{"id":"25d42488-ced2-41d7-8f65-d250e953d645","arxiv_id":"2507.14851","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"RONIN distills per-frame language descriptions of video degradations into lightweight input-conditioned prompts, achieving all-in-one video restoration without any text encoder or MLLM at inference and outperforming prior methods on four benchmarks.","lead":"This paper introduces RONIN, a video restoration method that uses a language model to describe each frame's degradations, then learns to produce the same guidance from the frame so the language model is not needed at inference. It also proposes standardized multi-degradation benchmarks and a new time-varying snow dataset, reporting state-of-the-art results across them.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Language grounding is never isolated: Eq. 3's text-embedding loss is not ablated, so the reported gains may be due to the learned prompt mechanism alone.","rationale":"The reader's weakest assumption focuses on Q-Instruct description accuracy and the validity of text embeddings as targets. That is a real issue, and Tab. 10 supports it. My concern is adjacent but more structural: even granting that the descriptions are accurate, the paper never demonstrates that the text-embedding target contributes to performance. The prompt is generated from latent features and first-encoder cross-attention; the text loss may simply be an auxiliary objective that is largely ignored (λ2=0.01) or that introduces noise. Ablations in Tab. 5 and the appendix show that prompt injection matters and that prompts align with text embeddings, but neither establishes causality for the language signal. If a λ2=0 ablation matches RONIN, the central contribution reduces to a PromptIR-style learned prompt plus a cross-attention injection, and the interpretability claim is largely retrospective (the inference-time prompt is a continuous vector, not text). The fixed candidate-degradation list in Algorithm 1 also weakens the 'no degradation knowledge' framing, though that is secondary. I would keep the reader's CONDITIONAL verdict: the method is promising and the comparisons are extensive, but the headline mechanism needs one direct ablation before the language-grounding claim can be accepted. Hence verdict_should_be is CONDITIONAL and agreement with the reader is partial.","tokens_in":21138,"tokens_out":7285,"duration_ms":83597,"concrete_test":"Train the identical RONIN architecture on the 3D benchmark with λ2=0 in Eq. 3, and in a second arm keep the loss but replace et(Ct) with shuffled same-norm text embeddings. Compare deblur/denoise/derain PSNR-SSIM to the published RONIN. If both text-target arms are statistically indistinguishable from λ2=0 (e.g., within 0.05 dB average), the language-grounding objective is not load-bearing and the method should be reframed as a learned-prompt method; if removing or shuffling the target drops average PSNR by a meaningful margin (e.g., >0.3 dB), the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"RONIN's central claim is that language-grounded prompts, learned by approximating frozen text embeddings et(Ct) in Eq. 3, improve all-in-one video restoration. The only language supervision is the second loss term with λ2=0.01, and Pt is produced from latent features plus first-encoder cross-attention (Eq. 1, Fig. 1), so the network could learn a blackbox-like prompt from image features alone. No experiment isolates this term. Tab. 5 ablates prompt placement but the 'No Prompt' setting retains the cross-attention and only removes injection, so it tests the injection module, not the text target. Supplementary tSNE, cosine-similarity, and perturbed-prompt tests show learned prompts are meaningful and the decoder depends on them, but a prompt trained purely by restoration loss would also be necessary and would also separate degradations; the high cosine similarity (0.985-0.991) is expected after training with the L1 target. Given Tab. 10 shows substantial MLLM mislabeling (e.g., 13,471 noise mentions and 10,549 blur mentions in desnow frames), the target is noisy, and at λ2=0.01 the term may be a weak, neutral, or even harmful regularizer. Without a λ2=0 or shuffled-target ablation, the causal role of language grounding in the SOTA numbers is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents RONIN, an all-in-one video restoration framework. At training time, the multimodal LLM Q-Instruct generates per-frame degradation descriptions, a frozen text encoder (BGE-Micro-v2) converts them into embeddings, and a lightweight prompt generator (Eq. 1) is trained with an L1 loss (Eq. 3) to approximate those embeddings from the latent