Dynamic-Image (DynImg): Adaptive Image Representation

• Dynamic-Image (DynImg) is a research motif encompassing adaptive image representations that condition processing on sample specifics, temporal order, or downstream objectives.
• It spans diverse applications such as 3D MRI compression, video representation with temporal prompts, and dynamic reconstruction, fusion, and retrieval in imaging.
• By leveraging sample-specific adaptations, DynImg methods deliver improved performance in classification, restoration, and multimodal tasks over static pipelines.

Dynamic-Image (DynImg) is not a single canonical operator in the arXiv literature, but a family of task-oriented image representations and adaptive image-processing schemes whose common feature is dynamic conditioning on temporal order, input content, or downstream objective. In different subfields, the term denotes: a recognition-oriented transformation learned jointly with a classifier; a 2D image obtained from an ordered 3D MRI volume by approximate rank pooling; a composite video representation in which key frames are augmented by non-key temporal prompts for multi-modal LLMs (MLLMs); and a broader class of dynamic reconstruction, fusion, and retrieval systems that replace fixed preprocessing or static weighting with sample-specific or region-specific adaptation (Sharma et al., 2017, Xing et al., 2020, Bao et al., 21 Jul 2025).

1. Terminological scope and recurrent design pattern

A common source of confusion is that “Dynamic-Image” and “DynImg” are used heterogeneously. In some papers the phrase denotes a concrete representation built from an ordered sequence; in others it denotes a task-driven dynamic transformation, dynamic weighting rule, or adaptive fusion mechanism. The surveyed literature therefore supports treating DynImg as a research motif rather than a single standardized algorithm.

Sense of DynImg	Core mechanism	Representative paper
Classification-driven enhancement	Sample-specific enhancement filters learned end to end before classification	(Sharma et al., 2017)
3D-to-2D MRI compression	Approximate rank pooling over ordered $z$-slices	(Xing et al., 2020)
MLLM video representation	Key frame plus non-key temporal prompts with 4D RoPE	(Bao et al., 21 Jul 2025)
Dynamic MRI reconstruction	Unsupervised or graph-based priors for time-varying images	(Yoo et al., 2019, Li et al., 2024)
Dynamic coding / low-light imaging	Time-varying spatiotemporal filtering or motion-aware photon-starved reconstruction	(Doutsi et al., 2021, Chi et al., 2020, Palladino et al., 11 May 2026)
Dynamic fusion / retrieval / advertising	Test-time dynamic weights, mutual-guided filters, or multi-objective query-image matching	(Cao et al., 2024, Guan et al., 2023, Wen et al., 2023)

Across these variants, the fixed preprocessing pipeline is typically replaced by an adaptive operator. A plausible implication is that DynImg research is best understood as part of a broader shift from static image formation or feature extraction toward representation learning in which the transformation itself is optimized for the end task.

2. Classification-conditioned dynamic image transformation

A closely related interpretation appears in "Classification Driven Dynamic Image Enhancement" (Sharma et al., 2017). The central argument is that conventional enhancement operators such as WLS, bilateral filtering, guided filtering, sharpening, and histogram equalization are designed to improve perceptual quality for a human observer, whereas classification depends on discriminative textures, edges, local contrast, and structure. The method therefore inserts an enhancement module in front of a classifier and trains the pipeline end to end so that enhancement is optimized jointly for classification accuracy and reconstruction to an enhancement target. The input RGB image $I$ is converted to $YCbCr$; only the luminance channel $Y$ is enhanced and then recombined with chrominance. For one enhancement method, the enhancement network predicts a sample-specific filter $f_\Theta$, producing

$Y'(i,j)=f_\Theta(Y(i,j)).$

The first formulation uses

$Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$

where $L$ is softmax cross-entropy. The paper then extends this to a frozen multi-branch formulation, Stat-CNN,

