A Neural Algorithm of Artistic Style

A Neural Algorithm of Artistic Style Leon A. Gatys Alexander S. Ecker Matthias Bethge

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

A Neural Algorithm of Artistic Style 2015 arXiv preprint 1508.06576 gatys2015-style-transfer gatys2015-style-transfer.txt, the project's staged copy 1a62e230dce3258bdb715c2b08baac4175d9a30b arxiv-open art-history

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them existence Quotation withheld because this record is not cleared for redistribution. separation of image content and style in CNN representations Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms separating and recombining the content and style representations that a CNN learns across its layers Quotation withheld because this record is not cleared for redistribution. image content and style cannot be completely disentangled what this limits: any claim that the method separates content from style: OF COURSE, IMAGE CONTENT AND STYLE CANNOT BE COMPLETELY DISENTANGLED Quotation withheld because this record is not cleared for redistribution. Content and style are not fully independent, so the separation is only partial. what this limits: the same scope as limitations[0] - the separation is partial, and the paper says so in its own framing Quotation withheld because this record is not cleared for redistribution. No image satisfies both the content and the style constraint perfectly; a trade-off has to be tuned per image pair. what this limits: every output image: THERE USUALLY DOES NOT EXIST AN IMAGE THAT PERFECTLY MATCHES BOTH CONSTRAINTS AT THE SAME TIME, so the result is a trade-off tuned per image pair Quotation withheld because this record is not cleared for redistribution. What the paper calls style is a texture statistic that discards the composition of the image, in a feature space built for something else. what this limits: any reading of this system's style as a style in the art-historical sense: it is computed in a feature space originally designed to capture texture information, and the resulting representation captures texture but not the global arrangement of the image Quotation withheld because this record is not cleared for redistribution. The headline separability claim is demonstrated in the objective function, which is where the two terms are separated. what this limits: the claim that content and style are separable: what is shown is that the minimised loss function contains two well-separated terms for content and style, which is a property of the objective the authors wrote Quotation withheld because this record is not cleared for redistribution. Which image you get is set by a balance the user picks for each pair of sources. what this limits: any output shown here as a result of the method: for a specific pair of source images the trade-off between content and style is adjusted by hand to create visually appealing images Quotation withheld because this record is not cleared for redistribution. The use for perception research is projected, not carried out. what this limits: the claim that this method is a tool for studying visual perception: the authors envision it being useful for such studies, from psychophysics to neural recordings, rather than reporting one Quotation withheld because this record is not cleared for redistribution. convolutional neural networks trained for object recognition, repurposed to synthesize artistic images Quotation withheld because this record is not cleared for redistribution. style style Quotation withheld because this record is not cleared for redistribution. style representation style representation Quotation withheld because this record is not cleared for redistribution. content representation content representation Quotation withheld because this record is not cleared for redistribution. Gram matrix Gram matrix Quotation withheld because this record is not cleared for redistribution. feature space feature space Quotation withheld because this record is not cleared for redistribution. style - AS CORRELATIONS BETWEEN FILTER RESPONSES: a Gram matrix over a layer's feature maps, which is a texture statistic and not an art-historical category, and which by construction throws away where things are in the picture Quotation withheld because this record is not cleared for redistribution. content - AS A LAYER'S FEATURE MAPS: what can be reconstructed from the network's responses in a particular layer, so 'the content of an image' means whatever survives that reconstruction Quotation withheld because this record is not cleared for redistribution. 'style transfer' IN THIS PAPER - THE PRIOR WORK IT IS NOT: the single occurrence names earlier non-parametric methods that manipulate pixels directly, and the sentence after it distinguishes this approach from them; the field's own name for this paper, 'neural style transfer', is absent Quotation withheld because this record is not cleared for redistribution. builds-on → Deep Neural Networks for object recognition Quotation withheld because this record is not cleared for redistribution. builds-on → the VGG-Network (ref. 22) Quotation withheld because this record is not cleared for redistribution. builds-on → the texture feature space of ref. 8 Quotation withheld because this record is not cleared for redistribution. extends → earlier content/style separation work (refs. 12, 13) Quotation withheld because this record is not cleared for redistribution. style representation —operationalised-as→ Gram matrix style representation Gram matrix strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: 'These feature correlations' = the style representation (previous sentence). Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · style representation —distinct-from→ content representation style representation content representation EVIDENCE CORRECTED: RE-GROUNDED. The old span named neither representation as the edge writes them. This sentence names both and is the one that separates them - the method's whole premise is that a photograph's CONTENT and an artwork's STYLE can be matched independently. — strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span writes 'the representations of content and style ... are separable', not 'content representation'. Quotation withheld because this record is not cleared for redistribution. The system uses neural representations to separate and recombine the content and style of arbitrary images, giving a neural algorithm for creating artistic images. about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. style representation style representation Quotation withheld because this record is not cleared for redistribution. convolutional neural network convolutional neural network Quotation withheld because this record is not cleared for redistribution. standard-error standard-error about this quote: DAMAGED GLYPHS IN THIS QUOTE: (cid:126). These are raw PDF font glyph indices that leaked out in place of the characters they stand for - usually a mathematical operator. The quote is verbatim to the staged text and its anchor is sound; the STAGED TEXT is what is damaged. NOT repaired: the cid-to-character mapping is defined by the font embedded in that particular PDF, so guessing it would produce a clean and unverifiable quote. Repair needs the source PDF and is a re-staging job. Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://arxiv.org/abs/1508.06576.