features of a Turtle-style U-Net. The resulting prompt is injected into the last two decoder stages through a learned soft mask. At inference, only the restoration network and prompt generator remain, with no MLLM or text encoder. The authors assemble and evaluate on four benchmarks—3D (deblur/denoise/derain), 4D (adds desnow), TUD (time-varying noise/blur/compression), and a new SnowyScenes benchmark with time-varying snow—and report PSNR/SSIM improvements over several image/video restoration baselines, together with qualitative results, prompt-perturbation tests, t-SNE, and cosine-similarity analyses.","tokens_in":21368,"tokens_out":7945,"duration_ms":86774,"significance":"If the central claim is established, this is a useful and timely contribution: the design decouples MLLM/text-encoder use from inference while preserving interpretable, language-style conditioning, and the proposed SnowyScenes benchmark addresses an underexplored time-varying weather setting. The paper is also transparent about annotation failures (Tab. 10) and limitations (Sec. 11), and the benchmark-standardization effort is valuable. However, the causal role of the language-grounding term is not isolated by the current experiments, and the 'no degradation knowledge' framing is in tension with the fixed candidate-degradation prompt procedure in Algorithm 1.","major_comments":[{"comment":"The paper does not isolate the language-grounding term. The prompt P_t is generated from latent features plus first-encoder cross-attention (Eq. 1, Fig. 1), so it could carry degradation information even if the text-embedding target e_t(C_t) were removed. The 'No Prompt' row in Tab. 5 only disables decoder injection; it keeps the prompt-generation path, and therefore cannot distinguish a learned blackbox-style prompt from a language-grounded one. The cosine-similarity and perturbed-prompt analyses in Appendix 7 show the prompt is informative and necessary, but any input-conditioned prompt trained only by restoration loss would also be informative and necessary. Given that Tab. 10 shows noisy targets (e.g., 13,471 noise mentions and 10,549 blur mentions in desnow frames), and λ2=0.01 is small, the text term may even be near-neutral. Please add (i) a λ2=0 baseline, (ii) a shuffled or mismatched text-target control, and (iii) a non-text control target (e.g., fixed random embeddings) to demonstrate that the language-aligned target, rather than the prompt mechanism, causes the reported gains.","section":"§3.3, Eq. (3), Tab. 5"},{"comment":"The abstract and Tab. 1 claim RONIN assumes no degradation knowledge at train or test time, but the annotation procedure in Algorithm 1 explicitly queries Q-Instruct with a fixed set of candidate degradations and thresholded Yes/No intensity questions. This means the benchmark's degradation vocabulary is known a priori and hard-coded into the training-data pipeline; a degradation outside the candidate set would not be queried and might not appear in the open-ended initial description. The claim should be qualified to 'no per-sample degradation labels at inference,' or the authors should provide evidence that the open-ended initial description alone recovers all relevant degradations, including those not in the candidate list.","section":"§10, Algorithm 1; §3.1"},{"comment":"AverNet appears in the 3D block of Tab. 2 but is absent from the 4D block without any explanation. Since AverNet is one of the two video-specific baselines and the 4D benchmark is a central contribution, this omission makes the 'state-of-the-art on all benchmarks' claim incomplete. Please add the missing results, or explicitly state and justify why they are not reported, and keep the method list consistent across benchmark blocks.","section":"Tab. 2, §4.1"}],"minor_comments":[{"comment":"Eq. (1) has an unmatched parenthesis and the dimension 'Rb×d' should be written as R^{b×d}; please fix the notation.","section":"§3.2, Eq. (1)"},{"comment":"The Fig. 6 caption says 'all tasks of the 4D benchmark,' but the figure shows SnowyScenes results; correct the caption.","section":"Fig. 6"},{"comment":"In Tab. 2, the 3D block appears to include four task columns, including Desnow, although Sec. 4.1 defines 3D as three tasks; some rows also lack an Average entry. Please align the columns with the benchmark definitions and report averages consistently.","section":"Tab. 2"},{"comment":"Tab. 10 is labeled 'Robustness Analysis' but reports only counts of degradation mentions in the MLLM text; rename it to something like 'Degradation Mention Statistics' and move the actual robustness evidence to an experiment that varies target noise or compares cleaned versus uncleaned targets.","section":"§10.1, Tab. 10"},{"comment":"Tab. 7 uses the same asterisk symbol for two different exclusions (optical flow and text encoder); use distinct markers and clarify the note.","section":"Tab. 7"},{"comment":"In Sec. 4.2, the +0.23 dB improvement over AverNet holds for Set8; on the DAVIS test set RONIN is below AverNet on PSNR, so the abstract's unqualified 'state-of-the-art on all benchmarks' should be qualified by metric and test split.