{\mathbf{D}}_f = \mathcal{F}_{\theta}(\mathbf{I}, \mathbf{D}_s),$$

where an RGB image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ and sparse or incomplete depth $\mathbf{D}_s$ are converted into dense full depth $\mathbf{D}_f \in \mathbb{R}^{H \times W}$. The defining claim of this formulation is not merely dense prediction, but domain generality and pattern agnosticism: robustness to unseen visual domains and to varying depth input patterns, including random sparsity, holes, LiDAR scan lines, SfM points, range truncation, and mixed defects (Zhou et al., 5 Mar 2026).

A broader reading, suggested by adjacent literature, treats Any2Full as a family of mappings from arbitrary partial evidence to a richer target state. In multimodal medical imaging, the target is a full clean multimodal stack $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$, while the available input may be any subset of clean modalities and/or measurement-space observations (Gan et al., 9 Feb 2026). In heterogeneous collaborative perception, the target is an ego-consistent feature representation in the desired modality space, obtained from arbitrary source-target modality pairs without retraining (Li et al., 18 May 2026). In feature upsampling, the target is a full-resolution dense feature grid $\hat{F}$ that remains in the source feature space, even though the input is a coarse feature map extracted by an arbitrary encoder (Wimmer et al., 14 Oct 2025).

This suggests a useful conceptual distinction. The “Any” component refers to heterogeneity of conditioning: arbitrary sparsity patterns, missing modalities, measurement operators, unseen feature modalities, or variable source resolutions. The “Full” component refers to a completed latent or output state: dense metric depth, the full multimodal image stack, a target-modality feature tensor suitable for fusion, or a dense high-resolution feature map. The paradigm is therefore less about a specific modality than about inference under partial observability with a reusable full-target prior.

2. Formal problem formulations

The explicit Any2Full depth-completion formulation is based on prompting a pretrained MDE model rather than learning RGB-depth fusion from scratch. Sparse metric depth is first converted to the relative-depth or disparity domain and normalized as

$\tilde{d}_i = \frac{d_i - \mu(d)}{\sigma(d)}.$

A Scale-aware Prompt Encoder $\mathcal{G}$ then produces scale prompts that condition the pretrained MDE model $\mathcal{M}$: $$\hat{\tilde{\mathbf{D}}}_f = \mathcal{M}\big(\mathbf{I} \mid \mathcal{G}(\tilde{\mathbf{D}}_s)\big).$$ The final dense metric depth is recovered by non-parametric least-squares alignment to the raw sparse anchors (Zhou et al., 5 Mar 2026). The critical modeling claim is that sparse depth should provide scale cues, while dense geometry should remain under the control of the pretrained MDE prior.

The multimodal medical-imaging formulation is more explicitly generative. The clean target is the full multimodal stack

$\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$0

and all available information is encoded by a linear observation model

$\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$1

Here $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$2 denotes a clean observed modality, $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$3 a missing modality, and a nontrivial $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$4 a measurement-space forward model such as undersampled MRI acquisition. This lets reconstruction and synthesis be treated as a single virtual inpainting problem (Gan et al., 9 Feb 2026).

UniTrans formulates heterogeneous collaborative perception at the intermediate-feature level. Each agent produces $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$5 with modality type $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$6, and a neighbor feature is translated into ego modality space as

$\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$7

The translator parameters are not trained per pair; instead, they are instantiated on the fly from source and target modality-intrinsic codes. With

$\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$8

the final translator is synthesized by

$\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$9

This gives a zero-shot any-to-any mapping mechanism for unseen modality pairs (Li et al., 18 May 2026).

AnyUp addresses densification rather than modal translation. Given a coarse feature map $\mathbf{D}_s$0 and a high-resolution RGB image $\mathbf{D}_s$1, it realizes

$\mathbf{D}_s$2

where each output vector is described as a learned weighted sum of a local set of coarse feature vectors, guided by image content and feature-informed local attention. The paper’s core canonicalization layer maps arbitrary input channels into a fixed-size structural code through per-channel basis convolutions and channel-averaged softmax responses: $\mathbf{D}_s$3 This removes dependence on encoder-specific channel dimensionality and semantics (Wimmer et al., 14 Oct 2025).

