{"id":"ceebb06f-1194-44fe-bca5-c233f5960af6","arxiv_id":"2502.05695","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes LD-ABS, an adaptive bitrate streaming framework that compresses I-frames with a latent diffusion model and reconstructs P and B frames at the receiver, claiming better QoE than existing ABR algorithms.","lead":"This paper describes a video streaming system that uses an AI model to shrink the key frames of a video into small codes and sends only tiny adjustment data for the other frames over a wireless network. The authors claim this saves bandwidth and keeps quality high, but their paper lacks the details needed to check whether it actually works.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QoE metric in Eq. (1) measures bitrate utility, not the visual fidelity of LDM-reconstructed frames; the claimed QoE advantage over baselines is therefore unestablished.","rationale":"The reader identified the weakest load-bearing premise: the QoE metric in Eq. (1) uses bitrate as a proxy for quality and never measures the fidelity of the LDM-reconstructed video. My reading of the paper confirms this. The evaluation in Section V compares LD-ABS against BOLA, RobustMPC, Comyco, and Pensieve using a virtual player and throughput traces, but the only reported outcome is the bitrate-based QoE of Eq. (1). The paper does not provide PSNR, SSIM, VMAF, or any other direct quality measurement of the reconstructed frames, nor does it compare LD-ABS's reconstructed quality against conventional codecs at equal bitrate. Since LD-ABS transmits only latent I-frames and motion metadata, the receiver's output is generatively synthesized; assuming that the selected bitrate ak maps to the same perceptual quality as a conventionally encoded chunk is exactly the unverified step. If LDM reconstruction introduces artifacts, the reported QoE advantage could vanish or reverse. This is not an internal inconsistency in the math, but it is a missing validation that is essential to the paper's headline claim. The corrupted placeholder strings in Section V-B and the absence of code or detailed experimental settings further weaken reproducibility, but the QoE metric issue is the most load-bearing technical concern. I therefore agree with the reader's assessment and recommend no change to the REJECT verdict.","tokens_in":13396,"tokens_out":3040,"duration_ms":31030,"concrete_test":"Re-run the Section V-C evaluation with a QoE metric that replaces bitrate utility with a fidelity-aware term, e.g., m(ak)=VMAF(reconstructed_chunk, source_chunk), while keeping smoothness and rebuffering penalties unchanged. If LD-ABS no longer outperforms the baselines on FCC/3GPP/OBOE traces, then the reported QoE advantage is an artifact of Eq. (1). A complementary check: report PSNR/SSIM/VMAF of LDM-reconstructed frames vs H.264/HEVC at the same transmitted bitrate for a sample of chunks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LD-ABS delivers high QoE and bandwidth savings while preserving visual quality. This is tested only through the bitrate-based QoE metric in Eq. (1) (Section V-C), where the bitrate utility m(ak)=log(ak/min(A)) is used as a proxy for perceived quality. For LD-ABS, however, ak is the bitrate of compressed I-frame latents plus P/B motion metadata; the actual video is reconstructed on the client by an LDM conditioned on that metadata. Nothing in the simulation measures the fidelity of those reconstructed frames: no PSNR, SSIM, or VMAF is reported, and no comparison is made between LDM reconstruction and conventional decoding at the same bitrate. If the diffusion reconstruction introduces blur, artifacts, or temporal flicker, then the true user-perceived quality at bitrate ak is lower than the utility assumed in Eq. (1), and the reported QoE advantage over BOLA/MPC/Pensieve could be an artifact of the metric. The claim 'high-quality video streaming' therefore rests on an unverified equivalence between bitrate and reconstructed perceptual quality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LD-ABS, a semantic-aware adaptive bitrate video streaming framework that uses a latent diffusion model (LDM) to compress I-frames into latent representations, transmits P/B frames as compact motion metadata, and reconstructs frames at the client with conditional denoising and video frame interpolation. A CNN-GRU based controller selects the next chunk's bitrate based on channel estimates, buffer state, and content features. The authors evaluate the