---
title: "Poli et al.'s 2022 NeurIPS paper 'Self-Similarity Priors' explicitly cites both the fractal/IFS and implicit-neural-representation canons and states its method belongs to both"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/abs/2204.07673"
source_sha: "e582a6c91bc834af0a006a450e30d70b2ebf797e4e2e20caaa99374f17cc6da0"
source_title: "Self-Similarity Priors: Neural Collages as Differentiable Fractal Representations"
source_author: "Michael Poli, Winnie Xu, Stefano Massaroli, Chenlin Meng, Kuno Kim, Stefano Ermon"
source_venue: "arXiv (cs.LG) preprint; NeurIPS 2022 (36th Conference on Neural Information Processing Systems), venue confirmed independently via neurips.cc/virtual/2022/poster/53631"
source_date: "2022-04-15T00:00:00.000Z"
source_quote: "(Sitzmann et al., 2020; Mildenhall et al., 2020; Dupont et al., 2021) parametrize the implicit functions via neural networks for use in downstream tasks such as compression."
source_quote_2: "Neural Collages belong to both classes of methods, being a fixed–point iteration whose parameters define data implicitly."
source_quote_3: "Jacquin et al. (1992) introduces more flexible fractal compression schemes for images based on partitioned iterated function systems (PIFSs)."
source_tier: 1
audit_status: "capture-verified — the batch worker fetched and read the paper's full text (24 pages including references and appendix, PDF) at capture time (2026-08-03) and verified all three quotes verbatim via quote_check against the extracted text; venue independently confirmed via a second Tier-1 primary (NeurIPS proceedings/poster page). Independent queen re-check not performed in this headless promotion run (no network access, by design). Freely fetchable primary — clean re-read target for the verifier bee. — Appended 2026-08-08, scheduled cross-model audit (auditor claude-opus-5; writer claude-sonnet-5): the queued re-read is now DONE and the queen re-check is no longer outstanding. PDF re-fetched via extract_pdf, tls verified; sha256 e582a6c9… matches frontmatter EXACTLY. All three source_quotes verified VERBATIM in §5 'Related Work and Discussion' (quote 1 and quote 2 in the 'Implicit representations and models' paragraph, quote 3 opening the 'Fractal compression' paragraph). arXiv v1 date 15 Apr 2022 confirmed on the PDF's own stamp; NeurIPS 2022 acceptance re-confirmed independently at neurips.cc/virtual/2022/poster/53631 (title and all six authors match). Barnsley & Demko 1985 Collage Theorem confirmed as Theorem 1, and present in the reference list as Proc. R. Soc. Lond. A 399(1817):243–275. TWO CORRECTIONS APPLIED to the body; the load-bearing citation-lineage claim and the title are unaffected and stand. (1) The body said the two canons sit in subsections placed 'side by side'; they are two labelled paragraphs of §5 separated by a third ('Attention operators and patches'). Corrected to describe the actual layout. (2) More substantively, the body reported the benchmark as 'narrowing the reconstruction-quality gap with both fractal compression and COIN'. The paper says the opposite about those two baselines — Neural Collages BEAT fractal compression and COIN on reconstruction quality ('less noticeable artifacts and improved color retention'); the gap being narrowed is with SPECTRAL compressors (block-DCT / JPEG), which the paper concedes still hold best PSNR at medium and high bpp. The speed figures were also loose ('roughly 10×' for the paper's 'up to 10×'); now quoted as written, with the Figure 7 caption pointer. The [unverified-quant] flag STANDS unchanged — this audit checked what the paper reports, not whether the numbers reproduce."
provenance: "Promotion from 10-inbox/raw/2026-08-03-does-any-implicit-neural-representation-paper-actually-cite.md, 2026-08-07"
origin: "batch"
derived_from: "10-inbox/raw/2026-08-03-does-any-implicit-neural-representation-paper-actually-cite.md"
date_created: "2026-08-07T00:00:00.000Z"
tags: ["implicit-neural-representations","fractal-compression","iterated-function-systems","neural-collages","michael-poli","image-compression","citation-lineage","history-of-machine-learning"]
seek_code_commit: "649b1a4"
---


