Does any implicit-neural-representation paper actually cite fractal/IFS/WFA image compression?
Short answer the claims below support: Mostly not — but not entirely. The two founding implicit-neural-representation (INR) papers and their direct compression follow-up trace their own lineage exclusively through 1990s–2010s neural-network work (compositional pattern-producing networks, 3D shape/scene representation, autoencoder-based neural compression) and never mention fractal, IFS, PIFS, or weighted-finite-automata (WFA) compression. But one peer-reviewed 2022 paper — working from the fractal-compression side rather than the INR side — explicitly cites both canons in the same related-work section, states outright that its method belongs to both classes, and benchmarks its compressor head-to-head against both classical fractal compression and COIN. The "image as a compact function" lineage is therefore real in the sense that a citation trail exists in the published literature — but it runs from fractal compression reaching toward INR work, not from a canonical INR paper reaching back to cite fractal/IFS/WFA compression. This updates the open flag on claim-fractal-compression-to-implicit-neural-representations-bridge, which as of 2026-07-09 found "no source in the capture or found since draws this line."
Claim: The two founding INR papers — SIREN (2020) and COIN (2021) — do not cite fractal, IFS, PIFS, or WFA image compression anywhere in their text or reference lists
Claim type: Historical/bibliographic (what a specific primary document's own reference list contains — verified by direct reading of the full text, not a technical-mechanism claim requiring interpretation).
Both papers' PDFs were fetched in full (SIREN, 35 pages including supplement; COIN, 12 pages) and read directly, including the complete numbered reference lists (40 entries for COIN). Neither paper's related-work section, body text, or bibliography contains the words "fractal," "Barnsley," "Jacquin," or "weighted finite automata" in any form.
SIREN's "Related Work" section frames the INR lineage as: "Recent work has demonstrated the potential of fully connected networks as continuous, memory-efficient implicit representations for shape parts [6, 7], objects [1, 4, 8, 9], or scenes [10–13]." — i.e. the 3D-vision/signed-distance-function tradition (DeepSDF, occupancy networks), not compression.
COIN's "Related Work" section frames its own lineage as: "Representing data with neural networks was originally proposed by [32] but has seen a recent surge in interest in the 3D vision community [27, 26, 6]." Reference [32] is Kenneth O. Stanley's 2007 paper on compositional pattern-producing networks (CPPNs) — the paper COIN itself credits as the origin point for "representing data with neural networks," not fractal image coding.
Sourcing floor check: Clears Tier 1 — both quotes are drawn from direct PDF reads of the papers' own text (arXiv primary documents), verified verbatim via quote_check against the extracted text.
| Field | Value |
|---|---|
| source_url | https://arxiv.org/abs/2103.03123 (PDF: arxiv.org/pdf/2103.03123) |
| source_title | COIN: COmpression with Implicit Neural representations |
| source_author | Emilien Dupont, Adam Goliński, Milad Alizadeh, Yee Whye Teh, Arnaud Doucet |
| source_date | 2021-03 (v1) |
| source_venue | arXiv (eess.IV) preprint |
| source_tier | 1 |
| source_sha | 35ee023ddcc0810df5aaaa2dbab8ab4f8dbf1e42b6d4efb9bb5fd69d7bf1a99b |
| source_quote | "Representing data with neural networks was originally proposed by [32] but has seen a recent surge in interest in the 3D vision community" |
| Field | Value |
|---|---|
| source_url | https://arxiv.org/abs/2006.09661 (PDF: arxiv.org/pdf/2006.09661) |
| source_title | Implicit Neural Representations with Periodic Activation Functions |
| source_author | Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, Gordon Wetzstein |
| source_date | 2020-06-17 |
| source_venue | arXiv (cs.CV) preprint; NeurIPS 2020 |
| source_tier | 1 |
| source_sha | 35e460d4bbf083d63291fc4d04e0bf7eff545e7dc4dc608dad6326af00f25286 |
| source_quote | "Recent work has demonstrated the potential of fully connected networks as continuous, memory-efficient implicit representations for shape parts" |
Related existing note: claim-siren-2020-implicit-neural-representations-signals-as-functions already covers SIREN's own definitional claim; this claim adds the negative citation-lineage finding, not a duplicate.
Claim: The direct INR-compression follow-up (Strümpler et al. 2022, ECCV) also omits fractal/IFS/WFA compression from its history, and credits COIN (2021) as "the first" INR-based image-compression approach
Claim type: Historical/bibliographic.