Across these formulations, the common structure is that the target representation is modeled jointly or canonically, while the conditioning pattern is deferred to prompting, operator constraints, routing in intrinsic space, or image-guided attention. That is the central mathematical signature of Any2Full-style systems.

3. Architectural patterns

The explicit Any2Full architecture is centered on a frozen Depth Anything v2 backbone plus a learned Scale-Aware Prompt Encoder (SAPE). SAPE receives normalized sparse depth through a mask-aware sparse depth embedding and is divided into Local Enrichment and Global Propagation. Local Enrichment couples patch-level sparse-depth features $\mathbf{D}_s$4 with corresponding MDE features $\mathbf{D}_s$5 through generalized FiLM: $\mathbf{D}_s$6 Global Propagation then updates $\mathbf{D}_s$7 with geometry-guided Transformer blocks whose Query and Key come only from MDE geometry features and whose Value comes from $\mathbf{D}_s$8. The resulting multi-level prompts are injected into decoder features by hierarchical FiLM modulation: $\mathbf{D}_s$9 The stated objective is to preserve the backbone’s domain-general geometric prior while steering it toward globally scale-consistent predictions (Zhou et al., 5 Mar 2026).

The Any2all architecture replaces prompting with an unconditional diffusion prior over complete multimodal stacks. Training uses a standard DDPM objective on fully observed stacks, with no task-specific masks or synthesis heads. Conditioning happens only at inference through samplers that enforce $\mathbf{D}_f \in \mathbb{R}^{H \times W}$0. Multimodal Posterior Sampling applies a data-consistency gradient

$\mathbf{D}_f \in \mathbb{R}^{H \times W}$1

during reverse diffusion, while Multimodal Decomposition Sampling solves

$\mathbf{D}_f \in \mathbb{R}^{H \times W}$2

before a DDIM-style recombination step. This yields soft preservation in MPS and potentially direct replacement of known clean slices in MDS (Gan et al., 9 Feb 2026).

UniTrans adopts a different architectural answer to universality: it factorizes the problem into modality identification and translator synthesis. A Modality-Intrinsic Encoder organizes a latent space $\mathbf{D}_f \in \mathbb{R}^{H \times W}$3 in which each modality occupies a compact region, using first-order channel statistics, second-order Gram descriptors, and global response features. A Modality Mapping Router predicts expert-mixture coefficients from source-target intrinsic geometry, and a Translator Parameter Bank stores reusable full translator parameter sets. The decisive technical choice is parameter mixing rather than output mixing: the model instantiates one translator and runs it once (Li et al., 18 May 2026).

AnyUp is architecturally lighter but conceptually parallel. It does not learn an encoder-specific decoder; instead, it converts arbitrary source features into a canonical structural code with the feature-agnostic layer, then uses image-guided local window attention to redistribute original low-resolution feature vectors into dense pixel positions. Because the attention values are the original, unprocessed low-resolution features, the upsampled output remains in the source feature space (Wimmer et al., 14 Oct 2025).

These architectures differ in substrate—prompted MDE, unconditional diffusion, mapping-conditioned translator synthesis, and image-guided attention—but converge on the same systems principle: universality is achieved by standardizing the interface between arbitrary inputs and a full target representation.

4. Representative systems and empirical behavior

The most explicit benchmark evidence for Any2Full comes from depth completion. Any2Full reports that it outperforms OMNI-DC by $\mathbf{D}_f \in \mathbb{R}^{H \times W}$4 in average AbsREL and is $\mathbf{D}_f \in \mathbb{R}^{H \times W}$5 faster than PriorDA with the same Depth Anything-L backbone (Zhou et al., 5 Mar 2026). In the main benchmark table, Any2Full(DA-L) obtains the best average rank, $\mathbf{D}_f \in \mathbb{R}^{H \times W}$6, and reports average results of 0.010 / 0.459 for Hole, 0.046 / 2.184 for Range, and 0.026 / 0.829 for Sparse. Representative dataset-pattern results include NYU Hole 0.008 / 0.067, iBims Range 0.030 / 0.334, KITTI Sparse-LiDAR 0.017 / 1.356, ETH3D Sparse-SfM 0.063 / 1.129, and Logistic-Black Mixed 0.010 / 0.029. The runtime table on IBims-1 reports 0.49 s for Any2Full(DA-L), compared with 0.68 s for PriorDA, 92.40 s for TestPromptDC, and 112.47 s for Marigold-DC.

