{"id":"8d113e28-e049-4afb-8f1a-fa5d28010064","arxiv_id":"2501.09350","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A zero-shot pipeline claims to generate dream videos from sleep fMRI by transferring a model trained on awake visual perception, with only three dream segments validated.","lead":"This paper claims to turn fMRI recordings of sleeping people into video stories of their dreams. It reuses an existing brain-to-image model trained on awake vision, then asks a language model to narrate the decoded frames, but the validation rests on a tiny dataset and weak similarity measures.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on untested zero-shot transfer from 7T awake NSD data to 3T sleep fMRI, and the only quantitative test is confounded by subject identity: positive and negative sets are different participants, so higher CLIP similarity may reflect subject-specific output biases rather than…","rationale":"The reader's verdict of REJECT is appropriate, and our stress-test pass identifies the same weak point but sharpens it: the zero-shot transfer assumption is not merely unproven; the paper's only quantitative evidence is structured so that it cannot distinguish dream-content decoding from subject-identity decoding. Section 6.5 defines positive samples as the dream-reporting participant's sleep data and negative samples as other participants' data. Because the NSD-trained model has never seen these participants, its outputs may carry large subject-specific biases (scanner field strength, head shape, physiological noise, sleep stage), and CLIP similarities to the dream label could track those biases. The authors' own exclusion of two dreams due to 'unavoidable domain gaps' (Section 6.1.2) further signals that the transfer is not robust. A within-subject comparison using Sub-1's and Sub-3's two dreams would settle the question and is a natural, cheap control. We do not object to the novelty of collecting dream fMRI data or to the pipeline idea; the concern is specifically that the evidence does not support the strong claim of high-fidelity dream reconstruction. We therefore keep the reader's REJECT verdict (UNCHANGED).","tokens_in":10588,"tokens_out":3632,"duration_ms":37166,"concrete_test":"Run a within-subject CLIP similarity analysis on the collected dream data, as the authors' Table 1 already provides two dreams for Sub-1 and two dreams for Sub-3. For each subject, take all decoded images I_dream_i from the two dream segments, compute CLIP similarity to both dream captions (e.g., 'some cats' vs 'some fruits, with plenty of grapes' for Sub-3), and test whether the correct caption wins within each segment (e.g., Mann-Whitney U or sign test on per-image similarities). If the within-subject discrimination fails, the cross-subject result in Table 2 is explained by subject identity rather than dream content, and Eq. (3) is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Eq. (3), I_dream_i = G(E(F_i)), recovers dream content from sleep fMRI. This requires the NSD-trained encoder E and generator G to transfer to 3T sleep data from new participants. The paper offers no adaptation or validation of this transfer, and Section 6.1.2 admits that two of five dreams were excluded due to 'unavoidable domain gaps,' indicating that the transfer is not robust. More importantly, the only quantitative evidence (Section 6.5, Table 2) is structurally confounded. Positive samples are all sleep-fMRI volumes from the participant who reported the dream; negative samples are volumes from the other participants. Thus the reported higher CLIP similarity for positive samples can be explained by any subject- or session-specific bias in the zero-shot outputs (for example, the model producing systematically different image distributions for each new subject), without any dream content being decoded. The within-subject control that would resolve this is available in the authors' own data: Sub-1 has two distinct dream reports ('skiing with a snowboard' and 'enjoying a cup of milk tea'), and Sub-3 has two ('some cats' and 'some fruits, with plenty of grapes'). If the method truly decodes dream imagery, decoded images from the 'cats' segment should score higher on 'cat' than decoded images from the 'fruits' segment for the same subject, and vice versa. The paper does not report this test; without it, the central claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage pipeline for turning sleep fMRI into a coherent dream video: (1) train a NeuroPictor-style fMRI-to-image reconstruction model on the 7T Natural Scenes Dataset, (2) apply this model zero-shot to 3T sleep fMRI windows via Eq. (3), I_dream_i = G(E(F_i)), to decode dream snapshots, and (3) use LENS image captioning and ChatGPT to assemble the snapshots into a narrative video. The authors collected 3.3 hours of sleep fMRI from three participants, with five clearly recalled dream reports, and analyze three of those segments. Quantitative evaluation in Section 6.5 compares CLIP similarity of decoded images to dream labels between positive and negative samples, followed by Mann-Whitney U tests.","tokens_in":10880,"tokens_out":3693,"duration_ms":81244,"significance":"If the main claim were established, this would be a noteworthy first step toward high-fidelity visual decoding of dream content and a novel multimedia application. The paper builds on the public NSD benchmark and an existing reconstruction model (NeuroPictor), and the authors report a new sleep-fMRI dataset, which are useful resources. However, the