$$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$

and to Dyn-CNN,

$$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$

The most important contribution is Dyn-CNN, which jointly learns $I$0 dynamic filters—WLS, bilateral filtering, guided filtering, histogram equalization, and image sharpening—together with sample-specific weights, plus an identity/original-image branch. The weights are derived from the relative MSE strengths of the enhancement branches and normalized to sum to one; MSE-based weighting is reported to work better than equal weighting. On CUB-200-2011, Pascal VOC2007, MIT-Indoor, and DTD, the dynamic formulation consistently outperforms both plain fine-tuning and static enhancement. On CUB, BN-Inception improves from $I$1 to $I$2 with Stat-CNN and $I$3 with Dyn-CNN; AlexNet improves from $I$4 to $I$5 and $I$6; GoogLeNet improves from $I$7 to $I$8 and $I$9. On Pascal VOC2007, Dyn-CNN reaches $YCbCr$0 mAP with AlexNet and $YCbCr$1 with VGG-16. On MIT-Indoor it reaches $YCbCr$2 with AlexNet and $YCbCr$3 with VGG-16, and on DTD it reaches $YCbCr$4 with AlexNet and $YCbCr$5 with VGG-VD.

Several implementation details clarify what “dynamic” means here. The enhancement network for one method has about 570k parameters, and the last fully connected layer outputs 36 parameters for a $YCbCr$6 filter. Among $YCbCr$7, $YCbCr$8, and $YCbCr$9, the $Y$0 filter performs best; filters larger than $Y$1 over-smooth images and hurt classification by about 2%, whereas smaller filters drop performance by about 3%. Traditional target generation takes 1–6 seconds per image per method, while the learned EnhanceNet produces all enhanced images in about 8 ms on GPU. The main limitation, however, is that the system remains tied to a finite set of hand-designed enhancement operators and relies on intermediate target images generated by them.

3. Sequence-derived dynamic images: from volumetric MRI to MLLM video prompts

A more literal DynImg construction appears in "Dynamic Image for 3D MRI Image Alzheimer’s Disease Classification" (Xing et al., 2020). Here a 3D MRI volume is treated as an ordered sequence along the $Y$2-axis, and approximate rank pooling compresses the volume into a single 2D dynamic image. For ordered slices $Y$3, the rank-pooling formulation defines

$Y$4

with later prefixes constrained to have higher scores. Because exact RankSVM is expensive, the paper uses approximate rank pooling,

$Y$5

In this implementation, the $Y$6-dimension is treated as the temporal dimension, and the 3D MRI is compressed over the height/$Y$7 dimension into a 2D dynamic image. Using ADNI “spatially normalized, masked, and N3-corrected T1 images” of size $Y$8, the method generates dynamic images of resolution $Y$9, feeds them to fine-tuned 2D CNN backbones, and adds an attention module with four $f_\Theta$0 convolutions and a three-layer classifier.

The empirical motivation is computational: training a 3D CNN is expensive, whereas the dynamic-image projection allows reuse of ImageNet-pretrained 2D backbones. On a dataset of 100 samples—51 CU and 49 AD—the best backbone is VGG11, achieving Acc $f_\Theta$1, ROC $f_\Theta$2, F1 $f_\Theta$3, Precision $f_\Theta$4, Recall $f_\Theta$5, and AP $f_\Theta$6. Against 3D baselines, the final model reaches Acc $f_\Theta$7, ROC $f_\Theta$8, F1 $f_\Theta$9, Precision $Y'(i,j)=f_\Theta(Y(i,j)).$0, Recall $Y'(i,j)=f_\Theta(Y(i,j)).$1, and AP $Y'(i,j)=f_\Theta(Y(i,j)).$2, corresponding to a reported $Y'(i,j)=f_\Theta(Y(i,j)).$3 accuracy improvement and $Y'(i,j)=f_\Theta(Y(i,j)).$4 better ROC over 3D-ResNet. Training for 150 epochs takes 414 s, compared with 2359 s for 3D-VGG and 3916 s for 3D-ResNet. A major caveat is that skull stripping is crucial: skull-included MRI volumes produce blurrier dynamic images, and the proposed model drops to $Y'(i,j)=f_\Theta(Y(i,j)).$5 accuracy on skull-included data.

A different sequence-derived DynImg appears in "DynImg: Key Frames with Visual Prompts are Good Representation for Multi-Modal Video Understanding" (Bao et al., 21 Jul 2025). This paper argues that many MLLM video pipelines separate spatial encoding from temporal reasoning, so motion-sensitive regions may already be underrepresented when temporal modeling begins. DynImg is therefore built as a composite image in which one key frame is placed above four nearby non-key frames used as temporal prompts. Videos are decomposed using the MPEG-4 notion of I-frames; four I-frames are evenly sampled as key frames, and for each key frame two preceding and two following I-frames are used as prompts. The prompts are resized and concatenated left-to-right beneath the key frame, allowing the visual encoder to perform spatiotemporal interaction before token compression.

