High-Fidelity Mamba Architectures

• High-Fidelity Mamba is a family of architectures that augments linear-time state-space models with domain-specific fidelity mechanisms to recover fine details.
• It has been applied across MRI reconstruction, image diffusion, 3D detection, speech editing, and video prediction to enhance performance and reduce computational costs.
• The design leverages techniques like frequency-aware dual-stream modeling and unrolled optimization to balance global context processing with local detail preservation.

High-Fidelity Mamba (HiFi-Mamba) denotes a class of Mamba- or selective state-space-based architectures whose central objective is to preserve fine-grained, high-frequency, or otherwise fidelity-critical information while retaining the linear-time, long-range modeling advantages of state-space models. In explicit usage, the term names MRI reconstruction networks that combine frequency decomposition, adaptive state-space modulation, and data-consistency-regularized unrolling (Chen et al., 7 Aug 2025, Fang et al., 23 Nov 2025). In adjacent work, the same design logic appears in latent diffusion, multimodal 3D detection, text-driven motion generation, speech editing and zero-shot TTS, video prediction, and long-context raw-audio understanding: global SSM aggregation is supplemented with local attention, spectral decoupling, geometry-aware serialization, diffusion refinement, or sparse attention bottlenecks to mitigate the information loss exhibited by pure Mamba backbones (Fu et al., 2024, Wang et al., 6 Jul 2025, Zhan et al., 10 Mar 2025, Mohammad et al., 6 Oct 2025, Xie et al., 9 Jul 2025, Khushiyant et al., 22 Mar 2026).

1. Terminology, scope, and representative systems

The label “HiFi-Mamba” is used most directly in MRI reconstruction, but the broader literature shows a convergent pattern: fidelity is improved by pairing Mamba’s global recurrent compression with mechanisms that explicitly restore or preserve local detail, spectral structure, or geometric correctness.

System	Domain	Fidelity mechanism
HiFi-Mamba	MRI reconstruction	Dual-stream WL block, adaptive state-space modulation, unidirectional traversal
HiFi-MambaV2	MRI reconstruction	SF-Lap, LSGP, hierarchical shared-routed MoE, DFSA
LaMamba-Diff	Image diffusion	VSSM SS2D plus local shifted-window attention in a U-Net
MambaFusion	Camera-LiDAR 3D detection	Height-Fidelity LiDAR Encoding and Hybrid Mamba Block
MAVE	Speech editing and zero-shot TTS	Cross-attentive Mamba with codec-token autoregression
DIFFUMA	Video prediction	Dual-path Mamba plus diffusion enhancement

In this usage, “high-fidelity” does not refer to a single metric family or a single modality. In MRI, it is tied to recovery of high-frequency anatomical detail, anatomical coherence, and PSNR/SSIM/NMSE gains (Chen et al., 7 Aug 2025, Fang et al., 23 Nov 2025). In image diffusion, it refers to retaining fine spatial detail at patch size $1$ while remaining approximately linear in sequence length (Fu et al., 2024). In camera-LiDAR fusion, it means preserving continuous height geometry so that sequence ordering for Mamba is not corrupted by discretization (Wang et al., 6 Jul 2025). In speech, it refers to perceptual naturalness, speaker consistency, and text-acoustic alignment at codec-token resolution (Mohammad et al., 6 Oct 2025).

A plausible synthesis is that HiFi-Mamba has become less a single architecture than a design family: Mamba remains the long-range backbone, but fidelity is protected by auxiliary structures targeted at the dominant information bottleneck of the application.

2. Core architectural principle

Across these works, the motivating problem is consistent. Pure self-attention captures all-pair interactions, but its quadratic complexity becomes prohibitive as token count grows. Pure Mamba or other SSM backbones offer linear complexity and global receptive fields, but compression into a recurrent hidden state or selective scan can suppress the very information needed for high-fidelity prediction or reconstruction. In image diffusion, this appears as loss of “fine-grained local information” and underperformance relative to DiT at similar scale (Fu et al., 2024). In MRI, it appears as “high-frequency insensitivity” and inadequate emphasis on sharp edges, fine trabecular structure, and small lesions (Chen et al., 7 Aug 2025). In multimodal 3D detection, naïve linear-complexity replacements degrade because height information loss leads to deviations in sequence order (Wang et al., 6 Jul 2025). In speech, a Mamba-only joint text-audio model is described as suffering from “fuzzy memory” over long text segments (Mohammad et al., 6 Oct 2025).

The underlying SSM formalism is standard. A continuous-time linear SSM can be written as

$h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$

with discretization yielding recurrent updates in linear time; Mamba then makes parameters such as $\Delta$, $B$, and $C$ input-dependent through selective scan (Fu et al., 2024). What HiFi-Mamba systems change is not this backbone principle, but the surrounding information path. The recurrent state continues to model long-range context, while a second mechanism preserves what recurrent compression alone tends to discard.