current evidence does not support the central claim: the evaluation is confounded by subject identity, the analysis is selective, no baselines are reported, and the narrative integration step is circular as evidence of decoding fidelity. The scientific claim of directly decoding dream images from fMRI therefore remains unsubstantiated.","major_comments":[{"comment":"The positive and negative sample comparison is confounded by participant identity. For a given dream label, positive samples come from the participant who reported that dream while negative samples come exclusively from the other participants. For example, for 'cat', positives are Sub-3's sleep volumes and negatives are Sub-1's and Sub-2's volumes. Higher CLIP similarity of the positive set could therefore reflect subject-specific output biases of the zero-shot decoder (e.g., a tendency to generate certain image statistics for a given subject) rather than decoded dream content. The paper does not report the necessary within-subject control, even though it is available in the authors' own data: Sub-1 has two distinct dream reports and Sub-3 has two distinct dream reports. Without such a control, the reported p-values in Table 2 do not test whether dream content was decoded.","section":"Section 6.5, Table 2"},{"comment":"Two of the five clearly recalled dream segments were excluded post hoc with the stated reason of 'unavoidable domain gaps and complex semantics.' This is a selection on the dependent variable and it removes exactly the data needed for the within-subject controls described above. The quantitative evaluation is therefore based on only three segments, and the exclusion criteria are not principled or pre-registered, which further weakens the support for the general claim that the method decodes dream imagery.","section":"Section 6.1.2"},{"comment":"No baseline or chance level is reported for the CLIP similarity evaluation. The paper only compares positive and negative samples to each other. To establish decoding, the authors would need to compare against baselines such as decoded images from temporally permuted fMRI volumes, images from a model with scrambled or shuffled fMRI inputs, or a chance-level similarity distribution. Additionally, the p-values in Table 2 are not corrected for multiple comparisons, and no sample sizes, effect sizes, or per-subject breakdowns are reported, so the statistical claims in Section 6.5 are incomplete.","section":"Section 6.3 and Section 6.5"},{"comment":"The central decoding pipeline assumes that the NSD-trained encoder E and generator G, trained on 7T awake visual stimulus data, transfer zero-shot to 3T sleep fMRI from unseen participants without any adaptation. The paper offers no validation of this cross-scanner, cross-state transfer. Section 6.1.2's admission that two of five segments failed due to domain gaps indicates that the transfer is not robust. At minimum, the authors would need a control experiment showing that the identical pipeline behaves appropriately on matched 3T awake data, or an explicit adaptation step. Without this, Eq. (3) rests on an untested assumption.","section":"Section 5.2, Eq. (3)"},{"comment":"The narrative integration stage is circular as evidence for decoding fidelity. LENS first generates captions from the decoded dream images, and ChatGPT then composes a coherent story from those captions. The vividness and coherence of the final narrative reflect the language model's generative capabilities, not the accuracy of the neural decoding. Consequently, the qualitative claims in Section 6.4 that the resulting videos 'align with participants' descriptions' cannot be taken as support for the central claim that dream images were directly decoded from fMRI.","section":"Section 5.3 and Section 6.4"}],"minor_comments":[{"comment":"There is a typo in 'fMRI encider'; it should be 'fMRI encoder.'","section":"Equation (1)"},{"comment":"The set of augmented COCO class labels is not specified. The authors should report the exact label list, including how 'skis' was derived from the dream caption 'skiing with a snowboard' and how other non-COCO labels were formed.","section":"Section 6.3"},{"comment":"The figure caption lacks axis labels, error bars, and a per-subject or per-segment breakdown, making it difficult to assess the magnitude and variability of the reported similarities.","section":"Figure 8"},{"comment":"The table reports only p-values. It should also report the number of positive and negative samples, the mean and standard deviation of the similarity scores, and the effect size for each comparison.","section":"Table 2"},{"comment":"The statement that inference used a single GTX 3090Ti GPU is not an implementation detail relevant to the method; please provide instead the fMRI window size used for temporal averaging, the number of decoded frames per dream segment, and the exact inference settings.","section":"Section 6.2"},{"comment":"The phrase 'brave new idea' is informal and repeated throughout; the footnote distinguishing the colloquial use of 'dreams' from neurology's precise sleep-stage definitions should be integrated into the main text for scientific clarity.","section":"Abstract and Section 4 footnote"}],"recommendation":"reject","confidential_remarks":"The rejection is driven by the structurally confounded evaluation and the post hoc exclusion of data, not by disagreement with the novelty of the goal. The central claim would need substantial new evidence, including within-subject analyses on