Strümpler, Postels, Yang, Van Gool & Tombari's "Implicit Neural Representations for Image Compression" (ECCV 2022) is the most-cited direct successor to COIN. Its "Related Work" section traces INR-based image compression's origin entirely within the neural-network tradition — DeepSDF → occupancy networks → NeRF → COIN — with no mention of fractal or WFA compression at any point, despite fractal image compression being, on its own historical terms, an earlier "encode-the-image-as-a-compact-function-not-a-grid" approach (see claim-fractal-compression-to-implicit-neural-representations-bridge).
Sourcing floor check: Clears Tier 1 — direct PDF read of the paper's own related-work text, quote verified via quote_check.
| Field | Value |
|---|---|
| source_url | https://arxiv.org/abs/2112.04267 (PDF: arxiv.org/pdf/2112.04267) |
| source_title | Implicit Neural Representations for Image Compression |
| source_author | Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, Federico Tombari |
| source_date | 2021-12 (v1); ECCV 2022 |
| source_venue | arXiv (eess.IV) preprint; European Conference on Computer Vision (ECCV) 2022 |
| source_tier | 1 |
| source_sha | 08e3ae7fdff5f263680991a25e1dbacf45e0623f7d0fca807dec574988900bba |
| source_quote | "Dupont et al. [19] propose the first INR-based image compression approach COIN, which overfits an INR's model weights to represent single" |
Claim: One peer-reviewed paper does explicitly draw the fractal/IFS ↔ implicit-neural-representation lineage — Poli et al.'s "Self-Similarity Priors: Neural Collages as Differentiable Fractal Representations" (NeurIPS 2022)
Claim type: Technical-mechanism / historical-bibliographic (what the paper's related-work section actually contains and argues).
This NeurIPS 2022 paper (arXiv:2204.07673, confirmed as a NeurIPS 2022 poster via neurips.cc/virtual/2022/poster/53631 and the official NeurIPS proceedings page) builds its method, Neural Collages, directly on Barnsley and Demko's 1985 Collage Theorem — the mathematical foundation of fractal/IFS image compression — and reformulates it as a differentiable, learnable operator trained via hypernetworks. Its "Related Work" section contains two adjacent subsections that cite the two canons side by side:
Under "Implicit representations and models": "(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." — directly citing SIREN, NeRF, and COIN.
Under "Fractal compression": "Jacquin et al. (1992) introduces more flexible fractal compression schemes for images based on partitioned iterated function systems (PIFSs)." — directly citing the classical fractal/PIFS canon (Barnsley & Demko 1985; Barnsley 1986; Jacquin et al. 1992; Fisher 2012; Welstead 1999).
The paper states explicitly that 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 directly against both traditions in a single results table (PSNR vs. bits-per-pixel) comparing "Fractal (no aug.)," "Fractal (augment.)," "COIN," "Neural Collage (ours)," and "block-DCT" on the same DOTA aerial-image crops, and reports Neural Collages as roughly 10× faster than fractal compression at training time and up to 100× faster at test time, while narrowing the reconstruction-quality gap with both fractal compression and COIN. (Those specific speedup/PSNR figures are quantitative claims resting on this single Tier-1 primary source and are recorded here as context for the citation-lineage finding, not independently re-verified against a second source — flagged [unverified-quant — needs independent replication] if promoted as standalone numbers.)
This is the direct answer to the core question: the lineage is not merely a structural rhyme invented by this vault. It is drawn explicitly, in a single peer-reviewed document, by researchers who read both literatures and built a method that cites and empirically compares against both. What is not found: no canonical INR paper (SIREN, COIN, NeRF, Strümpler et al.) cites fractal/IFS/WFA compression in the other direction, and no source found in this session — including this bridging paper's own reference list — cites weighted finite automata (WFA) image compression specifically (only IFS/PIFS, the Barnsley–Jacquin fractal-compression branch, not Culik-style WFA).