framework against BOLA, RobustMPC, Comyco, and Pensieve using a simulated player on FCC, 3GPP, and OBOE traces, reporting higher QoE and lower end-to-end latency.","tokens_in":13590,"tokens_out":3753,"duration_ms":41755,"significance":"If the claims were substantiated, the paper would address a timely problem: reducing the bandwidth footprint of adaptive streaming through generative semantic coding while preserving quality. The conceptual decomposition into latent I-frames plus motion metadata is a plausible direction, and the explicit use of I/P/B frame structure and video frame interpolation is a useful framing. However, the paper currently provides no implementation details, no training or evaluation code, and no measurement of the reconstructed video's perceptual quality. The reported QoE advantages rest entirely on a bitrate-based utility function, which does not validate the central 'high-quality video streaming' claim. The work is therefore at the level of an architecture proposal with an encouraging but unverified simulation story.","major_comments":[{"comment":"The QoE metric in Eq. (1) uses bitrate utility m(ak)=log(ak/min(A)) as a proxy for perceived quality, but for LD-ABS the transmitted bitrate ak consists of compressed I-frame latents plus B/P motion metadata, while the actual video is generated on the client by the LDM. The paper reports no PSNR, SSIM, VMAF, or any other measure of the reconstructed frames, and no rate-distortion comparison with conventional decoding at equal reconstructed quality. The abstract's claim of 'high-quality video streaming' is therefore not established; Fig. 5 and the '<3% QoE loss' sentence only reflect the simulation's bitrate-based objective.","section":"Section V-C / Eq. (1)"},{"comment":"No implementation details are provided for the VAE, denoising U-Net, conditional LDM, or the CNN-GRU selector: no architecture specifics beyond channel counts, no training data, no loss functions, no hyperparameters, no values for r, s, T, and no channel or noise model. Consequently, the numerical results in Section V cannot be reproduced or independently checked, and it is unclear which components are newly contributed and which are taken unchanged from reference [4].","section":"Section IV-B / Section V-A"},{"comment":"The comparison with BOLA, RobustMPC, Comyco, and Pensieve is not on equal terms: those baselines transmit conventional coded video bitstreams, whereas LD-ABS transmits latents and metadata whose bitrate does not correspond to the bitrate of the reconstructed video. Evaluating both on the same bitrate-utility QoE function can favor LD-ABS simply because it sends smaller chunks, even if the LDM-reconstructed frames have lower visual quality. A fair comparison requires measuring the perceptual quality of the final reconstructed video and comparing at matched quality levels.","section":"Section V-C / Fig. 5"},{"comment":"The claim 'With CSI error, LD-ABS loses <3% QoE, four times less than the closest competitor' is not backed by a described experiment: no channel error model, SNR regime, error percentage, or trace modification is given, and no error bars or confidence intervals are shown in Fig. 5. This specific quantitative assertion is therefore unverifiable as written.","section":"Section V-C"},{"comment":"The CNN-GRU bitrate selector is described conceptually, but the paper does not state how it is trained, what data or expert strategies fill the offline replay buffer, or how the test traces relate to the training distribution. Since the selector is trained to optimize the same QoE metric used in the evaluation, the reported improvement may largely reflect alignment with the simulator's objective; held-out evaluation with multiple random seeds and a train/test separation is needed to support a general policy claim.","section":"Section IV-A / Section V-C"}],"minor_comments":[{"comment":"The baseline description in Section V-B contains a long unreadable encoded character sequence (the '/uni00000026...' line) that obscures part of the text, and Figure 4's caption and table formatting appear corrupted. These must be repaired before publication.","section":"Section V-B"},{"comment":"There are several typos and formatting inconsistencies, including 'Recontruction' in the Section IV heading, 'cahnnel' in Section IV-B(iv), and inconsistent spacing in 'V AE'; a careful proofreading pass is needed.","section":"Throughout"},{"comment":"Figure 4(a) is described as showing chunk QoE, bitrate utility, rebuffering