MSPG-SEN: Multi-Scale Progressive GAN

• The paper introduces MSPG-SEN as a design pattern that integrates multi-scale progressive generation with a specialized enhancement network, improving performance across various GAN tasks.
• It employs multi-scale generators with progressive refinement and diverse SEN mechanisms—such as VN+SC, SE, and Fourier injection—to optimize reconstruction accuracy and control.
• Empirical results in frame interpolation, super-resolution, and synthesis show MSPG-SEN provides competitive quality while balancing efficiency and computational cost.

Searching arXiv for the named model and closely related formulations to ground the article in current records. Multi-Scale Progressive Generative Adversarial Network, abbreviated MSPG-SEN, denotes a non-uniform but increasingly useful label for adversarial architectures that combine multi-scale representation, progressive refinement or staged growth, and an additional enhancement component operating on channels, statistics, or scale conditioning. In the surveyed arXiv literature, the term is explicit in "Two-flow Feedback Multi-scale Progressive Generative Adversarial Network" (Weikai et al., 22 Aug 2025), but closely related formulations appear earlier under other names: FIGAN for frame interpolation (Amersfoort et al., 2017), a multi-scale recursive super-resolution system with a statistics-capturing discriminator (Michelini et al., 2018), BSD-GAN with scale-disentangled latent branches and a later SE-based mapping (Yi et al., 2018), PC-GANs for pan-sharpening (2207.14451), and a StyleGAN3-based scale-space generator recast as an MSPG-SEN-style system (Wolski et al., 2024). Across these works, MSPG-SEN is best understood as a design pattern rather than a single canonical network.

1. Terminological scope and lineage

The label is not used consistently across papers. In "Frame Interpolation with Multi-Scale Deep Loss Functions and Generative Adversarial Networks" (Amersfoort et al., 2017), the paper introduces FIGAN rather than MSPG-SEN, and its discriminator-generator design is mapped to a "Multi-Scale Progressive GAN" because the generator estimates motion and synthesis features in a coarse-to-fine manner. In "Multi-Scale Recursive and Perception-Distortion Controllable Image Super-Resolution" (Michelini et al., 2018), the system is explicitly interpreted as an MSPG-SEN because the discriminator contains a specialized statistics-capturing layer, namely Variance Normalization and Shift Correlator (VN+SC). In "PC-GANs: Progressive Compensation Generative Adversarial Networks for Pan-sharpening" (2207.14451), the core system corresponds to MSPG without SEN, while SE integration is described as an extension. In "Learning Images Across Scales Using Adversarial Training" (Wolski et al., 2024), the acronym does not appear in the paper, but the method is recast as MSPG-SEN because scale-aware conditioning, procedural Fourier injection, and cross-scale consistency act as a scale-enhancement component. The 2025 paper uses MSPG-SEN directly, but its concrete definition of SEN centers on APFL, GCTDRN, DEMA, and an auxiliary feature discriminator rather than on classical SE blocks (Weikai et al., 22 Aug 2025).

Source	Multi-scale/progressive mechanism	SEN interpretation
(Amersfoort et al., 2017)	Coarse-to-fine residual flow and synthesis refinement	Proposed extension via SE blocks
(Michelini et al., 2018)	Recursive multigrid SR over $2\times,4\times,8\times$	Statistics Extraction Network via VN+SC
(Yi et al., 2018)	Progressive growing with latent branch de-freezing	Mapped extension via SE gating
(2207.14451)	Triple-GAN progressive compensation	Suggested SE add-on, not original
(Wolski et al., 2024)	Progressive scale-bin training and Fourier injection	Scale-enhancement component
(Weikai et al., 22 Aug 2025)	Two-flow progressive GAN with feedback	DEMA/APFL-centered enhancement

A second source of ambiguity is the meaning of "SEN." In the super-resolution formulation, SEN is not a squeeze-and-excitation block; it is a Statistics Extraction Network formed by VN+SC. In the BSD-GAN mapping and the FIGAN extension, SEN refers instead to channel recalibration through standard SE modules. This suggests that MSPG-SEN functions as an umbrella term whose suffix identifies an enhancement mechanism, but not a unique one (Yi et al., 2018).

2. Core architectural pattern

The common architectural core is multi-scale generation coupled to a progressive procedure that either refines predictions from coarse to fine or grows the model through resolution stages. In FIGAN, two consecutive frames $I_0$ and $I_1$ are processed through a 3-level pyramid with downsampling factors $\times 8$, $\times 4$, and $\times 2$, then upsampled to full resolution. At each scale $s$, the network predicts synthesis features $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$, where $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$ is a learned bidirectional motion field and $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$ is a spatial blending weight. The coarsest estimate is produced by $I_0$0, while finer scales use $I_0$1 to predict residual corrections after warping the inputs with the upsampled coarser flow (Amersfoort et al., 2017).

