{"id":"c2444e1e-1177-4248-9736-c40cdcb65b64","arxiv_id":"2506.19465","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Reverse-stylized neural fractals, synthetic images generated by random complex networks then textured with features from a small real-image set, reduce the domain gap to real images and improve pre-training for autoencoding, diffusion, and classification.","lead":"This paper builds synthetic images from random complex-valued neural networks, calling them neural fractals, then transfers texture patterns from about 7,000 real photos onto them with neural style transfer. It reports that models pre-trained on these reverse-stylized images improve autoencoder reconstruction, diffusion generation, and classification accuracy compared with other synthetic datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper contradicts itself on which encoder is used for reverse stylization (Sec 3.2 says pre-trained on Unsplash; Sec 4.1 says the autoencoder trained on synthetic data), so the claimed mechanism of real-feature transfer is unverified.","rationale":"The reader's weakest assumption is that the 7k Unsplash sample's feature statistics generalize to the target domains. That is a valid representativeness concern. My concern is more fundamental: the paper internally contradicts itself about which encoder is used to compute the style-transfer features. Section 3.2 claims an Unsplash-pretrained encoder, while Section 4.1 indicates the encoder comes from an autoencoder trained on synthetic datasets. If the latter is true, the stylization is not using real-data-trained features, and the claim that reverse stylization 'transfers visual features from a small, license-free set of real images' is misleading. The downstream improvements might still occur, but the causal story — that the gains arise from aligning with real-image statistics via an encoder that has learned those statistics — would be unsupported. This is a concrete, verifiable inconsistency in the manuscript itself, and it sits at the heart of the method's claimed mechanism. The reader's verdict is already CONDITIONAL due to missing artifacts, error bars, and overstatement; this additional concern reinforces that conditioning rather than overturning it. The empirical results could survive scrutiny, but the method as described needs clarification and re-verification before the central claim can be accepted. I therefore recommend keeping the verdict CONDITIONAL (UNCHANGED), with the added condition that the encoder identity be resolved and the stylization re-evaluated accordingly.","tokens_in":17073,"tokens_out":9694,"duration_ms":102207,"concrete_test":"Check the released code or the authors' clarification to identify the exact encoder used in the reverse-stylization pipeline (the autoencoder trained on synthetic data vs. a separate encoder pre-trained on the 7k Unsplash images). If the former, re-run the stylization with a genuinely Unsplash-trained encoder and reproduce Table 5 (autoencoder loss, diffusion FID, DINO accuracy) and the KID values in Table 1. If the results change materially (e.g., the 11% Flowers FID improvement and 10.7% ImageNet-100 accuracy gain shrink or reverse), then the paper's stated mechanism is not what produced the reported numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that reverse stylization transfers visual features from a small set of real images to synthetic data, reducing the domain gap and improving downstream pre-training. The paper's stated mechanism depends on the encoder used for style transfer. Section 3.2 explicitly says: 'we replace it with an encoder network pre-trained on Unsplash to prevent any biases ... arising from using a network trained on large-scale real data.' However, Section 4.1, describing the autoencoder experiment, states: 'Importantly, we also use its encoder as a VGG replacement to generate features for style transfer' — where 'its' refers to the autoencoder trained on synthetic datasets in that section. These statements are mutually contradictory. If the actual implementation used the synthetic-trained autoencoder encoder, then the reverse stylization is not using real-image-trained features as claimed, and the mechanism by which the domain gap is reduced is different from (and potentially weaker than) the one described. The reported KID improvement and downstream gains may still hold, but the explanation that they arise from transferring statistics learned on the 7k Unsplash images is unsupported. This is a correctness risk central to the paper's contribution, not merely a reproducibility detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-step method for generating synthetic pretraining datasets. First, it introduces an improved neural fractal generator: a complex-valued MLP recursion with adaptive sampling, escape-time coloring, and automatic threshold adjustment. Second, it applies 'reverse stylization,' using neural style transfer to transfer visual features from a small set of 7,000 license-free Unsplash images onto the synthetic images. The authors evaluate the resulting datasets on a domain-gap measure (KID/FID vs. ImageNet) and on three downstream tasks: autoencoder reconstruction, EDM2 diffusion fine-tuning (Flowers, FFHQ), and DINO-based representation learning (ImageNet-100, Flowers, Food101). They report