To preserve order, the paper introduces a 4D video Rotary Position Embedding with coordinates $Y'(i,j)=f_\Theta(Y(i,j)).$6 and combined angle

$Y'(i,j)=f_\Theta(Y(i,j)).$7

The sequence dimension uses standard sinusoidal RoPE, $Y'(i,j)=f_\Theta(Y(i,j)).$8, while $Y'(i,j)=f_\Theta(Y(i,j)).$9 are learnable and initialized to zero. The model uses SigLIP-so400m-384 as visual encoder, a feedforward projection layer with the Adaptive Average Structure Pooling module from PLLaVA, pooling shape $Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$0, and Qwen2.5-7B-Instruct as the LLM. On MSVD, MSRVTT, TGIF, ActivityNet, Video-ChatGPT, and MVBench, the paper reports approximately 2% average improvement over prior methods. MSVD reaches 78.6 Acc and 4.2 Score; MSRVTT 64.1 and 3.5; ActivityNet 57.9 and 3.6; TGIF 77.5 and 4.0. On MVBench the average accuracy is 55.8, with especially large gains on Moving Direction, Moving Count, and Moving Attribute. Ablations show that prompt insertion before the encoder is markedly stronger than post-encoder fusion, and that four prompt frames and four DynImgs per video are the best settings among those tested.

4. Dynamic images in reconstruction, coding, and inverse problems

In dynamic MRI, DynImg is often embedded in unsupervised reconstruction rather than explicit frame pooling. "Time-Dependent Deep Image Prior for Dynamic MRI" introduces a generalized deep-image-prior method in which a fixed low-dimensional temporal manifold, a mapping network $Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$1, and a CNN generator $Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$2 jointly produce the image sequence from latent variables without prior training or ground-truth images (Yoo et al., 2019). For a cardiac cycle, the manifold may be a straight line,

$Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$3

or a helix,

$Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$4

The full model $Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$5 is optimized by a measurement-domain loss over sparse radial k-space data,

$Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$6

The method requires neither heartbeat marking nor spoke reordering and outperforms GRASP and RD on the retrospective dataset, with the best configuration “Helix + MapNet” reaching about 28.05 dB RSNR.

"Graph Image Prior for Unsupervised Dynamic Cardiac Cine MRI Reconstruction" reformulates the same inverse problem by separating image recovery from manifold discovery (Li et al., 2024). Instead of one pyramid-shaped generator shared across frames, GIP uses independent CNNs $Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$7 for per-frame structure and a graph convolutional network for spatio-temporal feature fusion. Reconstruction is posed as

$Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$8

with ADMM alternating between image update, network update, and multiplier update. On OCMR, GIP achieves PSNR 50.63 dB, SSIM 99.59%, MSE 0.98, and MAE 2.24 at $Loss_{Filters}=MSE(T,Y')+L(\mathbf P,\mathbf y),$9, and PSNR 43.23 dB, SSIM 97.84%, MSE 5.34, and MAE 5.33 at $L$0. The paper emphasizes that GIP substantially narrows the gap with supervised methods and generalizes better when transferred from OCMR to CMRxRecon without additional training data.

DynImg also appears in coding and low-light sensing. "Retinal-inspired Filtering for Dynamic Image Coding" defines a non-Separable sPAtioteMporal filter, or non-SPAM filter, with kernel

$L$1

where center and surround terms couple spatial Gaussians with gamma and exponential temporal filters (Doutsi et al., 2021). For a still image $L$2, the system reduces to a time-varying spatial filter $L$3 that behaves as a temporally evolving difference of Gaussians and yields frame coefficients satisfying a frame inequality. The analysis/synthesis system therefore has stable inversion and progressive reconstruction. On a $L$4 test image, MSE decreases from $L$5 with 20% of coefficients to $L$6 with 100%.

