Virtual-Width Dynamic Vision Encoder

• Virtual-Width DVE is a neural architecture that dynamically expands virtual channel width using context-dependent Fourier bases to improve feature discrimination.
• It employs a Frequency Dynamic Linear (FDLinear) operator that adaptively assembles weights based on a global descriptor, enhancing linear separability in complex visual tasks.
• The architecture integrates into multi-stage pipelines in medical imaging, optimizing diagnostic performance with minimal parameter overhead and efficient computation.

A Virtual-Width Dynamic Vision Encoder (DVE) is a neural architecture class designed to achieve high geometric capacity and efficient feature discrimination in visual encoding, without incurring the computational and parameter expansion typically associated with physically wider models. The DVE concept is central to recent advancements in specialized medical vision-language systems, particularly for tasks like dermatological diagnosis that demand precise separation of subtle pathological structures from noisy backgrounds (Liu et al., 14 Jan 2026). DVEs leverage dynamic, sample-conditioned transformations, virtual channel width expansion, and adaptive computation to improve both the efficiency and expressiveness of visual representations.

1. Conceptual Foundations and Motivation

The Virtual-Width DVE paradigm originated as a response to the “retina–brain” asymmetry in multimodal vision-LLMs—where the vision backbone is much narrower than the paired LLM—and the resulting geometric “Capacity Collapse.” In tasks characterized by an unbounded diversity of target manifolds (e.g., pathological skin textures), static encoders with fixed channel widths $d$ cannot linearly separate a large number of features: by Cover’s theorem, $P(N, d) \approx 0$ when $N \gg 2d$, causing essential diagnostic cues to be averaged out and irretrievably lost.

DVEs address this by virtually expanding the channel width to $K \cdot d$ via a small set of orthogonal, frequency-disjoint bases and dynamic, context-dependent weighting. This virtual expansion enables the encoder to "unfold" complex data manifolds in the embedding space, enhancing the probability of linear separability and making explicit and implicit features accessible to downstream decision modules. Unlike approaches that rely on brute-force parameter scaling, DVE methods maintain O($d^2$) compute per token, with only marginal increases in learned parameters and negligible impact on runtime (Liu et al., 14 Jan 2026).

2. Architectural Innovations

The DVE architecture in SkinFlow (Liu et al., 14 Jan 2026) implements its virtual width with a Frequency Dynamic Linear (FDLinear) operator, which substitutes for static Linear layers in select MLP blocks of the Vision Transformer (ViT) backbone (notably at layers 8, 16, 24, 32). Each FDLinear stores $K$ Fourier-derived basis matrices $\{B_1, ..., B_K\} \in \mathbb{R}^{d \times d}$.

For any input sample, a global descriptor $\bar{x}$ (e.g., channelwise mean) is computed. A lightweight, bottlenecked fully connected (FC) network predicts $K$ scalar coefficients $\{\alpha_k(\bar{x})\}$, forming a sample-adaptive mixture:

$W(\bar{x}) = \sum_{k=1}^K \alpha_k(\bar{x}) B_k$

Input tokens $x$ are then projected via $y = W(\bar{x})x$, enabling the weight matrix orientation to change on a per-sample basis and adapt to the local data geometry. This contrasts fundamentally with static projections ($y = W_\text{fixed}x$).

There is no need to explicitly compute all $K$ projections $(B_k x)$; by aggregating the weighted bases before applying to $x$, the method retains standard computational complexity. Parameter overhead is only in storing the $K$ basis matrices and the compact coefficient predictor (typically $<5\%$ increase).

3. Mathematical Formalism

The key principles are captured by the following equations and constructs:

• Cover’s theorem on linear separability:

$$P(N, d) \approx \begin{cases} 1, & N \leq 2d \ 0, & N \gg 2d \end{cases}$$

• Virtual width construction:
  • "Explicit" projection to all bases:

  $$\mathcal{H} = \mathrm{Concat}(B_1x, \dots, B_Kx) \in \mathbb{R}^{K \times d}$$ - "Implicit" dynamic aggregation:

  $$y = \left(\sum_{k=1}^K \alpha_k B_k\right)x \in \mathbb{R}^d$$

• Complexity:

  • Standard layer: O($d^2$) per token.
  • Virtual-Width FDLinear: O($d^2$) per token (no dependence on $K$ in runtime, only memory).
• Typical hyperparameters: $K = d/2$ (e.g., for $d=1280$, $K=640$ yields virtual width $K \cdot d \approx 81,920$), coefficient predictor with FC bottleneck $d \rightarrow d/4 \rightarrow K$, and FDLinear inserted at 4 points in the vision stack.