consistent improvements for stylized neural fractals over existing synthetic datasets, including roughly 19-24% lower autoencoder reconstruction loss, an 11% FID reduction on Flowers diffusion fine-tuning, and about 10.7 percentage points higher ImageNet-100 accuracy for a ViT-S trained with DINO.","tokens_in":17286,"tokens_out":6725,"duration_ms":66740,"significance":"If the results are reproducible, the paper offers a practical contribution to synthetic pretraining: it shows that a small, license-clean real-image set can be used to inject realistic visual statistics into synthetic data, narrowing the domain gap and improving performance on generative as well as discriminative downstream tasks. The evaluation is broad, covering multiple base synthetic datasets, three task families, and ablations such as dataset scaling, style-transfer algorithm choice, and a MixUp comparison. The reverse-stylization technique appears to generalize across all tested base datasets, which strengthens the claim that it is a general tool rather than a quirk of neural fractals. The paper does not ship code or release the datasets, however, and the central mechanism is undermined by internally contradictory statements about which feature encoder is used for the stylization step.","major_comments":[{"comment":"The specification of the feature encoder used for reverse stylization is contradictory. Section 3.2 states that the VGG network is replaced by 'an encoder network pre-trained on Unsplash to prevent any biases ... arising from using a network trained on large-scale real data,' while Section 4.1, in the autoencoder experiment, says 'Importantly, we also use its encoder as a VGG replacement to generate features for style transfer,' where 'its' refers to the autoencoder trained on synthetic data in that section. This is not a minor wording issue: the proposed mechanism for reducing the domain gap is that real-image statistics (from Unsplash) are transferred to synthetic images. If the encoder is instead an autoencoder trained on synthetic fractals, the claim that the improvement comes from transferring statistics learned on the 7k Unsplash images is unsupported. Please clarify which encoder was actually used for the stylization experiments. If it is the synthetic-trained autoencoder, revise the mechanism description and provide additional experiments that isolate the effect of encoder choice on the downstream results and on KID/FID.","section":"Sec. 3.2 vs. Sec. 4.1"},{"comment":"The neural fractal generator architecture is described inconsistently. Section 3.1 says 'We use a fully connected neural network with 6 hidden layers of 3 neurons each,' but Section 4.3 states 'we use a 3-layer, 6-neuron network,' and the supplementary material similarly reports 'a network with 3 layers and 6 neurons.' Since the network g in Eq. (1) determines the fractal structure of every rendered image, this discrepancy directly affects the reproducibility of the proposed synthetic dataset. Please correct the inconsistency and specify the precise architecture used to generate the reported 100K-image datasets.","section":"Sec. 3.1 vs. Sec. 4.3 and Appendix"},{"comment":"All experimental results are reported as single numbers with no error bars, standard deviations, or number of seeds. This is a particular concern because several claimed improvements are modest in magnitude (e.g., FID 18.3 to 16.3 on Flowers, DINO accuracy 48.5 to 59.2 on ImageNet-100). Without multiple seeds or variance estimates, it is impossible to determine whether the observed differences are statistically reliable or within training noise. Please provide results averaged over at least three seeds with standard deviations, or otherwise justify the stability of the single-run numbers, for the key comparisons in Tables 1, 3, 4, 5, and 6.","section":"Tables 1-6 (single-run results)"},{"comment":"The claim that reverse stylization reduces the domain gap is partly circular. The stylization objective—Gram-matrix matching of encoder features—explicitly aligns feature statistics of the synthetic images with those of the real Unsplash images, and the reported KID/FID are computed on Inception features. Although the evaluation target is ImageNet rather than Unsplash, the reduction in KID/FID is to a large degree the direct objective of the method, so it does not independently confirm that the method transfers useful visual information. The authors should either frame the KID/FID result as a sanity check rather than as evidence, or include a control experiment (for example, stylizing with an encoder trained on random images, or matching random feature statistics) to demonstrate that the KID reduction is not a trivial consequence of matching features.","section":"Sec. 4.3 'Domain Gap' and Table 10"}],"minor_comments":[{"comment":"The number of training iterations for the autoencoder is inconsistent: Section 4.1 says 'We train the network for 1M iterations,' while the supplementary material says 'We train the AE ... for 1.5M iterations.' Please align these numbers.","section":"Sec. 4.1 vs. Appendix Sec. 8"},{"comment":"The DINO training length is inconsistent: the main text states 'We train the encoder for 1000 epochs,' but the appendix specifies 200 epochs for the ViT-S encoder. Please reconcile.","section":"Sec. 4.1 vs. Appendix Sec. 