In multimodal medical imaging, Any2all is validated on an in-house PET/MR/CT brain dataset with five modalities—T1 MRI, T2FLAIR MRI, T2* MRI, CT, and PET ($\mathbf{D}_f \in \mathbb{R}^{H \times W}$7F-Florbetapir)—using 226 participants, split as 198 train, 7 validation, and 21 test (Gan et al., 9 Feb 2026). The paper reports that performance improves as more modalities are provided; MPS tends to give better FID and perceptual quality; MDS tends to give better PSNR/SSIM in reconstruction; and the unified model yields competitive distortion-based performance with superior perceptual quality over specialized methods. It also notes that for unconditional generation, Any2all obtains lower FID than mmGAN, mmResViT, and M2DN across modalities. In T2FLAIR reconstruction, auxiliary modalities improve quality markedly, and mixed clean-plus-measurement inputs are uniquely supported.

UniTrans reports strong zero-shot any-to-any translation results in heterogeneous collaborative perception. On OPV2V-H, the average AP over ego modalities is 0.716 / 0.605, compared with 0.662 / 0.538 for NegoCollab and 0.653 / 0.544 for Classic MoE (Li et al., 18 May 2026). On DAIR-V2X, the corresponding average is 0.553 / 0.421, again ahead of prior baselines. In the open-world zero-shot setting over held-out emerging modalities, UniTrans reports 0.716 / 0.605 on OPV2V-H and 0.553 / 0.421 on DAIR-V2X, whereas native baseline generalization is substantially worse. Efficiency profiling gives 109.3 GFLOPs, 6.865 ms CPU, 53.760 ms CUDA for UniTrans, compared with 245.5 GFLOPs, 89.078 ms CPU, 141.352 ms CUDA for Classic MoE.

AnyUp reports generalization across feature families and spatial scales. A model trained on DINOv2 ViT-S generalizes to SigLIP 2 and DINOv3, with the paper highlighting, for example, $\mathbf{D}_f \in \mathbb{R}^{H \times W}$8 mIoU and $\mathbf{D}_f \in \mathbb{R}^{H \times W}$9 depth RMSE on SigLIP 2 in the LoftUp setup, $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$0 mIoU on SigLIP 2 in the JAFAR setup, and essentially identical DINOv3 performance for DINOv2-trained and DINOv3-trained AnyUp, $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$1 versus $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$2 mIoU (Wimmer et al., 14 Oct 2025). On semantic segmentation, it reaches $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$3 mIoU / $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$4 accuracy on COCO and $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$5 on ADE20k. On depth estimation it reports the best absolute RMSE $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$6 and $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$7, and the best relative RMSE $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$8 and $\bm{x}=[\bm{x}_1,\dots,\bm{x}_n]^T$9. In arbitrary-resolution evaluation, it is best for $\hat{F}$0 and $\hat{F}$1 segmentation, and best across all three tested resolution changes for depth.

The empirical record therefore supports a recurring pattern: universality need not imply uniformly dominant distortion metrics, but these systems are consistently competitive while expanding the admissible conditioning space.

5. Generalization claims, and what they do not mean

A recurring misconception is that Any2Full implies training from arbitrary partial data. The surveyed formulations do not support that interpretation. Any2all is flexible at inference, but training requires complete multimodal stacks that are co-registered, normalized, and jointly available per subject (Gan et al., 9 Feb 2026). UniTrans is zero-shot at deployment for emerging modality pairs, but only after Stage 1 pretraining of the Modality-Intrinsic Encoder and Stage 2 training of the Modality Mapping Router and Translator Parameter Bank on training-visible modalities (Li et al., 18 May 2026). AnyUp is feature-agnostic at inference, yet it is still a trained module, optimized on ImageNet for 100,000 steps using AdamW, learning rate $\hat{F}$2, batch size 4, and 4 random local crops per image (Wimmer et al., 14 Oct 2025). Any2Full depth completion freezes the Depth Anything v2 backbone, but it still trains the prompt-related modules on 85K synthetic RGB-D images for 224K steps with Adam, learning rate $\hat{F}$3, batch size 16 (Zhou et al., 5 Mar 2026).