In the controllable super-resolution system, the generator is G-MGBP, a generative version of Multi-Grid Back-Projection. It operates recursively across scales $I_0$2, $I_0$3, and $I_0$4, and uses learned Analysis, Synthesis, Upscale, and Downscale modules with latent iterative back-projection. A single external amplitude parameter $I_0$5 scales Gaussian noise channels injected into the Upscale module, allowing inference-time traversal of the perception-distortion trade-off. All scales are produced in a single forward pass, and parameters are shared across scales (Michelini et al., 2018).

BSD-GAN realizes progression differently. Its generator is DCGAN-like, but the latent vector is partitioned into scale-specific sub-vectors $I_0$6. Training proceeds by progressively increasing resolution and de-freezing one latent branch at a time. Stage I at each scale trains only the newly added generator block, whereas Stage II unfreezes all generator layers and gradually activates the new branch by feeding $I_0$7 with $I_0$8 increasing from $I_0$9 to $I_1$0. The paper attributes scale separation to branch suppression, whereby previously trained branches retain coarse structure and newly activated branches are pushed toward finer features (Yi et al., 2018).

PC-GANs implement progression as compensation across spatial scales and directions. A Deep Multiscale Guidance module first produces a pre-fused image $I_1$1, then a Spatial-Spectral Residual Compensation module refines it with two reverse-architecture GANs: one coarse-to-fine and one fine-to-coarse. The coarse-to-fine update is

$I_1$2

with an analogous fine-to-coarse recursion used during training to close the cycle (2207.14451).

The continuous scale-space model generalizes multi-scale progression from discrete pyramids to a continuous scale parameter

$I_1$3

with typical $I_1$4, corresponding to zoom factors up to $I_1$5. Its generator is an alias-free StyleGAN3-R conditioned on continuous position and scale through

$I_1$6

and supplied with binned Fourier features whose blending weights vary with scale. Progression is enforced by curriculum sampling over scale bins and by an explicit scale-consistency loss across $I_1$7 and $I_1$8 (Wolski et al., 2024).

3. Generator, discriminator, and SEN mechanisms

The generator side of MSPG-SEN systems is uniformly multi-scale, but the discriminator side varies from a lightweight image-level classifier to recursive multi-scale discriminators and auxiliary feature critics. FIGAN uses a lightweight image discriminator with 32 initial filters and 8 Conv-BatchNorm-LeakyReLU blocks, alternating strides 2 and 1, and doubling channels at each stride-2 block. It is unconditional in the original paper and distinguishes real $I_1$9 from generated $\times 8$0. The mapped MSPG extension notes that conditioning $\times 8$1 on $\times 8$2 is straightforward (Amersfoort et al., 2017).

The super-resolution formulation provides the clearest example of SEN as a statistics extractor rather than channel attention. Its multi-scale recursive discriminator $\times 8$3 consumes generator outputs at all scales and begins with a VN+SC layer. Variance normalization is defined as

$\times 8$4

and the shift correlator constructs 49 channels through

$\times 8$5

These channels encode local correlation statistics that are described as close to Gaussian for natural images and sensitive to distortions. The VN+SC output is then passed through a 4-layer dense block, and the discriminator recurses across scales with shared parameters (Michelini et al., 2018).

In the SE-based interpretation of MSPG-SEN, the relevant operation is the standard squeeze-and-excitation block for a feature tensor $\times 8$6:

• squeeze: $\times 8$7,
• excitation: $\times 8$8,
• scale: $\times 8$9.

The mapped FIGAN extension proposes inserting SE after each convolution layer in $\times 4$0, $\times 4$1, and $\times 4$2, and after each discriminator block. The BSD-GAN-to-MSPG-SEN mapping proposes SE after each progressive upsampling or downsampling block, optionally conditioning the gate on the scale-specific latent branch $\times 4$3 (Yi et al., 2018).

PC-GANs do not include attention or SE in the original design. Their feature fusion is performed by concatenation inside generator blocks, and discriminators $\times 4$4 and $\times 4$5 are ordinary convolutional critics over coarse and fine domains. The paper explicitly treats SE as a possible extension and suggests placing channel-SE in the Deep Multiscale Guidance module, inside SSRC residual blocks, and optionally in the discriminators, with the warning that over-aggressive reweighting can induce spectral bias (2207.14451).