4. Implementation and Training Strategies

Implementation proceeds as follows:

• Initialization: Group Discrete Fourier Transform (DFT) components of a pretrained static weight matrix, with each basis masking disjoint frequency bands.
• Forward pass (see pseudocode in (Liu et al., 14 Jan 2026)):
  • Compute global average $\bar{x}$ across tokens.
  • Use a compact FC network to predict $K$ coefficients $\alpha_k(\bar{x})$.
  • Assemble $W_\text{dyn} = \sum_k \alpha_k B_k$.
  • Multiply $W_\text{dyn}$ with input tokens.
• Regularization: Apply a small $\ell_2$ penalty on $\alpha_k$ to avoid mode collapse.
• Optimization: Employ layer-wise learning rate warmup for dynamic modules to stabilize convergence.

5. Integration into Multistage Pipelines

In the SkinFlow pipeline (Liu et al., 14 Jan 2026), DVE modules are used in both:

• Stage I (Medical Caption Learning): The DVE’s virtual geometric capacity allows the encoder to compress explicit, describable visual features (e.g., color, scale, explicit lesion boundaries) into rich, linguistically-aligned embeddings. This facilitates more informative and precise medical report generation.
• Stage II (Diagnostic RL): The same DVE-augmented encoder supplies discriminative representations of implicit textures for a diagnostic RL policy (Generalized Reward Policy Optimization, GRPO). The improved linear separability accelerates convergence and enhances diagnostic reward, as the policy learns atop an "unfolded" vision manifold.

Downstream language decoders (e.g., LLMs) attend to these DVE-enhanced features. Empirical attention map analysis shows that the model’s high-confidence attention bins align closely with pathologically relevant regions (lesions), and background noise is substantially reduced.

6. Comparative Approaches and Generalizations

While the FDLinear-based DVE relies on channel-wise virtual expansion and dynamic projection, broader DVE concepts include:

• Patch-based virtual width: As formalized by Prisadnikov et al., virtual width $w_v = M+N$ is defined in terms of a fixed number of glimpse patches ($M$) per iteration and memory tokens ($N$), independent of input image size. The total budget is determined by the number of iterations $K$, which can be dynamically controlled by a gating criterion reflecting task difficulty or model confidence (Prisadnikov et al., 22 Aug 2025). This mechanism decouples per-step cost from input image resolution and matches biological strategies for visual attention.
• Token count adaptivity: DOVE (Dynamic Output Vision Encoder) generates a variable-length sequence of latent visual tokens per input, terminating early if sufficient semantic information has been extracted (Mao et al., 4 Jun 2025). While not a virtual-width mechanism in the FDLinear sense, DOVE likewise adapts encoder expressivity to image complexity or query demands.
• Task-driven computation: The DVE paradigm also covers adaptive early exit, patch selection, and multi-zoom cropping, thereby flexibly adjusting the representation complexity to the specific needs of the downstream task, rather than fixed architecture constraints.

Model	Virtual Width Mechanism	Backbone Modality	Dynamicity Level
SkinFlow DVE	FDLinear, channel expansion	ViT, MLP	Per-sample weights
Prisadnikov et al. DVE	Patch/iteration, glimpse-count	ViT-like, foveated	Task-driven early exit
DOVE	Variable output token length	VQGAN + Transformer	Query/image complexity

7. Empirical Outcomes and Impact

The introduction of DVE mechanisms has yielded substantial empirical improvements:

• SkinFlow (DVE with FDLinear) (Liu et al., 14 Jan 2026):
  • Fitzpatrick17k: Top-1 accuracy increased from 24.45% to 29.19%; Top-6 from 57.69% to 71.16%, surpassing much larger general-purpose models (e.g., Qwen3VL-235B, GPT-5.2).
  • Internal dermatology dataset: Top-1 from 35.64% to 36.63%; Top-6 from 74.75% to 79.21%.
• Model ablation experiments demonstrate that DVE is the predominant contributor to gains, particularly top-rank generalization on challenging datasets.
• Visualization: On synthetic benchmarks (Spirals, XOR, Moons, Circles), FDLinear enables dynamic projection field rotations, enabling non-linear separation unattainable by static layers.
• Attention analysis: A pronounced rightward shift in attention weight distributions on lesion regions, reflecting increased signal-to-noise and reduced background distraction.
• Computational efficiency: The virtual-width strategy delivers orders-of-magnitude savings in self-attention cost compared to traditional ViTs, with only modest accuracy reductions (or even accuracy gains when measured per FLOP) (Prisadnikov et al., 22 Aug 2025).

A plausible implication is that DVE-style architectures may generalize to other medical and scientific imaging domains where discriminative capacity and efficiency must coexist, particularly under severe annotation scarcity or class imbalance.

---

The Virtual-Width Dynamic Vision Encoder class encapsulates a family of methods for virtually expanding neural geometric capacity via dynamic sample-conditioned compositionality, without the physical and computational burdens of explicit channel width expansion. This architecture—manifested in specialized medical reasoning models, dynamic glimpse-based classification, and adaptive tokenization—demonstrates that optimizing information flow and adaptive computation provides superior diagnostic reasoning over raw parameter scaling, both theoretically and empirically (Liu et al., 14 Jan 2026, Prisadnikov et al., 22 Aug 2025, Mao et al., 4 Jun 2025).