Flash: Multi-Concepts in Imaging, Memory, & AI

• Flash is a multifaceted concept covering active illumination in imaging, non-volatile flash memory, media runtime environments, and specialized AI model architectures.
• In computational photography, flash techniques improve reflection removal and denoising, achieving performance gains such as over 5 dB PSNR improvement in reflection-free transmission.
• In memory and system architectures, flash drives ultrafast write/erase operations (~20 ns), supports in-flash computation, and enhances caching and processing through hardware-accelerated mapping.

In current technical literature, flash denotes several distinct concepts rather than a single technology. It refers to active illumination in computational photography and vision, non-volatile flash memory and its coding and system software, Flash Player as a browser-resident media runtime, and multiple unrelated acronyms in machine learning and robotics, including Few-shot Learning to Animate and Steer Humans, Fourier-Guided Latent Shifting I2V, Fast Latent-Aware Semi-Autoregressive Heuristics, and Fast Legendre-polynomial Action policy via Sparse History-anchored flow (Li et al., 1 Mar 2025, Ge et al., 29 Sep 2025, Wang et al., 19 May 2025, Bai et al., 15 May 2026).

1. Scope and principal meanings

Recent arXiv literature uses the term in at least four technically separate senses.

Usage of “flash”	Technical role	Representative papers
Active illumination	Flash/no-flash cues for reflection removal, denoising, and controllable relighting	(Lei et al., 2022, Wang et al., 2024, Xie et al., 2024, Xia et al., 2020, Maralan et al., 2023)
Non-volatile memory	Charge storage, threshold-voltage control, erase-before-program operation	(Liu et al., 2020, 0901.0702, Oh et al., 21 Jun 2025)
Storage-system substrate	SSD caching, FTL acceleration, in-flash computation, flash-based PIM	(Kang et al., 2012, Woo et al., 2017, Park et al., 2022, Oh et al., 21 Jun 2025)
Model or platform name	Acronym for specific image/video, multimodal, robotics, or browser-communication systems	(Li et al., 1 Mar 2025, Ge et al., 29 Sep 2025, Wang et al., 19 May 2025, Singh et al., 2011, Bai et al., 15 May 2026)

The usages are linked by name rather than by a shared mechanism. In imaging, flash is a controllable illumination source; in memory, it is a storage medium whose cell states are increased easily but decreased only by block erase; in machine learning, FLASH is primarily an acronym chosen for a specific method; in web communication, Flash Player is a runtime that provided audio/video capture, transport, and playback capabilities inside the browser (0901.0702, Singh et al., 2011).

2. Flash as active illumination in imaging and computational photography

In computational photography, flash is treated as a structured illumination perturbation. A flash photograph can be modeled as

$P = A + F,$

and, in an intrinsic formulation, as

$P = R \cdot (c_A S_A + S_F),$

where $R$ is reflectance, $S_A$ and $S_F$ are ambient and flash shadings, and $c_A$ is ambient color (Maralan et al., 2023). In reflection removal through glass, the standard no-flash and flash images are written as

$$\mathbf{I}_N = \mathbf{T}_N + \beta \circ \mathbf{R}, \qquad \mathbf{I}_F \approx (1+\alpha)\mathbf{T}_N + \beta \circ \mathbf{R},$$

so that

$$\mathbf{I}_F - \mathbf{I}_N \approx \alpha \mathbf{T}_N,$$

making the difference image approximately reflection-free transmission (Xie et al., 2024).

This cue underlies several modern reflection-removal systems. "Robust Reflection Removal with Flash-only Cues in the Wild" formalizes the flash-only image

$I_{fo}^{raw} = I_f^{raw} - I_a^{raw}$

and reports that its model outperforms state-of-the-art reflection removal approaches by more than $5.23\,\mathrm{dB}$ in PSNR, while its aligned variant outperforms the previous version by more than $P = R \cdot (c_A S_A + S_F),$0 in PSNR on a misaligned dataset (Lei et al., 2022). "Flash-Split" reformulates a single potentially misaligned flash/no-flash pair as latent diffusion separation, with a dual-branch diffusion stage and a cross-latent decoder; on the real flash/no-flash dataset of Lei et al., it reports 31.61 PSNR, 0.963 SSIM, and 0.048 LPIPS for transmission estimation (Wang et al., 2024). "Flash-Splat" extends flash/no-flash reflection separation to unpaired multi-view 3D reconstruction with four 3D Gaussian Splatting models $P = R \cdot (c_A S_A + S_F),$1, and reports 20.42 dB PSNR and 0.2868 LPIPS against paired difference images used as reference transmission (Xie et al., 2024).