The 2024 continuous scale-space model again diverges from classical SE. Its enhancement mechanism is the injection of procedural Fourier features in scale bins, assigned to layers according to per-layer Nyquist limits, combined with a cross-scale coherence loss. The 2025 explicit MSPG-SEN paper departs further: it defines a globally connected two-flow dynamic residual network, a Dynamic Embedded Attention Mechanism, and an auxiliary discriminator on intermediate generator features. Its fusion rule is

$\times 4$6

with $\times 4$7 produced dynamically, and with $\times 4$8 and $\times 4$9 providing joint adversarial feedback (Wolski et al., 2024, Weikai et al., 22 Aug 2025).

4. Objectives, supervision, and optimization

A defining property of MSPG-SEN systems is that adversarial training is rarely used in isolation; it is coupled to scale-aware reconstruction, perceptual, cycle, or consistency losses. FIGAN uses a multi-scale synthesis loss

$\times 2$0

with $\times 2$1 and $\times 2$2, plus a refinement loss on $\times 2$3. The distance

$\times 2$4

combines $\times 2$5 with VGG $\times 2$6 features, and the overall objective adds an image-level GAN term weighted by $\times 2$7. The paper states that multi-level supervision constrains the solution space, prevents degenerate decompositions, accelerates convergence, and stabilizes the GAN term by anchoring low-frequency structure early (Amersfoort et al., 2017).

The controllable super-resolution system uses deliberate loss routing by the control parameter $\times 2$8. Reconstruction and one branch of cycle loss are evaluated with $\times 2$9, while adversarial and contextual losses are evaluated with $s$0. Its total objective is

$s$1

The training schedule pre-trains $s$2 with $s$3 and only $s$4, then activates the full objective. Optimization uses Adam with initial learning rate $s$5, square-root decay, batch size 16, and $s$6 patches (Michelini et al., 2018).

BSD-GAN uses the non-saturating GAN objective and a two-stage schedule at every resolution. During Stage I, only the newest generator block is trainable; during Stage II, all generator layers are unfrozen and the new latent branch is gradually activated. This progressive de-freezing is the central training mechanism, rather than an auxiliary regularizer (Yi et al., 2018).

PC-GANs use a joint compensation loss

$s$7

with $s$8 and $s$9, combined with LSGAN adversarial terms for DMG, C2F, and F2C. The DMG module is first supervised by $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$0, then the triple-GAN system is trained cyclically with Adam for generators, mini-batch gradient descent for discriminators, learning rate $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$1, and 300 epochs (2207.14451).

The continuous scale-space model uses StyleGAN3 non-saturating logistic loss with R1 and ADA, augmented by the exact scale-consistency term

$\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$2

where $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$3 is implemented as a linear combination of $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$4 and LPIPS. The offset $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$5 is drawn from a Beta distribution conditioned on $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$6, and gradients are backpropagated through only one branch of the scale-consistency pair per iteration for stability (Wolski et al., 2024).

The explicit 2025 MSPG-SEN paper introduces APFL, which adapts learning rates and loss weights according to generator-quality and discriminator-accuracy signals. Its generator objective is written as

$\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$7

with AdamW, EMA decay $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$8, batch size 16, dropout $\Gamma^{(s)}=\{\Delta^{(s)},W^{(s)}\}$9, and a StepLR-like update scheme. This suggests an MSPG-SEN variant in which progression is controlled not only by resolution staging but also by adaptive feedback between generator and discriminator (Weikai et al., 22 Aug 2025).

5. Empirical behavior across application domains

In frame interpolation, FIGAN emphasizes the efficiency-accuracy trade-off. At $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$0, the paper reports a baseline CNN at $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$1 dB with $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$2k parameters and $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$3 G FLOPs; the multi-scale model with refinement reaches $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$4 dB at $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$5 G FLOPs and $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$6 s/frame; FIGAN itself, defined as MS+VGG+GAN, reaches $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$7 dB at $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$8 G FLOPs and $\Delta^{(s)}\in\mathbb{R}^{H_s\times W_s\times 2}$9 s; SepConv $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$0 reaches $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$1 dB but at $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$2 G FLOPs and $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$3 s/frame. The paper states that FIGAN provides subjective visual quality comparable to the best performing interpolation method at $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$4 faster runtime, while the best numeric PSNR reported in the paper is $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$5 dB for MS trained on 200k, without GAN (Amersfoort et al., 2017).

In super-resolution, the G-MGBP-based MSPG-SEN is explicitly controllable. It achieved 2nd best perceptual quality in PIRM Region 3 $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$6, 5th in Region 2 $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$7, and 7th in Region 1 $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$8. The full model uses approximately $W^{(s)}\in\mathbb{R}^{H_s\times W_s\times 1}$9k parameters, runs in approximately $I_0$00 s per image, and can traverse the perception-distortion plane by varying $I_0$01 at inference. The transition is reported as sharp near $I_0$02–$I_0$03, where contextual similarity improves initially while RMSE remains modest (Michelini et al., 2018).