penalty, and smoothness penalty, but the individual panels are not clearly labeled; please add explicit panel labels and axis titles.","section":"Figure 4(a)"},{"comment":"The QoE weights alpha=1 and beta=2.66 are taken from [6], but no sensitivity analysis is reported; given that the entire evaluation uses this metric, it would be useful to show how the relative ranking changes for reasonable alternative weights.","section":"Eq. (1) / Section V-C"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you can skip this one unless you have a tolerant reading group. The paper's architecture is a plausible integration of two known pieces: an LDM-based semantic codec from the authors' prior IEEE TWC paper and a neural ABR controller in the Pensieve/Comyco style, with a frame-type twist (I-frames go to latents, P/B frames become motion metadata, conditional LDM and VFI do the reconstruction). That is a real, if modest, extension of the existing program, and the survey of diffusion-based video streaming is competent.\n\nWhat the paper does not do is show that the system actually delivers high-quality video. The evaluation is a simulation that only reports the standard ABR QoE metric from Eq. (1), which is a function of bitrate, bitrate changes, and rebuffer time. For LD-ABS the bitrate is the compressed latent + metadata, and the visual output is produced by a generative decoder that is never evaluated. No PSNR, SSIM, or VMAF, no comparison of LDM reconstruction against a conventional decoder at the same bitrate. The claim that the method 'outperforms state-of-the-art' is thus an artifact of treating bitrate as a proxy for quality, which is exactly the assumption the method is supposed to invalidate. If the diffusion reconstruction is blurry or flickery, the whole advantage disappears.\n\nThere are other soft spots. Section V-B contains a long corrupted placeholder string; that is an editing failure that should have been caught before submission. I also found no equations for the diffusion process, no training details, no dataset, no code. The latency and QoE figures are plausible but not reproducible. The '<3% QoE loss with CSI error' claim is not backed with enough details to judge.\n\nCredit where due: the idea of adapting the denoising starting step to channel state and anchoring B/P-frame reconstruction to the refined I-frame is reasonable and could be a useful design principle. The survey is serviceable. But the paper is a position paper with a simulation sketch, not a verified system.\n\nMy recommendation: desk reject. This preprint does not deserve referee time until the authors provide direct measurements of reconstructed visual quality, a clearer system description, and a fixed text. If they address that, the idea might be worth a workshop or a short paper. For you personally, no need to read it unless you are tracking the sem-comm video area.","headline":"A plausible frame-type-aware integration of LDM compression with a neural ABR controller, but the central QoE claim is unsupported because no measurement of reconstructed video quality is reported.","tokens_in":14181,"tokens_out":3193,"would_cite":false,"duration_ms":32151,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LD-ABS transmits only compressed I-frame latents and motion metadata, then reconstructs video with latent diffusion, achieving higher QoE than BOLA, RobustMPC, Comyco, and Pensieve on FCC, 3GPP, and OBOE wireless traces.","keywords":["semantic communication","latent diffusion models","adaptive bitrate streaming","quality of experience","video frame interpolation","FFmpeg","wireless networks"],"falsifier":"Reconstruct video with LD-ABS at its reported operating points and compare the output frames against the original encoded video using a perceptual fidelity metric such as LPIPS or VMAF on high-motion clips; if fidelity is markedly worse while the Eq. 1 QoE score stays high, the central QoE claim does not reflect visual quality. A complementary test is to feed the system systematically biased, not just random, CSI errors and check whether the reported under-3-percent QoE loss still holds.","tokens_in":13150,"feed_emoji":"📹","tokens_out":9022,"duration_ms":80663,"temperature":0.7,"pith_summary":"The paper argues that adaptive-bitrate video streaming over wireless can be made far more efficient by exploiting the I/P/B frame structure semantically: send I-frames as compact latent representations and P/B frames as lightweight motion metadata, then reconstruct the