the existing data and meaningful baselines, before this is suitable for a journal. The reader's stress-test concern accurately identifies the main flaw: the positive-negative comparison cannot separate dream content from subject identity, and the paper's own data contain the controls needed to address this, but those controls are not reported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine first attempt at decoding dream imagery into video narratives, and the task formulation is new. The pipeline is clear: train NeuroPictor on 7T awake NSD data, apply it zero-shot to 3T sleep fMRI, caption the decoded frames with LENS, then let ChatGPT stitch them into a story. The authors are upfront that this is work in progress, and the dream dataset collection effort (3.3 hours of sleep fMRI with dream reports) is nontrivial. That part deserves credit.\n\nThe soft spots are load-bearing, though. Eq. (3) assumes the 7T awake multi-subject encoder transfers to 3T sleep data from new participants without adaptation. The paper gives no validation of that transfer, and Section 6.1.2 quietly admits two of five dreams were dropped due to 'unavoidable domain gaps.' That is a red flag, not a detail. More serious, the only quantitative test (Table 2, Fig. 8) is confounded by subject identity: positive samples come from the dreamer, negatives from the other participants. Any subject- or session-specific bias in the zero-shot outputs would produce the same pattern. The within-subject control is sitting in their own Table 1 (Sub-1 has two dreams, Sub-3 has two dreams) but they do not run it. Without that, the CLIP similarities support a subject effect, not dream decoding. The p-values come from small, non-independent samples, and there is no chance-level baseline. The ChatGPT narrative integration is circular as evidence of coherence, though acceptable as a generation method.\n\nThe paper is coherent on its own terms and honest about its exploratory status. I would not desk-reject it: the task is novel, the framework is reproducible enough to serve as a baseline, and the flaws are fixable in principle. I would send it to review with a clear request for the within-subject analysis, proper baselines, and no post-hoc exclusions. As it stands, the central claim is unsupported, but the paper is a reasonable starting point for a hard problem.","headline":"A genuine first attempt at decoding dream imagery into video narratives, but the central claim rests on an unvalidated zero-shot transfer and a subject-confounded evaluation; worth a serious referee, not acceptance as-is.","tokens_in":11411,"tokens_out":1330,"would_cite":false,"duration_ms":15144,"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 dreams can be reconstructed as viewable video frames directly from sleep fMRI, without dream-report labels.","keywords":["fMRI decoding","dream decoding","visual imagery","zero-shot transfer","diffusion models","large language models","sleep fMRI","video narrative generation"],"falsifier":"Decode sleep fMRI after shuffling the temporal order of the volumes within each dream segment, or decode fMRI recorded during quiet wakefulness with the same model: if the resulting images stay equally aligned with the dream descriptions, the pipeline is responding to generic model priors rather than dream-specific neural signal.","tokens_in":10360,"feed_emoji":"🧠","tokens_out":8144,"duration_ms":76278,"temperature":0.7,"pith_summary":"This paper tries to establish that the visual content of a dream can be recovered from sleep fMRI signals as viewable images, without relying on category-level dream labels. It builds on the premise that dreaming and waking perception share brain activity patterns in visual cortex, so a reconstruction model trained on awake visual-stimulus data can be transferred directly to sleep data. The authors propose a three-stage pipeline: reconstruct real visual stimuli from a large awake fMRI dataset, decode each sleep fMRI time point into a dream snapshot with the same model, then use language models to stitch the snapshots into a coherent narrated video. If the central claim holds, dreams become a new input modality for multimedia generation and a scientific window into subjective visual experience.","feed_headline":"Dreams decoded into video directly from sleep fMRI","feed_subtitle":"A model trained on awake brain scans reconstructs dream snapshots and weaves them into a narrated story.","key_machinery":"The load-bearing machinery is the cross-state transfer of a multi-subject fMRI-to-image reconstruction model: a transformer encoder $E$ maps fMRI surface images to a unified brain representation, and a diffusion-based generative network $G$, composed of a high-level guiding network and a low-level manipulation network, turns that representation into an image. Trained on thousands of awake visual-stimulus fMRI-image pairs, the same $G(E(\\cdot))$ is applied unchanged to averaged sleep fMRI windows in Eq. (3). The shared-activity-pattern assumption is what licenses this zero-shot transfer, and the language-model integration stage then converts the discrete decoded frames into a story.","core_discovery":"The paper's central claim is that Eq. (3), $I_i^{\\text{dream}} = G(E(F_i))$, yields a high-fidelity reconstruction of what a person was visually dreaming at time $i$: apply the pre-trained reconstruction function $G$ to the sleep fMRI volume $F_i$ encoded by $E$, with no training on dream data. The authors argue that because the same visual cortex processes both