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The missing λ2=0 / shuffled-target ablation is, in my view, mandatory before the central claim can be accepted; without it, the paper's main novelty—that language grounding, rather than the learned prompt mechanism, drives the gains—is not verified. The other required changes are clarifications and table corrections. I do not see a reproducibility or integrity concern: the paper openly documents annotation noise and limitations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RONIN is a solid all-in-one video restoration paper that deserves a serious referee, but the headline claim — that natural-language grounding is what drives the gains — is not yet proven by the experiments. The design is genuinely useful: per-frame MLLM descriptions are generated offline, distilled into input-conditioned prompts, and the MLLM and text encoder are removed at inference. That is a practical step beyond InstructIR-type methods. SnowyScenes is a reasonable new benchmark for time-varying snow, and the push to standardize 3D/4D video restoration evaluation is overdue. The benchmark comparisons are broad: 3D, 4D, TUD, SnowyScenes, with PSNR/SSIM and LPIPS, plus ablations on prompt placement, prompt perturbation, tSNE, and cosine similarity. The MLLM misclassification table (Tab. 10) and limitations section are honest and useful rather than buried.\n\nThe main soft spot is the missing ablation of the prompt approximation loss (lambda2 in Eq. 3). Nothing isolates the text-embedding target. The No Prompt row in Tab. 5 removes only the prompt injection, while the cross-attention from the first encoder remains; a prompt trained purely by restoration loss could in principle produce the same gains. The perturbation, tSNE, and cosine-similarity evidence show the prompts are meaningful and needed, but they do not show that the language target is the cause of the improvement. At lambda2=0.01 the loss may be little more than a gentle regularizer. A lambda2=0 run and a shuffled-label control would settle this. Without it, the claim that language grounding lifts performance is plausible but unverified.\n\nAlso minor: calling the method 'no degradation knowledge' is overstated because Algorithm 1 explicitly queries a fixed candidate list (noise, rain, blur, snow, haze). AverNet is missing from the 4D table without comment. Code and SnowyScenes data are not public yet, though the authors commit to release. The MLLM description noise shown in Tab. 10 is acknowledged; one experiment showing robustness to that noise would strengthen the causal story.\n\nBottom line: the engineering and benchmarking contributions are real, the paper is readable, and the authors are honest about limitations. It should go to peer review, but reviewers should ask for the missing ablation before publication. I would cite it for the benchmarks and the distillation design; whether I rely on the language-grounding claim depends on that ablation.","headline":"Solid all-in-one video restoration with practical offline distillation and useful new benchmarks, but the causal role of the language grounding is not isolated by the ablations.","tokens_in":21931,"tokens_out":2392,"would_cite":true,"duration_ms":25745,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that per-frame natural-language degradation descriptions can be distilled into a lightweight prompt generator, so an all-in-one video restoration network can be deployed with no multimodal language model or text encoder…","keywords":["all-in-one video restoration","language-grounded prompts","multimodal large language models","knowledge distillation","time-varying degradations","SnowyScenes benchmark","video restoration benchmarks"],"falsifier":"Train RONIN identically but replace each frozen grounded text embedding with a random or permuted embedding from the same text encoder; if PSNR/SSIM on 3D and 4D stay at RONIN levels, then the language grounding itself is not carrying the reported gains. A complementary check is to feed deliberately wrong descriptions (for example, noise captions for deblurring frames) and see whether performance drops roughly as much as the paper's prompt-perturbation experiment