8"},{"comment":"The Flowers FID for the 100K non-stylized neural fractal dataset is 18.3 in Table 3 and 18.4 in Table 6; please correct the discrepancy.","section":"Table 3 vs. Table 6"},{"comment":"The abstract says 'over a 10% improvement' in ImageNet-100 accuracy, but Table 5 reports a +10.7 percentage-point increase (48.5 to 59.2). Please state whether this is a relative improvement or a percentage-point difference to avoid ambiguity.","section":"Abstract and Table 5"},{"comment":"The text says KID was chosen because it 'has the ability to report confidence intervals,' but no confidence intervals are reported anywhere in the paper; either provide them or remove this justification.","section":"Sec. 4.3"},{"comment":"The attention-map cosine similarity metric is not defined: the paper does not specify which layer or attention head is used, how attention maps are extracted, or how similarity is averaged over images. Please provide a precise definition in the main text or appendix so that the numbers in Table 1 are reproducible.","section":"Table 1 and App. Sec. 9"},{"comment":"There are several typographical errors and inconsistent terminology, including 'signficant' (Sec. 2.2), 'generate generate' (Sec. 3.2), 'peform' (Sec. 4.3), 'pipline' (appendix), and the alternation between 'ImageNet-100' and 'ImageNet-100k.' Please proofread and standardize.","section":"Throughout"},{"comment":"The statement that Unsplash images are 'license-free' should be qualified: the Unsplash License permits use but imposes conditions (e.g., no compilation of the images to replicate a similar service). Please verify and describe the license terms more precisely.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the downstream evaluations are extensive, but the manuscript currently has a load-bearing internal contradiction about which encoder is used for reverse stylization (Sec. 3.2 vs. Sec. 4.1), as well as an inconsistency in the neural fractal architecture (Sec. 3.1 vs. Sec. 4.3). These must be resolved and the corresponding experiments clarified or rerun. The lack of error bars is also a significant weakness for a paper whose claims are quantitative. The KID/FID framing is partly circular and should be repositioned. I do not see an unresolvable flaw, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe two things you should know: (1) this is a surprisingly effective recipe for making synthetic pre-training data—neural fractals plus reverse stylization gives real gains across autoencoding, diffusion, and DINO pre-training; (2) the paper's own explanation of why it works is internally inconsistent, and the inconsistency sits right at the load-bearing point.\n\nWhat's new: the improved neural fractal renderer (adaptive sampling, escape-time coloring, auto threshold) and the specific application of neural style transfer in reverse—stylizing synthetic images with features from a small real-image set. The empirical results are consistent and often large: ~20% lower AE loss, 11% lower FID on Flowers diffusion, 10.7% higher ImageNet-100 accuracy for ViT-S trained with DINO, and KID drops from 0.162 to 0.120. They also show stylization helps other synthetic baselines, not just their own fractals. That's a useful recipe, especially for data-scarce settings. The ablations (MixUp comparison, dataset scaling) are good practice.\n\nNow the soft spots. The big one is the style-encoder contradiction. Sec 3.2 says they replace VGG with an encoder pre-trained on Unsplash to avoid biases from ImageNet-trained features. Sec 4.1, describing the autoencoder experiment, says they use 'its encoder'—the autoencoder trained on synthetic data—as the VGG replacement for style transfer. These are mutually exclusive. If the real implementation used the synthetic-trained encoder, then the claimed mechanism—transferring statistics learned from 7k real images—is not what's happening. The KID reduction and downstream gains might still hold, but the explanation would be wrong, and the method would lose its conceptual grounding. This needs to be clarified, and the actual implementation needs to be checked.\n\nOther issues are more standard: no seeds or error bars anywhere, so we can't tell if the 11% on Flowers is noise; the abstract highlights the 11% diffusion gain without mentioning that FFHQ only moves 1%; the attention-map cosine similarity metric is never fully defined; and the KID-reduction evidence is partly circular because the stylization is optimized to match feature statistics—though the downstream results are independent and do carry weight.\n\nVerdict: this deserves a serious referee, but not acceptance as-is. The reviewer should ask for multi-seed runs, a precise description of the style-transfer encoder, and ideally code/data release. If the encoder contradiction turns out to be a typo and the Unsplash-pretrained encoder was really used, the empirical story is solid. If not, the paper needs to be rethought.\n\nReading group: maybe—I'd want to see the encoder issue sorted first. Would I cite it? Possibly, if they release code. Otherwise the recipe is hard to reproduce.