real visual stimuli and dream imagery, the shared brain activity patterns allow the awake-trained model to generalize zero-shot to sleep-state fMRI. This would surpass prior dream decoding, which only classified dream content into semantic categories such as objects, by producing direct visualizations of the dream experience. They further claim that assembling these decoded snapshots into a narrative video through large language models captures the continuous story-like nature of dreams rather than isolated frames.","pith_inferences":["Until the zero-shot cross-scanner transfer is validated, a cautious reader should treat the decoded images as a blend of learned visual prior and sleep signal, because a strong generative prior can produce plausible scenes even from weak or noisy input.","A direct control experiment that decodes sleep fMRI with the temporal order of volumes scrambled, or decodes quiet-wakefulness fMRI with the same model, would separate dream-specific neural signal from generic prior-driven generation.","The narrative integration step may inflate apparent success, since a fluent language model can weave unrelated captions into a coherent story; story coherence should be assessed separately from frame-level decoding accuracy.","Because the quantitative evaluation uses text-image similarity against the participant's own dream caption, a forced-choice test among several candidate dream scenes would be a stronger check of whether the decoded frames truly carry the reported dream content."],"forward_implications":["Dream content can be visualized directly from sleep fMRI, bypassing the need for dream reports as training labels.","The same awake-trained reconstruction model can be carried across individuals and brain states, so dream decoding does not require per-person retraining on dream data.","Decoded dream frames, once captioned and narrated by language models, can be assembled into a video that reflects the reported dream sequence.","The reported quantitative results indicate that decoded frames from positive dream segments match the corresponding dream description better than frames from other participants, with significant differences for two of the three tested dream labels.","Dream experiences are treated as continuous narratives rather than isolated snapshots, aligning the decoding target with how dreams are actually recalled."],"supporting_citations":[{"why":"Supplies the large 7-tesla awake visual-stimulus fMRI-image pairs used to train the reconstruction model.","marker":"[1]"},{"why":"Establishes the premise that dream content shares visual brain activity patterns with stimulus perception and can be decoded from sleep fMRI.","marker":"[20]"},{"why":"Provides the multi-subject pre-trained fMRI-to-image reconstruction approach that the paper transfers zero-shot to dream decoding.","marker":"[21]"},{"why":"Provides the large-scale multi-subject fMRI encoder that lets the reconstruction model learn shared brain representations across individuals.","marker":"[29]"},{"why":"Describes the 3-tesla sleep fMRI data acquisition protocol followed for the collected dream dataset.","marker":"[12, 13]"},{"why":"Supplies the zero-shot text-image similarity model used to score decoded dream frames against dream descriptions.","marker":"[30]"},{"why":"Supplies the language-model-driven script and video narration pipeline adapted to turn dream snapshots into a story.","marker":"[37]"}],"fun_headline_variants":["Zero-shot dream video from sleep fMRI","Dream movies decoded from sleep fMRI","Awake-trained model yields dream video from fMRI","Video story from sleep fMRI without dream training","Dreams become watchable videos via fMRI zero-shot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The awake-trained model is assumed to work on sleep fMRI of different people recorded in a different scanner and brain state, without any adaptation.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot dream video from sleep fMRI","Dream movies decoded from sleep fMRI","Awake-trained model yields dream video from fMRI","Video story from sleep fMRI without dream training","Dreams become watchable videos via fMRI zero-shot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2680,"prompt_tokens":873,"completion_tokens":1807,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":1741}},"tokens_in":489,"tokens_out":1807,"duration_ms":16072,"temperature":1.0,"reasoning_tokens":1741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:05:23.371781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Decode sleep fMRI after shuffling the temporal order of the volumes within each dream segment, or decode fMRI recorded during quiet wakefulness with the same model: if the resulting images stay equally aligned with the dream descriptions, the pipeline is responding to generic model priors rather than dream-specific neural signal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the premise that dream content shares visual brain activity patterns with stimulus perception and can be decoded from sleep fMRI."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multi-subject pre-trained fMRI-to-image reconstruction approach that the paper transfers zero-shot to dream decoding."},{"cited_title":"Intelligent Director: An Automatic Framework for Dynamic Visual Composition using ChatGPT","cited_arxiv_id":"2402.15746","evidence_quote":"Supplies the language-model-driven script and video narration pipeline adapted to turn dream snapshots into a story."}],"review_version":1}