TinySR: Diffusion SR & Symbiotic Radio

• TinySR is a polysemous label that denotes two independent constructs: a compact diffusion-based model for image super-resolution and an ultra-low-power TIM-driven standalone RIS for symbiotic radio.
• In real-world image super-resolution, TinySR accelerates inference and reduces resource usage via techniques like depth pruning, VAE compression, and module removal, achieving lower MACs and latency relative to traditional methods.
• In wireless communications, TinySR employs time index modulation combined with a reconfigurable intelligent surface to enhance link reliability while enabling simultaneous RF energy harvesting for low-power IoT applications.

TinySR is a label that, in the recent arXiv literature, denotes two technically distinct constructs rather than a single unified method. In computer vision, TinySR is a compact one-step diffusion model for real-world image super-resolution (Real-ISR) that emphasizes depth pruning, VAE compression, and inference acceleration (Dong et al., 24 Aug 2025). In wireless communications, the term appears as an integrated narrative for an ultra-low-power symbiotic radio design in which a time index modulation (TIM)-driven primary transmitter powers and cooperates with a standalone reconfigurable intelligent surface (RIS) that harvests wireless energy and conveys its own information (Pihtili et al., 2024). The two usages share an emphasis on resource efficiency and compact operation, but they address different problem classes, design variables, and evaluation criteria.

1. Terminological scope and disambiguation

The term TinySR is polysemous in the provided literature. One usage belongs to Real-ISR, where TinySR is explicitly the title of "TinySR: Pruning Diffusion for Real-World Image Super-Resolution" (Dong et al., 24 Aug 2025). The other usage is embedded in the integrated narrative of "Time Index Modulation-Driven Standalone RIS Mechanism for Symbiotic Radio," where the proposed TIM-driven standalone RIS mechanism is stated to realize TinySR for ultra-low-power IoT environments (Pihtili et al., 2024).

Usage	Domain	Core objective
TinySR	Real-world image super-resolution	Real-time one-step diffusion SR with reduced parameters, MACs, and latency
TinySR	Symbiotic radio with standalone RIS	Joint communication and wireless energy harvesting under standalone RIS energy neutrality

A common source of confusion is to treat TinySR as a single cross-domain framework. The provided literature does not support that reading. Instead, it suggests two independent research directions that happen to share the same label: one in generative restoration and one in low-power wireless system design.

2. TinySR in real-world image super-resolution

In the Real-ISR setting, TinySR addresses low-resolution inputs affected by compound, unknown degradations, including sensor and ISO noise, non-linear spatially varying blur, ringing/compression artifacts, and mixed JPEG/WebP compression. The goal is to reconstruct high-resolution images with faithful content and realistic textures. The model is motivated by the observation that diffusion models exploit a strong generative prior and achieve high perceptual fidelity, but iterative denoising incurs substantial computational overhead, while existing one-step methods such as OSEDiff and TSD-SR remain large and compute-heavy (Dong et al., 24 Aug 2025).

TinySR adopts a teacher–student formulation derived from one-step diffusion SR. The background DDPM form is written as

$$q(x_t \mid x_{t-1})=\mathcal{N}(\sqrt{\alpha_t}x_{t-1},\beta_t I), \qquad q(x_t \mid x_0)=\mathcal{N}(\sqrt{\bar{\alpha}_t}x_0,(1-\bar{\alpha}_t)I),$$

with reverse denoising parameterized by $\epsilon_\theta$. One-step distillation learns a student $\epsilon_\theta(\cdot)$ that directly maps the degraded latent to a clean solution in a single call. TinySR keeps this one-step paradigm but compresses the teacher TSD-SR by combining depth pruning, VAE compression, removal of time and prompt modules, and pre-caching of modulation parameters.

Its stated core contributions are fivefold: Dynamic Inter-block Activation for depth pruning, an Expansion-Corrosion Strategy for pruning optimization, VAE compression through channel pruning, attention removal and lightweight SepConv, elimination of time- and prompt-related modules, and pre-caching techniques. Architecturally, the student is a pruned diffusion transformer initialized from TSD-SR; it retains approximately 50% of layers, implemented as “retain two of every four” via masks. The VAE uses max channels $=64$ across blocks, removes self-attention, uses SepConv in the encoder only, and keeps standard convolution in the decoder because SepConv in the decoder hurt restoration. On DIV2K-Val on V100, the reported footprint is 341 M parameters, 427 G MACs, and 0.026 s, compared with 2112 M parameters, 2700 G MACs, and 0.147 s for TSD-SR (Dong et al., 24 Aug 2025).