\n\nSo: send it out for review, with a major-revision expectation.","headline":"A genuinely new synthetic-pre-training recipe with real empirical gains, but the central mechanism contains an encoder contradiction that must be resolved before the claimed explanation can be trusted.","tokens_in":17837,"tokens_out":3958,"would_cite":false,"duration_ms":34195,"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":"Synthetic fractal images that absorb texture statistics from a small set of real photos come measurably closer to real-image data, improving pretraining for reconstruction, generation, and representation learning.","keywords":["neural fractals","synthetic data pretraining","reverse stylization","neural style transfer","domain gap","Kernel Inception Distance","diffusion model pretraining","self-supervised representation learning"],"falsifier":"Train the same reverse-stylization pipeline with a deliberately mismatched style set, such as only grayscale or only indoor images, and recompute KID to ImageNet plus the three downstream metrics. If KID stays near 0.120 and the downstream gains persist, representativeness of the style sample is not the deciding factor; if KID rises and the gains vanish, the premise is confirmed. A cheaper check is to ablate the style-set size, for example 7, 70, 700, and 7,000 images, and locate where the benefits actually appear.","tokens_in":16864,"feed_emoji":"🌀","tokens_out":9207,"duration_ms":86862,"temperature":0.7,"pith_summary":"The paper tries to show that synthetic images can be made substantially more useful for pretraining computer-vision models by giving them visual statistics borrowed from a small set of real photos. Its recipe is two-step: generate diverse fractal images from randomly initialized complex-valued neural networks, then run reverse stylization, transferring texture and color features from about 7,000 license-free photos onto the fractals. This step cuts the distribution gap to ImageNet, measured by Kernel Inception Distance, from 0.162 to 0.120. Across three evaluation pipelines, models pretrained on the stylized data improve over models trained on non-stylized neural fractals: roughly 20% lower autoencoder reconstruction loss, 11% lower FID in diffusion fine-tuning on Flowers, and 10.7 percentage points higher top-1 accuracy for a DINO-trained ViT-S on ImageNet-100.","feed_headline":"Fractals with real-photo textures narrow the synthetic-real gap","feed_subtitle":"Texture from 7,000 license-free photos narrows the synthetic-real gap and lifts downstream tasks.","key_machinery":"The argument rides on two pieces. First, improved neural fractals: instead of the Mandelbrot recurrence $z_{n+1}=z_n^2+c$, a randomly initialized complex-valued neural network $g$ defines $z_{n+1}=g(z_n)+c$, and escape-time coloring plus variance-guided adaptive Monte Carlo sampling render each unique pattern quickly. Second, reverse stylization: texture-based style transfer, either Gatys Gram-matrix matching or neural-neighbor matching, is run with a real photo as the style and a fractal as the content, using an encoder pretrained on about 7,000 Unsplash images rather than ImageNet. The lowered KID between the resulting dataset and real images is presented as the measurable mechanism connecting the reduced domain gap to downstream pretraining performance.","core_discovery":"The paper claims that reverse stylization, running neural style transfer with real photos as the style source and synthetic fractals as content, transfers enough real-image texture and color statistics onto synthetic data to substantially close the distribution gap. In their measurements, the Kernel Inception Distance between the generated dataset and ImageNet drops from 0.162 for plain neural fractals to 0.120 for the stylized version, and networks trained on the stylized data match the attention behavior of ImageNet-trained networks more closely. Downstream, this yields about 20% lower autoencoder reconstruction loss, 11% lower FID after diffusion fine-tuning on Flowers, and 10.7 percentage points higher ImageNet-100 accuracy for a DINO-trained ViT-S relative to the same models trained on non-stylized neural fractals. The authors also report that reverse stylization improves every base synthetic dataset they tried, indicating that the mechanism is not specific to fractals.","pith_inferences":["If the gains are driven mainly by transferable texture statistics rather than content semantics, a stylization source could eventually be synthesized or learned, removing the need for any photographed reference set.","The 7,000-image Unsplash set is generic; a natural test is whether a domain-matched style set, such as medical, satellite, or indoor images, is needed when the downstream target is a specialized domain.","The appendix shows stylization helps weaker base datasets even more, which suggests reverse stylization could act as a drop-in improvement for future procedural data generators and may compound with dataset-scaling laws.","The attention-map cosine-similarity results imply the benefit may extend beyond distribution matching into the inductive bias of the pretrained network, an effect worth checking across other architectures such as CNNs and larger vision transformers.",""],"forward_implications":["Autoencoder pretraining on reverse-stylized neural fractals reconstructs real images with about 20% lower loss than on plain neural fractals, making synthetic-only pretraining more viable for perceptual compression.","Diffusion models pretrained on stylized fractals and fine-tuned