"Self-Similarity Priors: Neural Collages as Differentiable Fractal Representations" (Poli, Xu, Massaroli, Meng, Kim & Ermon; NeurIPS 2022) builds its method — Neural Collages — directly on Barnsley and Demko's 1985 Collage Theorem, the mathematical foundation of fractal/IFS image compression, reformulated as a differentiable, hypernetwork-trained operator. Its §5 "Related Work and Discussion" gives each canon its own labelled paragraph (separated by a third, "Attention operators and patches"): under "Implicit representations and models," it cites SIREN, NeRF, and COIN together — "(Sitzmann et al., 2020; Mildenhall et al., 2020; Dupont et al., 2021) parametrize the implicit functions via neural networks for use in downstream tasks such as compression" — and under "Fractal compression," it cites the classical canon, including "Jacquin et al. (1992) introduces more flexible fractal compression schemes for images based on partitioned iterated function systems (PIFSs)." The paper states its own method sits at the intersection: "Neural Collages belong to both classes of methods, being a fixed–point iteration whose parameters define data implicitly." It then benchmarks its compressor against both classical fractal compression and COIN on the same image set — figures not independently replicated here; see below.

This is a peer-reviewed, published paper (not a preprint left unrefereed), and it is the one document found in this vault's research that draws the fractal↔implicit-neural-representation lineage explicitly, resolving the open half of [[question-verify-fractal-compression-inr-lineage]] and updating [[claim-fractal-compression-to-implicit-neural-representations-bridge]]. The lineage runs one direction only: no canonical INR paper cites back the other way — see [[claim-siren-2020-does-not-cite-fractal-ifs-wfa-compression]], [[claim-coin-2021-does-not-cite-fractal-ifs-wfa-compression]], [[claim-strumpler-2022-omits-fractal-compression-credits-coin-as-first]]. No source found, including this paper's own reference list, connects weighted finite automata (WFA) compression specifically — only IFS/PIFS, the Barnsley–Jacquin branch — to any INR paper.

The paper's reported benchmark numbers — encoding speedups of "up to 10× (accounting for training time) and 100× (at test time)" over fractal compression, the 100× restated in Figure 7's caption as "At test time, encoding for Neural Collage compressors is 100× faster than fractal compression" — rest on this single Tier-1 primary source and have not been independently replicated or cross-checked against a second source. The reconstruction-quality comparison points in a different direction from the speed one, and the two should not be merged: against fractal compression and COIN the paper claims Neural Collages are *better* on quality ("less noticeable artifacts and improved color retention"), while the gap it describes itself as *narrowing* is with spectral compressors — block-DCT and the JPEG-family codecs built on it — which the paper concedes still hold best PSNR at medium and high bpp ("Although spectral lossy compressors common in state–of–the–art codecs perform with best PSNR in medium and high bpp settings, Collages narrows the gap in terms of reconstruction quality as well as encoding speed"). **[unverified-quant — needs independent replication]** if these figures are ever cited as standalone numbers rather than as context for the citation-lineage finding above.

> [!note] Seek's commentary:
> The honest shape of the answer is asymmetric, and that asymmetry is the finding. SIREN, COIN, and Strümpler et al. all cite each other and the 3D-shape lineage, never Barnsley or Jacquin — but one paper read both literatures and said, in print, at NeurIPS, "these are doing the same thing." Poli et al. didn't discover a coincidence; they built a bridge on purpose, from the older field looking at the younger one. Nobody downstream — Strümpler et al. included, though it postdates this paper by only months — picked the citation back up in the other direction. A lineage that only one side acknowledges is still a lineage. It's just a lineage with a blind spot built into it, and now the blind spot has a paper trail too.