"Dynamic Low-light Imaging with Quanta Image Sensors" addresses photon-starved dynamic scenes using a student-teacher framework with a motion teacher based on a modified KPN, a denoising teacher based on a modified REDNet, and a student network with two encoders and a 15-layer decoder (Chi et al., 2020). The QIS observation model includes Poisson arrivals, dark current, read noise, and ADC quantization, and the method reconstructs dynamic scenes at 1 photon per pixel per frame. The overall objective is

$L$7

At 2 ppp with 28 pixels of global translation across 8 frames, the reported PSNRs are 23.04 dB for BM4D, 25.45 dB for KPN, 26.42 dB for mRED, and 29.39 dB for the proposed method.

"DynGhost: Temporally-Modelled Transformer for Dynamic Ghost Imaging with Quantum Detectors" extends dynamic-image reconstruction to ghost imaging (Palladino et al., 11 May 2026). The architecture alternates spatial attention across the $L$8 measurement tokens within a frame and temporal attention across the $L$9 frames for a fixed pattern index. Training combines

$$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$0

and quantum-aware simulation models Poisson counts, dark counts, afterpulsing, and crosstalk for SNSPDs, SPADs, and SiPMs, with Anscombe and Freeman–Tukey normalization used to stabilize photon-count noise. On Moving MNIST, DynGhost reaches SSIM $$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$1, MSE $$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$2, and 7.3 ms inference time, outperforming GhostGPT, DGI, PI, and FISTA. The paper nevertheless notes weaker gains on Kvasir/endoscopic video and identifies quadratic temporal attention, synthetic-data dependence, and flux-sensitive normalization as limitations.

A more abstract DynImg usage occurs in "Dynamic Implicit Image Function for Efficient Arbitrary-Scale Image Representation" (He et al., 2023). DIIF replaces per-coordinate implicit decoding with coordinate grouping, coordinate slicing, and a Coarse-to-Fine MLP. Dynamic slicing adapts the slice interval to scale factor, reducing compute dramatically relative to LIIF. With EDSR-baseline on DIV2K, LIIF requires 5.117T MACs and 3.923 s, whereas DIIF requires 0.790T MACs and 0.336 s; at $$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$3, LIIF requires 45.97T MACs and 39.20 s, whereas DIIF requires 6.28T MACs and 3.43 s. This suggests that, in some uses, DynImg designates dynamic decoding strategy rather than temporal image content.

5. Fusion, retrieval, and industrial-scale dynamic imaging

In industrial search advertising, "Enhancing Dynamic Image Advertising with Vision-Language Pre-training" uses DynImg/DIA to denote a pipeline that matches queries to ad images and generates multimodal ads (Wen et al., 2023). The system has two stages: candidate image retrieval with ANN search and relevance-based reranking. The proposed framework consists of a pre-trained base model, a fine-tuned relevance model, and a multitask retrieval model. The base model uses ViT-B/16 as vision encoder, RoBERTa$$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$4 as language encoder, and a RoBERTa$$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$5-based fusion encoder trained with InfoNCE-based multi-view image-text contrastive learning, MLM, and ITM. Fine-tuning freezes the vision encoder. The retrieval model is optimized by bidirectional cross-modal contrastive loss on clicked pairs together with knowledge distillation from the relevance model, and relevance labels lie in $$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$6. Pre-training uses 20B image-text pairs; the relevance model uses about 500K labeled query-image pairs; the retrieval model uses 1.3B clicked pairs. In a 15-day online A/B test in Baidu’s "Phoneix Nest", the deployed system improves CPM by 1.04%, CTR by 1.865%, and P97 latency by 0.05%.

A mathematically explicit dynamic-weighting formulation is given by "Test-Time Dynamic Image Fusion" (Cao et al., 2024). For source images $$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$7, the fused image is

$$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$8

with the weights normalized so that $$Loss_{Stat}=\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y),$$9. TTD computes source-wise reconstruction losses

$$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$0

and defines Relative Dominability by

$$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$1

The paper proves that reducing the generalization-error upper bound hinges on negative correlation between the RD-based fusion weight and the uni-source reconstruction loss. TTD is training-free, parameter-free, and operates only at test time, but approximately doubles inference time because it requires an extra pass for weight estimation.