suggests.","tokens_in":20908,"feed_emoji":"🎬","tokens_out":5310,"duration_ms":50487,"temperature":0.7,"pith_summary":"This paper tries to establish that the degradation information needed to restore a video can be captured in natural language during training and then distilled into the restoration network so thoroughly that the language model is no longer needed at inference. The proposed method, RONIN, queries a multimodal large language model (MLLM) to describe the degradations in each frame, converts those descriptions into frozen text embeddings, and trains a small prompt generator to approximate them while learning to restore the frame. A sympathetic reading of the results is that this language grounding gives the network interpretable per-frame conditioning, lets it handle composite and time-varying degradations without knowing what they are, and outperforms prior all-in-one image and video restoration methods on the benchmarks tested. The paper also standardizes all-in-one video restoration evaluation with 3D and 4D benchmarks plus a new SnowyScenes benchmark with time-varying snow.","feed_headline":"Video restoration ditches the language model at test time","feed_subtitle":"RONIN trains on per-frame degradation captions, then restores blur, noise, rain, and snow on its own with no extra cost.","key_machinery":"The load-bearing device is the prompt approximation objective (Eq. 3), which combines a restoration loss with a small L1 loss forcing the generated prompt $P_t$ to match the frozen text embedding $e_t(C_t)$ from the grounded description. The prompt itself is produced from latent frame features by a two-layer projection after global average pooling, with cross-attention from the first encoder stage to reintroduce degradation information, and it is injected into the last two decoder stages as a sigmoid-gated per-channel soft mask. This setup separates training support from the deployed network: the language pipeline is only needed offline, while inference runs on the prompt generator and U-Net-style backbone alone.","core_discovery":"RONIN grounds each degraded frame in natural language: a multimodal large language model (Q-Instruct) produces a per-frame description of image quality and degradations, a text encoder converts it into an embedding, and during training the restoration network must generate a prompt that is close to that embedding while also restoring the frame. At inference, the MLLM and text encoder are removed, and the prompt generator alone conditions the decoders through a per-channel soft mask. The paper's central claim is that this distilled language grounding is what allows one network to handle blur, noise, rain, snow, and time-varying mixtures of them, with no degradation assumption at train or test time and no extra inference cost, and that it achieves state-of-the-art PSNR/SSIM on 3D, 4D, TUD, and SnowyScenes benchmarks.","pith_inferences":["Beyond the paper: if the improvement persists when the grounded text embeddings are replaced by random embeddings from the same encoder, that would suggest the gains come from the input-conditioned prompt structure rather than the language grounding itself; the paper's own misclassification table makes this a live distinction.","If language grounding is what carries the gain, the same training-time distillation could transfer to other conditioning signals such as region-level quality descriptions or expert-written degradation templates.","A testable extension would be to corrupt the MLLM descriptions (for example, feeding noise descriptions to deblurring frames) and measure the drop in PSNR; the paper predicts a substantial drop, whereas a pure blackbox-prompt account predicts little change."],"forward_implications":["All-in-one video restoration can be deployed without an MLLM or text encoder at inference, removing per-frame language querying and its cost.","Because conditioning is per frame and text is generated automatically, the method can handle composite degradations and time-varying degradation without any label of which degradation is present.","The 3D, 4D, TUD, and SnowyScenes results imply that a single network can restore blur, noise, rain, and snow with better or comparable PSNR/SSIM than prior image and video restoration baselines.","The proposed SnowyScenes benchmark extends time-varying degradation evaluation to weather conditions, giving a common testbed for future methods."],"supporting_citations":[{"why":"Q-Instruct supplies the per-frame degradation descriptions that ground each prompt in natural language.","marker":"[49]"},{"why":"Turtle provides the streaming U-Net backbone and training recipe that RONIN builds on.","marker":"[8]"},{"why":"InstructIR