on Flowers improve FID by about 11%, and scaling the stylized set to 1M images reduces FID by roughly another 30%.","A DINO-trained ViT-S pretrained without labels on stylized fractals gains 10.7 percentage points of top-1 accuracy on ImageNet-100 over plain fractals, narrowing the gap toward models pretrained on curated real data.","Reverse stylization improves all tested base synthetic datasets, so it should transfer to other procedural or noise-based data generators, not just fractals.","Because the style-reference set is small and license-free, the method offers a path to pretraining corpora with fewer privacy and copyright constraints.",""],"supporting_citations":[{"why":"Introduces the neural-dynamical-system fractals whose rendering the paper improves with escape-time coloring, adaptive sampling, and automatic threshold selection.","marker":"[3]"},{"why":"Establishes the procedural and statistical-noise synthetic pretraining baselines and the 100K-image generation protocol the paper follows.","marker":"[4]"},{"why":"Provides the DINO self-supervised pipeline used to evaluate representation learning with a ViT-S encoder.","marker":"[9]"},{"why":"Serves as the real-image reference for KID and FID domain-gap measurements and as an evaluation target.","marker":"[11]"},{"why":"Supplies the Gram-matrix style-transfer objective that reverse stylization runs in reverse to inject real-image texture statistics into synthetic images.","marker":"[13]"},{"why":"Provides the EDM2 diffusion framework used to evaluate pretraining quality through fine-tuned FID on Flowers and FFHQ.","marker":"[17]"},{"why":"Supplies the neural-neighbor style transfer used as the second, often stronger, stylization algorithm for the synthetic datasets.","marker":"[20]"},{"why":"Defines the autoencoder architecture used to measure reconstruction loss and whose encoder is repurposed as the style-transfer feature extractor.","marker":"[24]"}],"fun_headline_variants":["Reverse stylization narrows synthetic-to-real gap","Real-photo textures on fractals lift downstream tasks","Neural fractals + real textures beat existing synthetic data","Stylized fractals reduce domain gap, improve ImageNet accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the texture and color statistics extracted from roughly 7,000 Unsplash photos are representative enough of the real-image domains, such as ImageNet, COCO, Flowers, and FFHQ, that transferring those statistics onto synthetic fractals actually narrows the gap; if this small sample carries the wrong statistics, the reported gains would shrink or reverse.","fun_headline_variants_meta":{"raw":{"variants":["Reverse stylization narrows synthetic-to-real gap","Real-photo textures on fractals lift downstream tasks","Neural fractals + real textures beat existing synthetic data","Stylized fractals reduce domain gap, improve ImageNet accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1465,"prompt_tokens":975,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":424}},"tokens_in":591,"tokens_out":490,"duration_ms":5306,"temperature":1.0,"reasoning_tokens":424,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:32:00.406472+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same reverse-stylization pipeline with a deliberately mismatched style set, such as only grayscale or only indoor images, and recompute KID to ImageNet plus the three downstream metrics. If KID stays near 0.120 and the downstream gains persist, representativeness of the style sample is not the deciding factor; if KID rises and the gains vanish, the premise is confirmed. A cheaper check is to ablate the style-set size, for example 7, 70, 700, and 7,000 images, and locate where the benefits actually appear.","supporting_citations":[{"cited_title":"Neuralfractal - a visual exploration of neu- ral dynamical systems","cited_arxiv_id":null,"evidence_quote":"Introduces the neural-dynamical-system fractals whose rendering the paper improves with escape-time coloring, adaptive sampling, and automatic threshold selection."},{"cited_title":"Learning to see by looking at noise","cited_arxiv_id":null,"evidence_quote":"Establishes the procedural and statistical-noise synthetic pretraining baselines and the 100K-image generation protocol the paper follows."},{"cited_title":"Emerg- ing properties in self-supervised vision transformers","cited_arxiv_id":null,"evidence_quote":"Provides the DINO self-supervised pipeline used to evaluate representation learning with a ViT-S encoder."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"Serves as the real-image reference for KID and FID domain-gap measurements and as an evaluation target."},{"cited_title":"Analyzing and improving the training dynamics of diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the EDM2 diffusion framework used to evaluate pretraining quality through fine-tuned FID on Flowers and FFHQ."},{"cited_title":"Neural neighbor style transfer","cited_arxiv_id":null,"evidence_quote":"Supplies the neural-neighbor style transfer used as the second, often stronger, stylization algorithm for the synthetic datasets."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Defines the autoencoder architecture used to measure reconstruction loss and whose encoder is repurposed as the style-transfer feature extractor."}],"review_version":2}