Flash/no-flash pairs are also used to denoise extremely noisy ambient captures without inheriting flash illumination artifacts. "Deep Denoising of Flash and No-Flash Pairs for Photography in Low-Light Environments" outputs a kernel-field estimate $P = R \cdot (c_A S_A + S_F),$2 and a gain map $P = R \cdot (c_A S_A + S_F),$3, then forms the final image as

$P = R \cdot (c_A S_A + S_F),$4

The method is trained on raw linear data with a heteroscedastic Gaussian noise model and is explicitly designed to preserve the color and mood from the ambient illumination of the noisy no-flash image while recovering surface texture and detail revealed by the flash input (Xia et al., 2020). A related line, "Computational Flash Photography through Intrinsics," treats flash not as a binary capture-time decision but as a decomposable and generatable illumination layer, enabling flash strength control, ambient color editing, flash decomposition, and flash generation from no-flash photographs (Maralan et al., 2023).

A recurring limitation in this literature is that flash cues are physically informative but fragile under misalignment, glare, severe specularities, weak flash contribution, or dynamic scenes. The trajectory of the field has therefore been from tightly aligned 2D differencing toward latent-space separation, robustness to handheld capture, and fully 3D inverse rendering (Wang et al., 2024, Xie et al., 2024).

3. FLASH in image, video, and multimodal generation

In generative modeling, FLASH commonly names acceleration or conditioning mechanisms rather than illumination. "Learning to Animate Images from A Few Videos to Portray Delicate Human Actions" introduces FLASH, short for Few-shot Learning to Animate and Steer Humans, a diffusion-based image-to-video framework that takes a single reference image and an action description and generates a 16-frame video in which the person performs a specific action. It is trained with 16 or fewer example videos per action class and uses a Motion Alignment Module together with a Detail Enhancement Decoder to learn motion that is invariant to appearance changes and to propagate reference-image details through time. On 16 HAA500 actions, it reports Cosine RGB 85.48, Cosine Flow 77.42, CD-FVD 815.11, Temporal Consistency 95.81, and a human preference rate of 65.78% of 488 valid responses over baselines (Li et al., 1 Mar 2025).

"FlashI2V: Fourier-Guided Latent Shifting Prevents Conditional Image Leakage in Image-to-Video Generation" addresses a different problem: conditional image leakage in image-to-video models. Its central modification is to replace concatenation of the clean condition latent with a latent shift

$P = R \cdot (c_A S_A + S_F),$5

and to supplement it with high-frequency magnitude features $P = R \cdot (c_A S_A + S_F),$6 derived by Fourier analysis without phase. This prevents the denoiser from shortcutting through the first-frame latent. With 1.3B parameters, FlashI2V reports a dynamic degree score of 53.01 on VBench-I2V, surpassing CogVideoX1.5-5B-I2V and Wan2.1-I2V-14B-480P, while also showing the best generalization on out-of-domain data among the paradigms it compares (Ge et al., 29 Sep 2025).

"FLASH: Latent-Aware Semi-Autoregressive Speculative Decoding for Multimodal Tasks" addresses multimodal decoding latency rather than video synthesis quality. It combines a latent-aware visual token compression

$P = R \cdot (c_A S_A + S_F),$7

with a semi-autoregressive draft model that predicts multiple future tokens per forward pass. In video captioning and visual instruction tuning, it reports up to 2.68$P = R \cdot (c_A S_A + S_F),$8 speed-up and 2.55$P = R \cdot (c_A S_A + S_F),$9 speed-up, respectively, over the original LMM, while preserving the target model’s output quality through standard speculative verification (Wang et al., 19 May 2025).

Across these works, FLASH is associated with the same general design pattern: shift expensive full-model inference toward a cheaper intermediate representation—motion channels and correspondences, shifted latents and Fourier magnitudes, or compressed multimodal hidden states—while preserving the behavior of the larger target model (Li et al., 1 Mar 2025, Ge et al., 29 Sep 2025, Wang et al., 19 May 2025).