Sourcing floor check: Clears Tier 1 — direct PDF read of the paper's own text (24 pages including references and appendix), all quotes verified verbatim via quote_check. Venue independently confirmed via a second Tier-1 primary (the NeurIPS proceedings/poster page), not resting on a single unrefereed preprint — this is a refereed, published conference paper.
| Field | Value |
|---|---|
| source_url | https://arxiv.org/abs/2204.07673 (PDF: arxiv.org/pdf/2204.07673) |
| 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_date | 2022-04-15 (v1); NeurIPS 2022 |
| source_venue | arXiv (cs.LG) preprint; NeurIPS 2022 (36th Conference on Neural Information Processing Systems) |
| source_tier | 1 |
| source_sha | e582a6c91bc834af0a006a450e30d70b2ebf797e4e2e20caaa99374f17cc6da0 |
| 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)." |
Corroborating venue check (Tier 1, no quote extracted — used only to confirm publication venue, not to ground a claim):
| Field | Value |
|---|---|
| source_url | https://neurips.cc/virtual/2022/poster/53631 |
| source_title | Self-Similarity Priors: Neural Collages as Differentiable Fractal Representations — NeurIPS Poster |
| source_venue | NeurIPS 2022 official virtual proceedings |
| source_tier | 1 |
This directly resolves — with a positive citation, not just continued absence — the open flag on claim-fractal-compression-to-implicit-neural-representations-bridge ("no source in the capture or found since draws this line," recorded 2026-07-09). It also intersects claim-djvu-1998-paper-coauthored-by-lecun-and-bengio's image-compression cluster and, more loosely, claim-complex-systems-issue-one-paired-nettalk-with-cellular-automaton's note on Karel Culik II's WFA compression as a formally distinct member of the same "images as compact functions" family — a family this paper does not itself connect to WFA specifically.
Further leads
- No source found in this session connects weighted finite automata (WFA) image compression (Culik & Kari; the vault's existing claim-complex-systems-issue-one-paired-nettalk-with-cellular-automaton thread) to any INR paper, including the Poli et al. bridge — searches tried:
"weighted finite automata" image compression neural network implicit representation, plus general WFA/INR combinations. This looks like a genuine gap, not just an unsearched corner, but the search was not exhaustive. - Poli et al. cite Kataoka et al. (2020), "Pre-training without natural images," which uses IFS-generated fractal images as a synthetic pretraining dataset for vision models — a different fractal→neural-network bridge (pretraining data, not representation format) worth its own capture.
- Poli et al.'s GitHub repo (github.com/ermongroup/self-similarity-prior) and project page (zymrael.github.io/self-similarity-prior/) were not fetched this session — could hold additional framing or a blog-style explanation of the fractal/INR connection in plainer language.
- Yuval Fisher's Fractal Image Compression: Theory and Application (Springer, 2012) is cited by Poli et al. as the standard overview reference for the whole fractal-compression field — not independently sourced this session, would be the natural next primary text to check for any forward-looking mention of neural-network parameterizations.
- The Poli et al. paper's own reference list contains a bibliographic artifact — a Mitra, Murthy & Kundu (2000) citation with what appears to be a stray BibTeX key fragment ("itera-tivdupont2022coine function system") mid-title — a benign LaTeX/reference-manager glitch in the source PDF, not adversarial content, noted here only for anyone re-citing that specific reference to check the correct title independently.
Entity candidates
- Michael Barnsley — person — foundational mathematician of iterated-function-system (IFS) fractal compression (Collage Theorem, Fractal Image Compression); the older figure the 2022 Neural Collages paper's whole lineage claim is built on top of and measured against.
- Arnaud (Alan) Jacquin — person — introduced partitioned iterated function systems (PIFS), the practical fractal-image-coding scheme that both the classical fractal-compression literature and the Poli et al. bridge paper cite as the field's key refinement.
- Yuval Fisher — person — author of the standard fractal-image-compression reference text (Fractal Image Compression: Theory and Application, Springer 2012), repeatedly cited by name as the field's overview across sources in this capture.
- Michael Poli — person — lead author of the one paper found in this session that explicitly draws the fractal-compression ↔ implicit-neural-representation lineage.
- Stefano Ermon — person — Stanford faculty, senior/corresponding author on the Neural Collages paper.
- Vincent Sitzmann — person — lead author of SIREN (2020), the paper most often cited as founding the INR field.
- Emilien Dupont — person — lead author of COIN (2021), the first INR-based image-compression method.
- Collage Theorem — concept — Barnsley & Demko's 1985 result bounding an attractor's distance from data; the specific mathematical object Poli et al. reframe as a differentiable neural operator, making it the actual technical hinge between the two traditions.
Sources (4)
claude-sonnet-5 · batch run 2026-08-03; web research via WebSearch, mcp__seek__archive_page, mcp__seek__extract_pdf; direct PDF reads of SIREN, COIN, Strümpler et al., and Poli et al., with quotes verified via mcp__seek__quote_check against the extracted text · raw markdown