3. Pruning, compression, and acceleration mechanisms in TinySR

The central pruning mechanism is Dynamic Inter-block Activation (DIA). For an $N$-layer transformer with parameters $\Phi=[\phi_1,\ldots,\phi_N]$ and binary mask $m\in\{0,1\}^N$, TinySR gates each layer through masked residual bypass:

$x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$

Rather than using a static importance rule, the method jointly optimizes the mask distribution and the weights via a bi-level objective:

$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$

where $G$ denotes Gumbel-Softmax sampling with temperature $\epsilon_\theta$0. The paper argues that naive blockwise pruning drastically reduces the feasible mask space; the cited example is pruning 12 of 24 layers, where the full space contains 2,704,156 possibilities but the constrained blockwise scheme reaches only 46,656, about 1.725% of the full $\epsilon_\theta$1 space (Dong et al., 24 Aug 2025).

DIA relaxes that restriction by introducing learned transformation operators $\epsilon_\theta$2 governed by $\epsilon_\theta$3 that move $\epsilon_\theta$4 active layers between adjacent blocks while keeping the global budget fixed. Expansion-Corrosion Strategy (ECS) then uses learned marginal probabilities

$\epsilon_\theta$5

to decide which layers to expand and corrode. The decision rule first selects the transform via $\epsilon_\theta$6 and then ranks $\epsilon_\theta$7 so that expansion targets high-$\epsilon_\theta$8 layers while corrosion removes low-$\epsilon_\theta$9 ones. This makes pruning recoverability-driven rather than purely heuristic.

The VAE compression path targets three bottlenecks identified in the teacher VAE: wide channels, expensive attention, and standard convolution dominating compute by more than 85%. The ablation on DRealSR reports the following progression: teacher VAE at 95.36 ms and 1781.91 G MACs with PSNR 27.77, SSIM 0.7559, LPIPS 0.2967; +Channels 64 at 22.79 ms and 146.56 G with PSNR 27.49 and LPIPS 0.3052; +Remove attention at 11.22 ms and 131.91 G with LPIPS 0.3222; SnapGen-style SepConv at 10.75 ms and 89.74 G with LPIPS 0.3054; and the final encoder-only SepConv variant at 9.25 ms and 78.48 G with PSNR 27.56, SSIM 0.7514, and LPIPS 0.3038 (Dong et al., 24 Aug 2025). The paper therefore attributes most of the VAE acceleration to channel pruning and attention removal, while restricting SepConv to the encoder to preserve restoration quality.

Additional acceleration comes from removing time-conditioning and prompt/text-conditioning paths. On RealSR, the ablation reports that the configuration with text and time has CLIPIQA 0.7081, MANIQA 0.6283, 35.7 ms, 535 G MACs, and 1010 M parameters, while removal yields CLIPIQA 0.7035, MANIQA 0.6235, 27.4 ms, 427 G MACs, and 516 M parameters. TinySR also pre-computes and caches adaLN-Zero modulation parameters because they are reported to become largely input-invariant after training in the one-step setting.

4. Training protocol, empirical performance, and limitations of the Real-ISR model

TinySR is trained in a teacher–student setup in which the teacher is TSD-SR and the student is the pruned compact model. The training recipe separates VAE training, pruning decision learning, latent distillation, and image-space fine-tuning (Dong et al., 24 Aug 2025). The VAE encoder uses

$\epsilon_\theta(\cdot)$0

and the decoder uses LPIPS plus GAN:

$\epsilon_\theta(\cdot)$1

with $\epsilon_\theta(\cdot)$2 and $\epsilon_\theta(\cdot)$3. For pruning, the joint objective is

$\epsilon_\theta(\cdot)$4

with $\epsilon_\theta(\cdot)$5 and $\epsilon_\theta(\cdot)$6. Restoration then proceeds with Stage 1 latent distillation,

$\epsilon_\theta(\cdot)$7

followed by Stage 2 image-space fine-tuning,

$\epsilon_\theta(\cdot)$8

The reported coefficients are $\epsilon_\theta(\cdot)$9, $=64$0, and $=64$1.

The optimization setup uses 8× NVIDIA V100 GPUs, fp16 precision, and random seed 80. The data are DIV2K, Flickr2K, LSDIR, and FFHQ, with degradation via Real-ESRGAN and scale $=64$2 from $=64$3 to $=64$4. The pruning decision stage initializes from TSD-SR weights, uses prune rate 50%, LoRA rank 64, AdamW with learning rate $=64$5, global batch 8, and 100k iterations. Restoration Stage 1 uses global batch 128, LoRA rank 64, learning rate $=64$6, and 150k steps. Stage 2 uses global batch 96, student learning rate $=64$7, discriminator learning rate $=64$8, LoRA rank 64, and 50k steps.