4. Flash memory as a physical medium and coding object

In memory systems, flash is a non-volatile storage technology in which information is stored as charge that shifts a transistor threshold voltage. The coding-theoretic abstraction models a flash block as

$R$0

with the crucial asymmetry that increasing a cell level is easy, whereas reducing a cell level requires erasing the entire block. This motivates flash codes, whose objective is to maximize the number of bit writes before an erase operation is required. For an $R$1 flash code that guarantees $R$2 writes, the write deficiency is

$R$3

and "Multidimensional Flash Codes" gives a multidimensional recursive construction with deficiency

$R$4

for $R$5 bits (0901.0702).

At the device level, "Ultrafast non-volatile flash memory based on van der Waals heterostructures" demonstrates a MoS$R$6/h-BN/multi-layer graphene flash cell in which two-layer MoS$R$7 is the semiconductor channel, h-BN is the tunneling layer, and multilayer graphene is the floating gate. The device achieves ultrafast writing/erasing speed (~20 ns), surpassing the reported state-of-the-art flash memory (~100 ns), shows a maximum memory window of 53 V at $R$8 V, and extrapolates 10-year retention with 50.6% and 50.4% of the initial threshold difference remaining for written and erased states, respectively (Liu et al., 2020).

The vdW flash results also make explicit the standard speed–retention trade-off. With 7.5 nm h-BN, retention degrades because leakage becomes too large; with 18 nm h-BN, the fastest write speed observed is 10 µs at $R$9 V; the 10.5 nm barrier is the reported compromise that enables both nanosecond-scale switching and non-volatile storage (Liu et al., 2020). This suggests that flash, even in materially novel realizations, remains governed by the same fundamental compromise between tunneling efficiency, leakage, and endurance that structured classical floating-gate models sought to manage algorithmically (0901.0702, Liu et al., 2020).

5. Flash-aware architectures, in-flash computation, and reliability optimization

At the system level, flash is not only a medium but also a performance bottleneck and an opportunity for architectural specialization. "Flash-based Extended Cache for Higher Throughput and Faster Recovery" introduces FaCE, which uses flash SSDs as a persistent cache layer between the DRAM buffer and disk. Its main flash-aware mechanisms are Multi-Version FIFO (mvFIFO) replacement and Group Second Chance, both designed to favor sequential flash writes. In PostgreSQL TPC-C experiments, FaCE delivers 2× (or more) throughput improvements, 4–5× shorter restart, and shows that a small flash cache of about 10% of database size can outperform storing the entire database on flash (Kang et al., 2012).

For SSD controllers, the dominant issue is often address translation overhead. "FMMU: A Hardware-Automated Flash Map Management Unit for Scalable Performance of NAND Flash-Based SSDs" moves mapping-table management from embedded processors into hardware. It reports that FMMU reduces FTL execution time in the map-cache hit case and the miss case by 44% and 37%, respectively, compared with a software-based approach operating in 4-core, and that it prevents FTL from acting as a performance bottleneck for up to 32-channel, 8-way SSD using PCIe 3.0 x32 host interface (Woo et al., 2017).

Flash can also be used as a compute substrate. "Flash-Cosmos: In-Flash Bulk Bitwise Operations Using Inherent Computation Capability of NAND Flash Memory" exploits NAND structure via Multi-Wordline Sensing (MWS) and Enhanced SLC-mode Programming (ESP). Intra-block MWS performs multi-operand bitwise AND in a single sensing operation, inter-block MWS performs multi-operand OR, inverse-read supports NOT, and internal latch logic supports XOR/XNOR. The paper reports zero bit errors across more than $S_A$0 bits, corresponding to an empirical bound of RBER < $S_A$1, and average 3.5×/32× and 3.3×/95× improvements in performance and energy efficiency over the state-of-the-art in-flash and outside-storage techniques across three real-world applications (Park et al., 2022).