video at the client with a latent diffusion model. Because only these compressed artifacts cross the wireless link, bandwidth and storage demands drop, and the receiver's denoiser is positioned to absorb channel noise by starting its reverse diffusion from a point tied to channel state. Using a CNN-GRU controller to pick the next chunk's bitrate, the framework reports higher QoE than BOLA, RobustMPC, Comyco, and Pensieve on FCC, 3GPP, and OBOE traces, and under 3 percent QoE loss when channel state information is imperfect. If correct, this points toward generative reconstruction, not just higher bitrates, as the lever for streaming quality in 5G and future wireless networks.","feed_headline":"Latent diffusion streams video as I-frame latents plus motion data","feed_subtitle":"LD-ABS tops BOLA, RobustMPC, Comyco, and Pensieve in QoE with fewer bytes on the air.","key_machinery":"The load-bearing mechanism is a frame-type-aware conditional latent diffusion pipeline. A VAE maps each I-frame into a compact latent L-I-frame for transmission, while P/B frames are reduced to light motion-vector adjustment metadata; a short reverse-diffusion run, started at a step chosen from the estimated channel state, denoises the received L-I-frame, and a conditional LDM then regenerates the P/B frames using the motion metadata as the conditioning signal, with video frame interpolation keeping the sequence temporally coherent. On the network side, a 1D-CNN plus GRU controller picks the next chunk's bitrate from channel estimates, latent content size, and playback state. This pipeline is what converts bandwidth savings into the reported QoE gains, and it is also what limits damage from channel errors: denoising happens in a low-dimensional latent space in few steps, and non-key frames are anchored to the already-refined I-frame.","core_discovery":"The paper claims that the standard I/P/B structure of compressed video can be turned into a semantic transmission scheme. Instead of sending full frames at a chosen bitrate, the transmitter encodes I-frames into low-dimensional latent representations with a VAE, transmits only those latents plus compact motion-vector metadata for P and B frames, and lets a latent diffusion model finish the job at the receiver: it denoises the received I-frame latents, with the reverse-diffusion starting point set by channel state information, then reconstructs P/B frames conditioned on the motion metadata, assisted by video frame interpolation. In simulations on FCC, 3GPP, and OBOE throughput traces, with low- and high-motion content, the paper reports that this LD-ABS framework yields higher average chunk QoE than BOLA, RobustMPC, Comyco, and Pensieve, and that under CSI error its QoE loss stays below 3 percent, about four times smaller than the closest competitor.","pith_inferences":["Editorial inference: the reported QoE advantage would be more convincing if the metric also scored perceptual or pixel fidelity of the reconstructed frames, since Eq. 1 is blind to diffusion artifacts; adding LPIPS, VMAF, or a user study could change the ranking.","Editorial inference: the paper itself notes that its LDMs are trained on a specific dataset and may lack generalization across diverse content and network conditions, so the reported QoE levels may not carry over to live news or sports streaming without domain adaptation.","Editorial inference: bandwidth savings and reconstruction difficulty both depend on motion; on action-heavy or camera-panning content the motion metadata grows and the conditional diffusion must work harder, so the reported margins should be stress-tested on high-motion clips beyond the low/high split shown.","Editorial inference: the sub-3 percent QoE loss under CSI error is demonstrated for the tested error model; if real-world CSI errors are biased or bursty, the denoising start point would be miscalibrated, so the robustness claim needs validation under richer channel models."],"forward_implications":["On the paper's evidence, a streaming client can reconstruct watchable video while the network only carries I-frame latent tensors and compact P/B motion metadata, so per-chunk bandwidth and CDN storage shrink relative to sending full-frame encodings at the same quality level.","Since the reverse diffusion runs only a few steps in a low-dimensional latent space, the reported end-to-end latency, about 2 to 3 seconds per chunk, stays within practical bounds while earlier pixel-space diffusion approaches take 5 to 8 