For unconditional image synthesis, BSD-GAN contributes scale-disentangled latent control rather than distortion metrics. Its evidence includes variance-by-scale distributions, latent mixing across scales, and the smallest average minimum modified iGAN objective values across tested datasets relative to DCGAN, PGGAN, InfoGAN, and StyleGAN. Qualitative control is described as coarse branches affecting global color and layout while finer branches modify facial details, hair, lighting, or analogous high-frequency attributes in other classes (Yi et al., 2018).

In pan-sharpening, PC-GANs report best reduced-resolution metrics on both QuickBird and WorldView-4 among the compared methods. On QB, the paper reports $I_0$04, $I_0$05, and $I_0$06; on WV-4, $I_0$07, $I_0$08, and $I_0$09. In full-resolution evaluation, the paper reports on QB $I_0$10, $I_0$11, $I_0$12, and on WV-4 $I_0$13, $I_0$14, $I_0$15. The SSRC module also improves other pan-sharpeners when used as a post-processor (2207.14451).

The continuous scale-space model extends the empirical scope of MSPG-SEN-like systems beyond fixed discrete resolution conversion. It demonstrates zoom-in factors up to $I_0$16 and interactive navigation at approximately $I_0$17 fps. Reported FID values for reconstruction from unstructured patches include Himalayas $I_0$18, Spain $I_0$19, Milkyway $I_0$20, Moon $I_0$21, and Rembrandt $I_0$22; for generative multiscale modeling, the paper reports, for example, MoonGen FID $I_0$23 versus AnyresGAN $I_0$24, together with stronger scale-consistency statistics (Wolski et al., 2024).

The 2025 explicit MSPG-SEN paper claims state-of-the-art generation results on datasets named INKK $I_0$25, AWUN $I_0$26, IONJ $I_0$27, POKL $I_0$28, and OPIN $I_0$29, and attributes gains to APFL, GCTDRN, DEMA, and AFE. However, the text does not define what these percentages measure and does not provide standard metrics such as FID, IS, or LPIPS. That omission is itself a substantive empirical characteristic of the paper (Weikai et al., 22 Aug 2025).

6. Limitations, misconceptions, and open directions

A common misconception is that MSPG-SEN denotes a single settled architecture. The literature instead shows at least three distinct families of "SEN" mechanisms: a Statistics Extraction Network via VN+SC in super-resolution, Squeeze-and-Excitation-style channel gating in mapped variants of frame interpolation and scale-disentangled GANs, and more general scale-enhancement or attention-feedback machinery in continuous scale-space modeling and the 2025 explicit MSPG-SEN formulation (Michelini et al., 2018, Amersfoort et al., 2017, Wolski et al., 2024).

Another misconception is that multi-scale progression alone resolves instability or realism-fidelity trade-offs. The surveyed papers repeatedly report residual failure modes. In frame interpolation, fast or large motion, layered scenes, occlusions and disocclusions, and texture flicker remain problematic; blending weights help but do not eliminate ghosting or holes, and single-step GAN or perceptual losses do not enforce multi-frame consistency (Amersfoort et al., 2017). In super-resolution, high $I_0$30 can over-hallucinate textures, the contextual loss can eventually behave like a distortion metric, and the adversarial-plus-VN+SC setup can overfit non-reference metrics (Michelini et al., 2018). In BSD-GAN, disentanglement is scale-aware rather than semantic-aware, added layers can still alter earlier weights, and too many branches or overly fast de-freezing can cause leakage across scales (Yi et al., 2018). In pan-sharpening, SE integration is not free: the paper warns that over-aggressive channel reweighting can induce spectral bias, which is especially critical because spectral fidelity is the overriding requirement (2207.14451). In the continuous scale-space model, occasional faint parallel-line overlays or saturated blobs appear at finest scales, and the method still benefits from coarse scale labels even though it is robust to noisy ones (Wolski et al., 2024). In the explicit 2025 MSPG-SEN paper, ambiguous metrics, missing standardized comparisons, and absent reproducibility details constrain interpretation of the reported gains (Weikai et al., 22 Aug 2025).

The open directions proposed across the literature are relatively coherent. Suggested extensions include conditional or multi-scale discriminators, explicit forward-backward consistency losses for motion fields, explicit occlusion-aware masks, temporal adversarial or cycle-consistency losses for video, spatial attention such as CBAM in conjunction with SE, deformable convolutions in synthesis refinement, curriculum schemes over motion magnitude, and standardized FID/IS/precision-recall reporting when evaluating scale-disentangled GANs (Amersfoort et al., 2017, Yi et al., 2018). This suggests that the most stable future definition of MSPG-SEN may be methodological rather than taxonomic: a multi-scale adversarial system in which progression is coupled to an explicit enhancement operator that constrains what each scale should represent.