Those second mechanisms differ by domain but are structurally analogous. LaMamba-Diff inserts local shifted-window self-attention after SS2D-based VSSM blocks (Fu et al., 2024). HiFi-Mamba for MRI splits features into low- and high-frequency streams, then lets high-frequency guidance modulate SSM parameters $B$ and $C$ while leaving $\Delta$ untouched (Chen et al., 7 Aug 2025). HiFi-MambaV2 adds frequency-consistent decomposition, a lightweight global context path, and per-pixel expert routing (Fang et al., 23 Nov 2025). MAVE retains Mamba on the long audio axis but injects text via cross-attention after each Mamba block (Mohammad et al., 6 Oct 2025). HELIX retains an overwhelmingly Mamba-based stack but inserts a single attention bottleneck to recover global mixing at very long sequence lengths (Khushiyant et al., 22 Mar 2026).

This suggests a common formulation: HiFi-Mamba is a corrective architecture for the mismatch between linear-time recurrent summarization and fidelity-critical local, spectral, or cross-modal constraints.

3. MRI reconstruction as the canonical explicit formulation

The clearest explicit definition comes from MRI reconstruction. HiFi-Mamba addresses accelerated MRI under the forward model

$y = P F x + \text{noise},$

where $y$ is undersampled k-space, $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$0 the fully sampled image, $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$1 the Fourier transform, and $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$2 the sampling mask. The original HiFi-Mamba adopts an unrolled optimization architecture with $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$3 cascaded HiFi-Mamba Groups, each containing two Mamba Units and a Data Consistency block. Input is a complex undersampled image $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$4, patch-embedded and refined through a dual-stream frequency-aware pipeline (Chen et al., 7 Aug 2025).

Inside each Mamba Unit, a $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$5-Laplacian block performs “fidelity-preserving spectral decoupling.” Features are split into $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$6 and $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$7; $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$8 undergoes CRM refinement and channel-wise 2D DWT, yielding $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$9, after which

$\Delta$0

The high-frequency stream is then formed as $\Delta$1. A CRM produces anatomical guidance $\Delta$2, and the HiFi-Mamba block processes the low-frequency stream as $\Delta$3, with adaptive state-space modulation applied only to $\Delta$4 and $\Delta$5:

$\Delta$6

Depth-wise 1D convolutions with kernel size $\Delta$7 make $\Delta$8, $\Delta$9, and $B$0 spatially aware. The block then uses a streamlined unidirectional traversal rather than four-directional scanning, giving an approximately $B$1 FLOPs reduction within the SSM backbone while preserving long-range modeling (Chen et al., 7 Aug 2025).

The reported results establish the architecture as a concrete high-fidelity Mamba. On fastMRI knee at AF=8, HiFi-Mamba (P1) reaches $B$2 dB PSNR, SSIM $B$3, and NMSE $B$4; on CC359 brain at AF=8, it reaches $B$5 dB, SSIM $B$6, and NMSE $B$7. Component ablations on CC359 show that WL, HiFi-Mamba, DSFA, and CRM are complementary, with the full configuration improving from $B$8 to $B$9 in PSNR/SSIM/NMSE (Chen et al., 7 Aug 2025).

HiFi-MambaV2 advances this formulation by redefining the frequency and routing machinery inside a larger unrolled framework. It uses eight cascaded HiFi-Mamba Groups, each containing two HiFi-Mamba Units and a DC block. The new Separable Frequency-consistent Laplacian pyramid employs the normalized 5-tap binomial kernel

$C$0

reflect padding, depthwise separable horizontal and vertical filtering, and bilinear upsampling to form low- and high-frequency streams:

$C$1

A Lightweight SE-Guided Global Context Path computes

$C$2

so global context is injected before frequency splitting. The Hierarchical Shared-Routed MoE then performs per-pixel top-1 sparse dispatch with shared experts always active and routed experts selected by

$C$3

A load-balancing term $C$4 is added to the reconstruction loss (Fang et al., 23 Nov 2025).

Empirically, V2 improves on V1 and earlier Mamba baselines across fastMRI, CC359, Prostate158, ACDC, and M4Raw. On CC359 at AF=8, HiFi-MambaV2(P1) reports $C$5 dB PSNR, SSIM $C$6, and NMSE $C$7, compared with $C$8 for LMO and $C$9 for HiFi-MambaV1(P1). On ACDC at AF=8 with radial masks, V2(P1) reports $B$0. The paper explicitly frames these gains as improved high-frequency detail recovery together with preserved anatomical coherence (Fang et al., 23 Nov 2025).