Evaluation covers synthetic DIV2K-Val and real paired RealSR and DRealSR. Metrics include PSNR/SSIM on Y, LPIPS, DISTS, FID, NIQE, MUSIQ, MANIQA, CLIPIQA, TOPIQ, and Q-Align. On DIV2K-Val, TinySR reports SSIM 0.5725, LPIPS 0.2793, DISTS 0.1883, FID 22.44, NIQE 4.1500, MUSIQ 69.90, CLIPIQA 0.7201, 0.026 s, 427 G MACs, and 341 M parameters. In the same table, TSD-SR reports SSIM 0.5808, LPIPS 0.2673, DISTS 0.1821, FID 29.16, NIQE 4.3224, MUSIQ 71.69, CLIPIQA 0.7416, 0.147 s, 2700 G MACs, and 2112 M parameters. The paper therefore characterizes TinySR as achieving the fastest inference and lowest MACs while remaining competitive in perceptual quality; it slightly trails TSD-SR on LPIPS and DISTS but improves FID and NIQE.

The depth-pruning ablation attributes the best DISTS, FID, MANIQA, and CLIPIQA among the listed pruning methods to DIA+ECS. The pruning-ratio study indicates graceful degradation from 33% to 50% pruning, with LPIPS around 0.279 and DISTS around 0.188, but sharp losses at 67% pruning, where LPIPS becomes 0.2984 and DISTS 0.2110. The paper also notes that distilling directly from ground-truth HR in either image or latent space produces smooth or blurry outputs, whereas distilling from the teacher improves perceptual metrics. The main limitations recorded are reduced flexibility after removing time and prompt conditioning, a practical pruning limit near 50% depth reduction, and unresolved opportunities for multi-scale training, mobile NPU or SoC benchmarking, quantization, and operator-level optimization.

5. TinySR as a TIM-driven standalone RIS mechanism for symbiotic radio

In the wireless-communications usage, TinySR is realized through a symbiotic radio environment comprising a single-antenna primary transmitter (PTx), an $=64$9-antenna primary receiver (PRx), a standalone RIS, and a single-antenna energy harvester (EH) (Pihtili et al., 2024). The PTx simultaneously delivers user data and RF power using a TIM waveform. The RIS is partitioned into three groups of unit cells, $N$0. During information slots, $N$1 assists the PTx$N$2PRx link, $N$3 conveys its own bit by phase modulation, and $N$4 acts as absorbers whose combined RF input powers a rectenna. In power slots, all reflectors are phase-aligned to maximize forward RF power transfer while $N$5 continues absorption. The mutualistic interpretation is explicit: the PTx provides spectrum and energy, the RIS improves PTx$N$6PRx reliability and conveys its own low-rate information without an external power supply, and the EH is powered by the same waveform.

The TIM frame has $N$7 time slots, of which $N$8 are information slots and $N$9 are power slots, with duty cycle $\Phi=[\phi_1,\ldots,\phi_N]$0. Time indexing generates $\Phi=[\phi_1,\ldots,\phi_N]$1 patterns, of which the PTx uses $\Phi=[\phi_1,\ldots,\phi_N]$2 legitimate patterns to carry

$\Phi=[\phi_1,\ldots,\phi_N]$3

additional time-index bits. In information slots the PTx transmits symbols $\Phi=[\phi_1,\ldots,\phi_N]$4 from an $\Phi=[\phi_1,\ldots,\phi_N]$5-PSK or QAM constellation, conveying $\Phi=[\phi_1,\ldots,\phi_N]$6 bits per block; the total block rate is $\Phi=[\phi_1,\ldots,\phi_N]$7.

The transmit signal is defined by the binary index vector $\Phi=[\phi_1,\ldots,\phi_N]$8:

$\Phi=[\phi_1,\ldots,\phi_N]$9

The direct PTx$m\in\{0,1\}^N$0PRx link is $m\in\{0,1\}^N$1 with Rician fading and 3GPP InH path loss. The composite RIS-assisted channel is represented through per-element coefficients $m\in\{0,1\}^N$2 and the cascaded channel

$m\in\{0,1\}^N$3

The received signal at PRx in slot $m\in\{0,1\}^N$4 is

$m\in\{0,1\}^N$5

where $m\in\{0,1\}^N$6. The RIS bit is embedded in the phase configuration of the $m\in\{0,1\}^N$7 unit cells, which hold an invariant phase over the $m\in\{0,1\}^N$8 information slots and thereby create spreading gain $m\in\{0,1\}^N$9.