is the whitebox text-conditioned baseline that RONIN contrasts against to show its text encoder can be removed.","marker":"[6]"},{"why":"AverNet introduces the time-varying unknown degradation (TUD) benchmark and is the strongest video baseline on it.","marker":"[55]"},{"why":"PromptIR is the blackbox prompt baseline that RONIN compares against and whose decoder-prompt injection style it adapts.","marker":"[36]"},{"why":"AirNet contributes the contrastive all-in-one restoration baseline and the multi-degradation benchmark convention used for 3D and 4D.","marker":"[18]"},{"why":"VRDS supplies the video deraining dataset in the 3D and 4D benchmarks.","marker":"[48]"},{"why":"DAVIS supplies the video denoising test set with Gaussian noise for the 3D, 4D, and TUD benchmarks.","marker":"[35]"},{"why":"GoPro supplies the video deblurring data and is one of the two sources for SnowyScenes.","marker":"[32]"},{"why":"RVSD supplies the video desnowing and dehazing data used in the 4D benchmark.","marker":"[2]"}],"fun_headline_variants":["Video restoration ditches LLM at test time after language grounding","Language-grounded training, LLM-free inference for all-in-one video restoration","Learn degradation captions, then restore with no language model at runtime","Ground video degradations in text, then drop the text model at test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The largest assumption is that the multimodal language model's per-frame degradation descriptions are accurate enough that their frozen text embeddings are a trustworthy training target; the paper itself counts misclassifications (noise and blur appearing in descriptions for desnowing and deblurring clips), and if those descriptions are systematically wrong, the prompt approximation loss may train prompts toward misleading language and the interpretability benefit would weaken.","fun_headline_variants_meta":{"raw":{"variants":["Video restoration ditches LLM at test time after language grounding","Language-grounded training, LLM-free inference for all-in-one video restoration","Learn degradation captions, then restore with no language model at runtime","Ground video degradations in text, then drop the text model at test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"total_tokens":2787,"prompt_tokens":853,"completion_tokens":1934,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":1857}},"tokens_in":469,"tokens_out":1934,"duration_ms":17942,"temperature":1.0,"reasoning_tokens":1857,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:45:26.302011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train RONIN identically but replace each frozen grounded text embedding with a random or permuted embedding from the same text encoder; if PSNR/SSIM on 3D and 4D stay at RONIN levels, then the language grounding itself is not carrying the reported gains. A complementary check is to feed deliberately wrong descriptions (for example, noise captions for deblurring frames) and see whether performance drops roughly as much as the paper's prompt-perturbation experiment suggests.","supporting_citations":[{"cited_title":"Learning truncated causal history model for video restoration","cited_arxiv_id":null,"evidence_quote":"Turtle provides the streaming U-Net backbone and training recipe that RONIN builds on."},{"cited_title":"Avernet: All-in-one video restoration for time-varying unknown degradations","cited_arxiv_id":null,"evidence_quote":"AverNet introduces the time-varying unknown degradation (TUD) benchmark and is the strongest video baseline on it."},{"cited_title":"Promptir: Prompting for all-in-one image restoration","cited_arxiv_id":null,"evidence_quote":"PromptIR is the blackbox prompt baseline that RONIN compares against and whose decoder-prompt injection style it adapts."},{"cited_title":"All-in-one image restoration for unknown corruption","cited_arxiv_id":null,"evidence_quote":"AirNet contributes the contrastive all-in-one restoration baseline and the multi-degradation benchmark convention used for 3D and 4D."},{"cited_title":"Mask-guided progressive network for joint raindrop and rain streak removal in videos","cited_arxiv_id":null,"evidence_quote":"VRDS supplies the video deraining dataset in the 3D and 4D benchmarks."},{"cited_title":"Deep multi-scale convolutional neural network for dynamic scene deblurring","cited_arxiv_id":null,"evidence_quote":"GoPro supplies the video deblurring data and is one of the two sources for SnowyScenes."},{"cited_title":"Snow removal in video: A new dataset and a novel method","cited_arxiv_id":null,"evidence_quote":"RVSD supplies the video desnowing and dehazing data used in the 4D benchmark."}],"review_version":1}