Other fusion papers use “dynamic” in the sense of adaptive filtering and degradation modeling. "Dynamic Image Restoration and Fusion Based on Dynamic Degradation" introduces DDRF-Net, which represents degradation as a weighted mixture of motion blur, isotropic Gaussian, and anisotropic Gaussian kernels,

$$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$2

and combines this with dynamic convolution for infrared-visible fusion (Fang et al., 2021). On FLIR, DDRF-Net reports best mean score 10.59, AG 9.43, and PSNR 43.66; on RGBT, mean score 11.23, AG 21.02, and EN 7.71. "Mutual-Guided Dynamic Network for Image Fusion" instead uses mutual-guided cross-attention and a dynamic filter predictor that outputs per-pixel kernels $$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$3, together with a masked mutual-information loss for redundancy suppression (Guan et al., 2023). MGDN reports best or competitive results across MEF, MFF, HDR deghosting, and RGB-guided depth map super-resolution, including CE 2.3624, $$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$4 0.4883, and VIF 0.8538 on MEF, and $$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$5, $$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$6 on HDR deghosting.

The dynamic-scene restoration literature further expands the concept. "Exploring and Evaluating Image Restoration Potential in Dynamic Scenes" introduces image restoration potential (IRP) as an image attribute that measures how much restoration quality can be unlocked from an input, rather than its current perceptual quality (Zhang et al., 2022). The DS-IRP dataset contains $$Loss_{Dyn}=\sum_{k=1}^{K} MSE_k(T_k,Y'_k)+\sum_{k=1}^{K+1} W_k\,L_k(\mathbf P,\mathbf y).$$7 labeled samples, and the proposed predictor reaches scene-average SRCC 0.9340 and PLCC 0.9412, with overall SRCC 0.8461 and PLCC 0.8687. In this setting, dynamic-image analysis is not frame aggregation but task-oriented evaluation of which frame in a dynamic scene is most valuable for restoration.

6. Boundary cases, misconceptions, and recurrent limitations

The literature makes clear that DynImg should not be reduced to a single historical meaning. It is not always a rank-pooled video surrogate, and it is not always a temporal representation. In some works, “dynamic” refers to adaptive measurement rate, adaptive ISP parameters, or dynamic coordinate slicing rather than temporal aggregation. "Compressed Domain Image Classification Using a Dynamic-Rate Neural Network" studies one network that operates over a range of compressive-sensing measurement rates and explicitly notes that this is not Dynamic-Image in the video/activity-recognition sense (Xu et al., 2019). "PQDynamicISP: Dynamically Controlled Image Signal Processor for Any Image Sensors Pursuing Perceptual Quality" likewise states that it is not a classic Dynamic-Image method, but a dynamically controlled ISP whose parameters vary per image, per sensor, and locally per region (Yoshimura et al., 2024).

Several limitations recur across otherwise different DynImg formulations. Methods based on task-driven enhancement may still depend on hand-designed target operators and carefully tuned weighting schemes (Sharma et al., 2017). Rank-pooled 2D surrogates for 3D MRI are sensitive to anatomical nuisance structure such as skull content and use an offline, non-trainable pooling stage (Xing et al., 2020). Prompt-based MLLM DynImg improves motion sensitivity, but too many prompt frames reduce per-frame resolution and weaken the signal (Bao et al., 21 Jul 2025). Unsupervised dynamic reconstruction methods often trade expressivity against optimization stability and may rely on specific manifold assumptions, graph construction choices, or ADMM schedules (Yoo et al., 2019, Li et al., 2024). Physics-aware photon-starved imaging methods still face synthetic-to-real gaps, detector-specific constraints, and computational overhead from temporal modeling or normalization (Chi et al., 2020, Palladino et al., 11 May 2026). Test-time dynamic fusion offers theoretical guarantees but remains dependent on the representational capacity of the frozen baseline model and incurs extra inference cost (Cao et al., 2024). Industrial DIA systems must cope with noisy web pairs, short and abstract ad queries, and limited diversity in business data (Wen et al., 2023).

Taken together, these works suggest that the durable core of DynImg research is not a particular image formula, but a methodological principle: the representation or transformation should be computed in a way that is conditional on the actual sample, its temporal context, or its downstream objective. In one branch this yields a single 2D image summarizing an ordered volume; in another it yields pre-encoder temporal prompts for video understanding; in others it yields test-time fusion weights, dynamic restoration kernels, graph-based temporal priors, or multi-objective cross-modal retrieval. The term therefore designates a heterogeneous but coherent research direction centered on replacing fixed image operators with dynamically constructed ones.