A more recent reliability question appears in flash-based processing-in-memory. "Quantizing for Noisy Flash Memory Channels" argues that verify-level optimization for conventional storage is misaligned with PIM workloads, where MSE and PSNR matter more than BER. It formulates a joint optimization over quantization thresholds, reproduction values, and verify-level-derived separations under a fixed threshold-voltage window $S_A$2, and solves the verify subproblem as a convex program. On BSD68 image retrieval, the proposed joint design improves a sample image from 17.92 dB to 23.13 dB at $S_A$3; on SwinIR-based super-resolution, it improves a Set14 image from 16.12 dB to 31.62 dB at $S_A$4 (Oh et al., 21 Jun 2025).

Taken together, these works shift the meaning of flash from passive storage toward a programmable substrate whose performance depends jointly on replacement policy, translation hardware, in-array computation, and signal-distortion-aware write design (Kang et al., 2012, Woo et al., 2017, Park et al., 2022, Oh et al., 21 Jun 2025).

6. Flash as runtime environment and as robotic-policy acronym

A historically separate usage is Flash Player as a browser plugin for audio and video communication. "Flash-based Audio and Video Communication in the Cloud" describes Flash Player as an execution environment with built-in support for audio/video capture, encoding, playback, RTMP over TCP, and RTMFP over UDP. Its generic VideoIO.swf exposes a simple src-based JavaScript API and supports one-to-one calls, conferencing, broadcast, recording, and SIP interoperability through a SIP–RTMP gateway. The paper emphasizes both the advantages of ubiquity and device access and the limitations of proprietary protocols, transport constraints, and codec mismatch, especially for video interoperability with native SIP endpoints (Singh et al., 2011).

In robotics, "FLASH: Efficient Visuomotor Policy via Sparse Sampling" uses the acronym Fast Legendre-polynomial Action policy via Sparse History-anchored flow. Rather than generating discrete action chunks, it represents future actions as Legendre-polynomial trajectories,

$S_A$5

predicts future coefficients from history coefficients in a single flow-matching step, and obtains desired velocity feed-forward by analytic differentiation. Across five simulated and two real-world manipulation tasks, it reports state-of-the-art success rates ($S_A$6 across all tasks), per-episode inference time of $S_A$7, up to $S_A$8 faster inference than diffusion policies, $S_A$9 faster inference than prior flow-matching policies, up to $S_F$0 faster training convergence than ACT, and $S_F$1 to $S_F$2 reduction in controller tracking error relative to discrete-action baselines (Bai et al., 15 May 2026).

These two usages are terminologically adjacent but technically unrelated. Flash Player denotes a media runtime that historically bridged browser and telephony constraints, whereas FLASH Policy denotes a high-speed visuomotor controller built around sparse polynomial trajectory prediction (Singh et al., 2011, Bai et al., 15 May 2026). Their coexistence in the literature illustrates the broader status of flash as a reused term whose meaning is determined entirely by domain context.

7. Cross-domain patterns and limits

Across vision, generative modeling, memory systems, and control, flash repeatedly appears where a system benefits from a fast, structured auxiliary mechanism: active illumination in photography, speculative drafts in multimodal decoding, a floating-gate state in memory, an SSD cache between DRAM and disk, or a history-anchored polynomial prior in robotics. This does not imply a shared technical foundation, but it does explain why the label recurs in systems that prioritize low latency, controllability, or non-volatility.

The limitations are likewise domain-specific. Active-illumination methods remain sensitive to glare, weak flash contribution, or large motion (Lei et al., 2022, Xie et al., 2024). Image-to-video FLASH methods still report failures on complex human–object interactions, out-of-distribution initial states, and conditional leakage or overfitting when conditioning is handled poorly (Li et al., 1 Mar 2025, Ge et al., 29 Sep 2025). Flash memory prototypes still face high operating voltages, endurance constraints, and manufacturing-scale integration challenges (Liu et al., 2020). Flash-based system software must reconcile flash asymmetry, translation overhead, and analog noise rather than treating the medium as simply “faster disk” (Kang et al., 2012, Woo et al., 2017, Oh et al., 21 Jun 2025). Flash Player–based communication, finally, was constrained by proprietary transports and plugin dependence even when it provided a practical precursor to later browser real-time communication stacks (Singh et al., 2011).

In contemporary research usage, therefore, flash is best understood not as a single topic but as a family of domain-specific concepts: a controllable source of illumination, a non-volatile electronic memory, a browser communication substrate, and a recurrent acronym for systems that trade structural prior knowledge for speed.