The RIS hardware model is deliberately low power. The reflection coefficient set is discrete,

$x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$0

with

$x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$1

because one state is reserved for absorption. The details specify RF-switch-based unit cells with 2-bit resolution: one RF-switch port is connected to an absorber, and the remaining settings provide discrete phase states. This avoids DACs and reduces static power relative to varactor-based designs. The control chips are integrated, with $x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$2 unit cells per chip, and phase tables are precomputed and remain invariant across channel realizations for RF-switch hardware.

6. Detection, energy neutrality, design trade-offs, and reported performance in wireless TinySR

The PRx performs TIM slot detection with an LLR that compares the “information” and “power” hypotheses. The per-slot statistic is

$x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$3

with numerically stable accumulation performed through log-sum-exp recursion. The LLR vector $x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$4 is then mapped to a legitimate time-index pattern by

$x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$5

The paper states that this prevents illegal index selections and reduces complexity from ML’s $x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$6 to $x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$7 (Pihtili et al., 2024).

Wireless energy harvesting at the RIS uses the absorber subset $x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$8 to feed an RF-combining network and nonlinear rectenna. The RIS input RF power per slot is

$x_{i+1}=m_i\cdot \phi_i(x_i)+(1-m_i)\cdot x_i.$9

and the harvested DC power follows the constant-linear-constant model

$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$0

Energy neutrality of the standalone RIS requires

$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$1

where the integrated-controller power model is

$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$2

The studied configuration assumes Rician fading with $$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$3, 3GPP InH path loss at $$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$4 GHz, distances of 5 m for PTx$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$5RIS and 10 m for RIS$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$6PRx and RIS$$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$7EH, $$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$8, $$\min_{p(m)}\min_{\Delta\Phi}\; \mathbb{E}_{x,m\sim G(p(m))}\big[L(x,\Phi+\Delta\Phi,m)\big],$$9, $G$0, and $G$1 dBm, $G$2 dBm, with 4-QAM unless noted. The controller uses $G$3, $G$4, and RF-switch power $G$5. The rectenna parameters are $G$6; for the RIS, $G$7 and $G$8; for the EH, $G$9 and $\epsilon_\theta$00.

The reported results emphasize three trade-offs. First, harvested RIS DC power increases with $\epsilon_\theta$01 and with more power slots, and with RF switches the minimum $\epsilon_\theta$02 required for standalone operation is about half the varactor case because controller power is much lower. Around $\epsilon_\theta$03, the RIS CLC rectenna saturates for all TIM setups, which the paper states suggests reconfiguring $\epsilon_\theta$04 when $\epsilon_\theta$05 is large. Second, increasing $\epsilon_\theta$06 reduces EH harvested power because fewer reflectors remain, making the EH–RIS trade-off explicit. Third, BER at PRx is improved by the RIS coherent term $\epsilon_\theta$07 and by the TIM structure: LLR and ML yield nearly identical BER for both PTx and RIS information, while the LLR reduces complexity by about 86% versus ML for TIM with $\epsilon_\theta$08 and $\epsilon_\theta$09.

The design guidance is correspondingly parametric. The paper recommends choosing $\epsilon_\theta$10 and $\epsilon_\theta$11 to balance data rate and WEH, starting with $\epsilon_\theta$12 and $\epsilon_\theta$13 depending on target RIS or EH power; using deterministic $\epsilon_\theta$14 with $\epsilon_\theta$15; preferring RF-switch unit cells and integrated control; and selecting $\epsilon_\theta$16 so that the standalone constraint is met. For the stated example parameters, $\epsilon_\theta$17 already achieves standalone, leaving more unit cells for reflection. Further guidance includes allocating $\epsilon_\theta$18 for beamforming assistance and $\epsilon_\theta$19 for RIS-bit transmission, placing the RIS to ensure strong line-of-sight PTx$\epsilon_\theta$20RIS and RIS$\epsilon_\theta$21PRx components, and tuning $\epsilon_\theta$22 and $\epsilon_\theta$23 to the expected RF-input distribution. The stated limitations are perfect or near-perfect CSI assumptions, invariant discrete phase shifts consistent with RF-switch hardware, use of the CLC harvesting model rather than other nonlinear models, omission of hardware impairments and controller signaling overhead, and restriction to one RIS bit per block. Future work is framed as joint optimization of TIM parameters, RIS partitioning, phase selection, and EH parameters under energy-neutrality constraints, along with adaptive LLR detectors, hardware co-design, and multi-user or mobile TinySR extensions.