4. Cross-domain extensions of the HiFi-Mamba pattern

In image generation, LaMamba-Diff functions as a direct precursor to the broader HiFi-Mamba idea. The model introduces Local Attentional Mamba blocks that combine VSSM with SS2D scanning and local window self-attention in fixed-size windows, using a Swin-style shifted window scheme, AdaLN conditioning, zero-initialized residual scaling, and a U-Net latent diffusion backbone. All components remain linear in sequence length $B$1, enabling patch size $B$2 in latent space. The ablation most relevant to HiFi-Mamba is that removing local attention degrades FID from $B$3 to $B$4, while replacing local with global attention improves FID only slightly, from $B$5 to $B$6, at $B$7 higher FLOPs. The paper’s own characterization is that LaMamba-Diff is “exactly what one would expect from a ‘HiFi-Mamba’ block” (Fu et al., 2024).

In camera-LiDAR 3D detection, MambaFusion reinterprets fidelity as geometric correctness, especially along the height axis. Height-Fidelity LiDAR Encoding replaces discrete voxel centers with continuous centers computed by $B$8, propagates these continuously through downsampling, and adds a conflict test for generated voxels. Features are serialized with a 3D Hilbert curve and processed by a Hybrid Mamba Block that stacks Local Mamba and Global Mamba. Quantitatively, the full system reaches $B$9 NDS on nuScenes validation and $C$0 on test, while ablations show that Mamba fusion in raw space without HFL hurts performance and that Hilbert ordering outperforms Morton and simple coordinate order (Wang et al., 6 Jul 2025).

In text-driven motion generation, HiSTF Mamba implements a fidelity-oriented decomposition of body and time. Dual-Spatial Mamba runs part-based and whole-based branches in parallel over six body parts, Bi-Temporal Mamba scans forward and backward through time, and the Dynamic Spatiotemporal Fusion Module removes redundant temporal information before fusing it with spatial features. On HumanML3D, HiSTF Mamba with 10 sampling steps reaches FID $C$1, compared with $C$2 for Motion Mamba, while preserving competitive R-precision and MM Dist; ablations show that removing the part-based branch raises FID to $C$3, and removing DSFM raises it to $C$4 (Zhan et al., 10 Mar 2025).

In speech, MAVE uses a cross-attentive Mamba backbone for codec-token autoregression. Audio is tokenized by X-Codec at $C$5 Hz with $C$6 codebooks of size $C$7, text is phonemized and encoded by a 4-layer Transformer, and a 12-layer Mamba decoder applies cross-attention to text after each Mamba block. Editing is trained via CM3-style causal masking and sequence rearrangement; zero-shot TTS emerges by prepending a reference utterance. On RealEdit, $C$8 of listeners rate MAVE-edited speech as perceptually equal to the original, with $C$9 preferring the original and $\Delta$0 preferring MAVE; on LibriTTS zero-shot TTS, MAVE improves over VoiceCraft in WER, SIM, and MOS naturalness (Mohammad et al., 6 Oct 2025).

In video prediction, DIFFUMA makes the Mamba/detail split explicit. A Mamba module produces a coarse but temporally coherent prediction, while a diffusion module, conditioned on temporal context from Mamba through AdaLN, predicts a detail residual $\Delta$1 so that

$\Delta$2

On the CHDL semiconductor wafer dicing benchmark, the abstract reports a $\Delta$3 reduction in MSE and SSIM improvement from $\Delta$4 to $\Delta$5; the main table reports CHDL performance improving over SimVP from MSE $\Delta$6, MAE $\Delta$7, SSIM $\Delta$8 to MSE $\Delta$9, MAE $y = P F x + \text{noise},$0, SSIM $y = P F x + \text{noise},$1 (Xie et al., 9 Jul 2025).

In long-context raw-audio understanding, HELIX provides an austere hybridization result: five Bidirectional Mamba blocks and one mid-stack attention layer at fixed total parameter count of about $y = P F x + \text{noise},$2M. Pure attention fails with out-of-memory errors at $y = P F x + \text{noise},$3 tokens, pure Mamba reaches $y = P F x + \text{noise},$4 on VoxPopuli 300 s speaker ID, and HELIX reaches $y = P F x + \text{noise},$5, closing an $y = P F x + \text{noise},$6-point gap over pure Mamba. This suggests that one sparse global mixing layer can materially improve fidelity of long-range audio representations without abandoning Mamba’s scaling regime (Khushiyant et al., 22 Mar 2026).

5. Computational profile and efficiency rationale

A central feature of HiFi-Mamba systems is that fidelity restoration is usually engineered without reverting to full quadratic attention everywhere. LaMamba-Diff states this most explicitly: SS2D is $y = P F x + \text{noise},$7, windowed self-attention is $y = P F x + \text{noise},$8 for fixed window size $y = P F x + \text{noise},$9, and the FFN is $y$0, whereas global MSA would add the quadratic term $y$1. In practice, LaMamba-Diff-XL reports $y$2 GFLOPs at ImageNet $y$3 against $y$4 for DiT-XL/2, and $y$5 GFLOPs at $y$6 against $y$7, despite using $y$8 longer sequences (Fu et al., 2024).