seconds.","Channel errors matter less than in conventional ABS: with CSI error the framework loses under 3 percent QoE, roughly four times less than its closest benchmark, because the few-step latent denoising limits error propagation and P/B reconstruction is anchored to the cleaned I-frame.","The CNN-GRU bitrate selector produces an adaptive policy that can be evaluated on the same virtual player and trace sets used for BOLA, RobustMPC, Comyco, and Pensieve, making the comparison direct."],"supporting_citations":[{"why":"Supplies the latent diffusion semantic-communication building block, namely VAE encoding, diffusion denoising, and channel-noise handling, that LD-ABS adapts to I-frame latents.","marker":"[4]"},{"why":"Provides the QoE metric in Eq. 1, the penalty weights alpha=1 and beta=2.66, and the benchmark algorithms and trace sets used for comparison.","marker":"[6]"},{"why":"Provides the latent-diffusion video frame interpolation method used to restore temporal coherence between reconstructed frames.","marker":"[10]"},{"why":"Is the Comyco neural ABR baseline and a source of the virtual-player evaluation setup used in the experiments.","marker":"[15]"},{"why":"Is CaDM, the codec-aware diffusion streaming method that LD-ABS extends by preserving the I/P/B frame structure.","marker":"[1]"}],"fun_headline_variants":["Latent diffusion video: send I-frame latents, not full frames","LD-ABS: I-frame latents plus motion beat adaptive bitrate baselines","Diffusion-based streaming: latents cut bytes, boost QoE","Semantic video via latent diffusion: I latents, P/B motion, high QoE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparisons rest on a QoE score that rewards high bitrate, smooth switches, and no stalls but never measures the actual visual fidelity of the diffusion-reconstructed frames, so if those frames show artifacts or blur, the reported QoE advantage may not survive.","fun_headline_variants_meta":{"raw":{"variants":["Latent diffusion video: send I-frame latents, not full frames","LD-ABS: I-frame latents plus motion beat adaptive bitrate baselines","Diffusion-based streaming: latents cut bytes, boost QoE","Semantic video via latent diffusion: I latents, P/B motion, high QoE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000526,"raw_usage":{"total_tokens":2545,"prompt_tokens":956,"completion_tokens":1589,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1503}},"tokens_in":572,"tokens_out":1589,"duration_ms":12703,"temperature":1.0,"reasoning_tokens":1503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:19:19.266055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct video with LD-ABS at its reported operating points and compare the output frames against the original encoded video using a perceptual fidelity metric such as LPIPS or VMAF on high-motion clips; if fidelity is markedly worse while the Eq. 1 QoE score stays high, the central QoE claim does not reflect visual quality. A complementary test is to feed the system systematically biased, not just random, CSI errors and check whether the reported under-3-percent QoE loss still holds.","supporting_citations":[{"cited_title":"Latent diffusion model-enabled real-time semantic communication considering semantic ambiguities and channel noises,","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion semantic-communication building block, namely VAE encoding, diffusion denoising, and channel-noise handling, that LD-ABS adapts to I-frame latents."},{"cited_title":"Improving generalization for neural adaptive video streaming via meta reinforce- ment learning,","cited_arxiv_id":null,"evidence_quote":"Provides the QoE metric in Eq. 1, the penalty weights alpha=1 and beta=2.66, and the benchmark algorithms and trace sets used for comparison."},{"cited_title":"Ldmvfi: Video frame interpolation with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the latent-diffusion video frame interpolation method used to restore temporal coherence between reconstructed frames."},{"cited_title":"Comyco: Quality-aware adaptive video streaming via imitation learning,","cited_arxiv_id":null,"evidence_quote":"Is the Comyco neural ABR baseline and a source of the virtual-player evaluation setup used in the experiments."},{"cited_title":"CaDM: Codec-aware Diffusion Modeling for Neural-enhanced Video Streaming","cited_arxiv_id":"2211.08428","evidence_quote":"Is CaDM, the codec-aware diffusion streaming method that LD-ABS extends by preserving the I/P/B frame structure."}],"review_version":1}