A second misconception is that universality eliminates explicit inductive bias. In fact, each system derives its generalization from a carefully chosen invariant interface. Any2Full depth completion discards output-level alignment and learns scale prompting in the relative domain. Any2all standardizes all conditioning into the operator form $\hat{F}$4. UniTrans routes through a modality-intrinsic latent space rather than raw feature tensors. AnyUp removes encoder dependence through per-channel basis responses and preserves source semantics by using original low-resolution features as attention values. This suggests that Any2Full-style generalization is not “model agnosticism” in a naive sense; it is interface engineering around the right latent variables.

A third misconception is that “full” always means exact recovery of all missing information. The reviewed papers are more precise. Any2Full depth completion aims for globally scale-consistent dense depth, then applies closed-form least-squares alignment to sparse metric anchors (Zhou et al., 5 Mar 2026). Any2all infers a full clean multimodal stack, but target-wise quantitative reporting dominates its evaluation and the paper does not provide one explicit benchmark for recovering all five modalities simultaneously from arbitrary subsets (Gan et al., 9 Feb 2026). UniTrans instantiates translators in a learned compositional sense and explicitly notes no guarantee for arbitrarily extreme unseen modalities (Li et al., 18 May 2026). AnyUp assumes that dense features can be formed as local linear combinations of coarse features and therefore does not claim unconstrained synthesis of entirely new latent detail (Wimmer et al., 14 Oct 2025).

6. Limitations, scope conditions, and likely future directions

The explicit Any2Full depth-completion framework inherits the strengths and weaknesses of pretrained MDE priors. Its advantages depend on the quality of Depth Anything v2, and the paper notes that the relative-SSIM term can slightly hurt zero-shot KITTI performance (Zhou et al., 5 Mar 2026). The method is not always numerically first in every single benchmark cell, even though it is strongest in cross-pattern consistency. The discussion points toward integration with Depth Anything v3, newer metric-depth models, and possible extension to arbitrary-view depth completion.

Any2all demonstrates a particularly strong Any2Full interpretation, but its scope is limited to medical imaging, specifically PET/MR/CT brain data (Gan et al., 9 Feb 2026). It requires complete paired multimodal stacks during training, uses iterative diffusion inference with MPS at 1000 steps and MDS at 150 steps, and assumes a linear block-diagonal observation model $\hat{F}$5. These are powerful abstractions, but they remain restrictive. The paper also notes that quantitative evaluation is target-wise rather than explicit full-stack completion.

UniTrans establishes universal any-to-any translation under open-world modality emergence, but several failure modes remain explicit. Camera-target settings $\hat{F}$6 and $\hat{F}$7 degrade for all methods because of the severe LiDAR-camera gap and limited camera field of view (Li et al., 18 May 2026). Performance depends on the learned intrinsic space, degrades with overly large latent dimension or translator-bank size, is affected by localization noise, and remains vulnerable to distribution shifts. Its universality is therefore empirical and compositional, not a theoretical guarantee.

AnyUp is highly practical, but its own limitations are clear. The paper explicitly notes that representing dense features as linear combinations of low-resolution input features may be fundamentally limiting, because patch features can encode subpatch information in complex distributed ways that are not fully recoverable by local linear mixing (Wimmer et al., 14 Oct 2025). It also implies sensitivity when RGB edges do not align with semantic or geometric feature transitions, and it does not evaluate combinations with model-specific feature-denoising methods such as FeatSharp.

A plausible implication is that future Any2Full systems will combine the strengths of these four lines: foundation-model priors, explicit operator conditioning, latent-space routing for heterogeneous inputs, and feature-space-preserving densification. The present literature already shows that the central challenge is not merely predicting missing values; it is defining a conditioning interface through which arbitrary partial evidence can steer a reusable model toward a full, task-compatible representation without pair-specific retraining.