The MRI lineage achieves analogous savings by targeting redundancy inside the SSM itself. HiFi-Mamba replaces four-directional scanning with a streamlined unidirectional traversal, and on fastMRI AF=8 at $y$9 reports $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$00G FLOPs for HiFi-Mamba (P1) and $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$01G for HiFi-Mamba (P2), compared with $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$02G for ReconFormer, $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$03G for FpsFormer, and $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$04G for the four-directional Mamba baseline LMO. HiFi-Mamba (P2) therefore attains similar or better accuracy than Transformer baselines with roughly $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$05–$h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$06 fewer FLOPs and better accuracy than LMO at much lower cost (Chen et al., 7 Aug 2025).

In speech, MAVE formalizes the asymptotic advantage of cross-attentive Mamba over decoder-only Transformers. For audio length $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$07 and text length $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$08, the Mamba decoder contributes $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$09, whereas decoder-only Transformer inference with KV caching scales as $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$10, which is effectively quadratic in $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$11 when $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$12. On RealEdit, MAVE reports average and max GPU memory of $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$13 GB, compared with $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$14 GB for VoiceCraft with KV cache, while maintaining similar wall-clock time (Mohammad et al., 6 Oct 2025).

HELIX shows the same principle in a controlled classification setting. The hybrid pays the $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$15 cost only once, with the rest of the stack remaining linear in sequence length. This is enough to keep $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$16-token training feasible on a single RTX 6000 Pro, whereas the 6-layer pure attention model OOMs. A plausible implication is that HiFi-Mamba systems often scale by making fidelity corrections sparse, localized, or conditional rather than globally attention-dense (Khushiyant et al., 22 Mar 2026).

6. Limitations, misconceptions, and open directions

One common misconception is that replacing attention with Mamba is by itself sufficient for high fidelity. Multiple papers argue the opposite. LaMamba-Diff shows that Mamba alone is not sufficient for HiFi image generation because removing local attention sharply worsens FID (Fu et al., 2024). MambaFusion shows that naïve linear-complexity replacements can degrade performance unless geometry and sequence design are corrected first (Wang et al., 6 Jul 2025). HiFi-Mamba for MRI similarly argues that high-frequency anatomy is not recovered by generic vision Mamba variants without explicit spectral handling and guided modulation (Chen et al., 7 Aug 2025).

Another misconception is that “high fidelity” is reducible to local sharpness alone. The MRI papers define it jointly in terms of high-frequency detail, anatomical coherence, and numerical fidelity; HiFi-MambaV2 emphasizes that a pure Mamba block is low-pass, a Conv block is overly cross-shaped in frequency response, and the HiFi-Mamba block seeks a balanced spectrum with dominant low frequencies while retaining moderate high-frequency energy (Fang et al., 23 Nov 2025). MAVE likewise ties fidelity not only to naturalness but to speaker consistency and text-conditioned editing accuracy (Mohammad et al., 6 Oct 2025).

The limits reported by the papers are domain-specific but structurally similar. HiFi-Mamba for MRI is evaluated in single-coil, equispaced 1D Cartesian sampling, and broader validation under multi-coil, non-Cartesian, or variable-density sampling is left open (Chen et al., 7 Aug 2025). HiFi-MambaV2 improves fidelity but increases complexity, with the higher-capacity P1 variant reaching $h'(t)=Ah(t)+Bx(t), \qquad y(t)=Ch(t)+\bar{D}x(t),$17G FLOPs (Fang et al., 23 Nov 2025). MambaFusion still depends on handcrafted Hilbert ordering and fixed voxel sizes, and extension to temporal perception is left as future work (Wang et al., 6 Jul 2025). MAVE degrades on very long target transcripts and raises explicit deepfake concerns (Mohammad et al., 6 Oct 2025). HELIX evaluates only one hybrid attention placement and focuses on classification rather than generation (Khushiyant et al., 22 Mar 2026). DIFFUMA’s dual-path architecture is heavier than a standalone Mamba predictor, and the paper itself points toward further efficiency optimizations such as latent-space diffusion (Xie et al., 9 Jul 2025).

Taken together, these works indicate that HiFi-Mamba is best understood not as “Mamba with more capacity,” but as a recurrent design principle: preserve the linear-time, long-context strength of selective SSMs, then add the smallest domain-appropriate mechanism that restores the information those SSMs systematically attenuate. In MRI that mechanism is frequency-aware dual-stream modeling with data consistency; in diffusion it is local attention; in 3D detection it is height-fidelity geometry and Hilbert serialization; in speech it is cross-attention to phoneme embeddings; in video it is diffusion-style detail enhancement; and in long audio it is a sparse attention bottleneck.