Pith. sign in

REVIEW 3 major objections 5 minor 7 references

Fine-tuned Protogen latent diffusion generates novel Batak Ulos motifs that stay culturally faithful, beating Stable Diffusion by roughly tenfold on FID.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Fine-tuned Protogen v3.4 generates novel Ulos motifs with ~10.5 imes lower FID and 2× higher IS than Stable Diffusion v1.4, with guidance scale 5–9 balancing fidelity and diversity.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Solid applied fine-tuning study with clear Protogen-vs-SD numbers and weaver feedback; the ~10.5× FID claim and guidance recommendation rest on a tiny closed motif set and an unablated custom loss, so treat them as in-distribution results rather than general heritage claims. the 3 major comments →

arxiv 2607.06590 v1 pith:ECNNNXGD submitted 2026-07-06 cs.CV cs.LG

AI for Cultural Heritage Textiles: Fine-Tuned Latent Diffusion for Novel Ulos Motif Synthesis

Humasak Tommy Argo Simanjuntak , Jesika Purba , Sitogab Girsang , Widya Manurung , Samuel Situmeang , Arlinta Barus , Daniel Oranova Siahaan This is my paper
classification cs.CV cs.LG
keywords latent diffusionUlos motif generationcultural heritage textilesProtogenStable DiffusionFIDInception Scorefine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Traditional Batak Ulos weaving is constrained by a narrow motif repertoire and slow hand design. This paper shows that fine-tuning two pretrained latent diffusion models on a curated set of 88 motif fragments (augmented to 299) can produce new designs that remain recognisably Ulos. Protogen v3.4 substantially outperforms Stable Diffusion v1.4, reaching roughly 10.5 times lower FID and twice the Inception Score, and therefore sits closer to the real motif distribution while still offering visual variety. Lower denoising strength preserves fidelity; higher strength buys diversity at the cost of realism. Across every tested setting a guidance scale between 5 and 9 stabilises the quality metrics and is recommended as the practical operating range. Traditional weavers and a public panel both prefer the Protogen outputs, confirming that the generated motifs look weaveable and culturally coherent. The result is offered as evidence that carefully conditioned generative models can expand living textile traditions without erasing their symbolic structure.

Core claim

Fine-tuning Protogen v3.4 on a modest, annotated Ulos motif corpus yields culturally consistent yet novel designs that are statistically far closer to authentic Batak patterns (FID ~10.5 imes lower) and more diverse (IS ~2 imes higher) than the same fine-tuning applied to Stable Diffusion v1.4, with guidance scale 5–9 providing the most reliable fidelity–diversity balance.

What carries the argument

Text-conditioned latent diffusion (image-to-image fine-tuning of Protogen v3.4 and Stable Diffusion v1.4) that injects descriptive Ulos prompts via cross-attention and is regularised by an additional perceptual motif-loss term against the 299 training samples.

Load-bearing premise

That 88 unique motif fragments, even after simple geometric and brightness augmentation to 299 images, are representative enough of the full structural, symbolic and ethnic range of Batak Ulos for the reported metrics and weaver judgements to generalise.

What would settle it

Generate a held-out set of real Ulos motifs never seen during fine-tuning and measure whether Protogen’s FID remains an order of magnitude lower than Stable Diffusion’s and whether weavers still rate the new samples as weaveable and culturally authentic.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A guidance-scale window of 5–9 can be used as a default operating range when adapting latent diffusion to other heritage textile motifs.
  • Lower strength settings should be preferred when cultural fidelity is paramount; higher strength when controlled novelty is desired.
  • Protogen-style stylised backbones appear better starting points than general-purpose Stable Diffusion for geometric, high-contrast cultural patterns.
  • The same prompt-plus-perceptual-loss recipe can be reused to enlarge motif libraries for digital weaving platforms without requiring massive new photography campaigns.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pipeline could be extended to other Indonesian weaving traditions (Songket, Ikat, Tenun) that share geometric regularity, provided comparable motif-fragment datasets are assembled.
  • Because weaver feedback already selects for loom feasibility, the generated motifs are natural candidates for a human-in-the-loop design tool that lets artisans accept, reject or further edit AI proposals.
  • If the 10× FID gap holds on larger multi-ethnic corpora, diffusion fine-tuning may become a standard first step for any low-resource cultural textile digitisation project.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper fine-tunes two pretrained latent diffusion models (Protogen v3.4 and Stable Diffusion v1.4) on a curated set of 88 unique Batak Ulos motif fragments (augmented to 299) using text conditioning and an added Ulos-motif perceptual loss. It claims that Protogen consistently outperforms SD-v1.4 by roughly 10.5× lower FID and 2× higher IS, that lower strength improves fidelity while higher strength trades fidelity for diversity, and that a guidance scale of 5–9 stabilises FID/KID/IS and is the recommended operating range for culturally coherent novel Ulos motifs. Quantitative results are reported across three controlled scenarios (structure, colour, complexity) and are supplemented by interviews with three traditional weavers plus a 30-person public survey.

Significance. If the comparative claims and operating-range recommendation hold under proper generalisation checks, the work supplies a concrete, reproducible pipeline for applying text-conditioned LDMs to an under-represented intangible-heritage domain, together with a systematic strength/guidance ablation and dual quantitative–qualitative validation that includes practising weavers. The explicit fidelity–diversity trade-off analysis and the demonstration that a stylistically specialised backbone (Protogen) adapts more readily than a general-purpose one are useful empirical findings for the cultural-heritage AI community. The public release of the motif-generation protocol and the emphasis on keeping artisans in the loop further increase potential impact.

major comments (3)
  1. [§2.1 / Table 1 / Figure 2] §2.1, Table 1 and Figure 2 document only 88 unique motif fragments (heavily skewed: 56 Toba, 24 Karo, 8 Simalungun) that are then augmented to 299 images; all subsequent FID/IS/KID numbers and the weaver rankings appear to be computed against this same closed collection. No held-out real Ulos motifs, external Batak corpus, or leave-one-ethnic-group-out protocol is described. Consequently the headline ~10.5× FID / 2× IS advantage and the guidance-scale recommendation of 5–9 certify reconstruction of the training distribution rather than alignment with the broader Ulos motif space claimed in the abstract and §4. A minimal remedy is a clearly separated reference set (or an external public textile archive) together with error bars or bootstrap intervals on the reported ratios.
  2. [§2.3 / Eq. (1)] Equation (1) introduces an additional Ulos-motif loss term λ·L_Ulos(299) that penalises the perceptual distance of the decoded clean latent to its nearest neighbour among the identical 299 training images. Combined with the absence of a held-out reference, this term further encourages memorisation of the fine-tuning set rather than synthesis of novel yet culturally coherent motifs. The paper should either ablate λ = 0 versus λ > 0 on a true held-out set or replace the nearest-neighbour term with a distributional regulariser that does not reference the training images directly.
  3. [Abstract / §3.1] The abstract and §3.1 assert a ~10.5× FID reduction and 2× IS improvement without reporting variance across seeds, multiple independent fine-tunes, or any statistical test. Given the tiny unique-motif count and the stochastic nature of both diffusion sampling and the Inception feature extractor, these point estimates alone cannot support the strong comparative claim. At least three independent runs with standard deviations (or a non-parametric test) are required before the magnitude of the gap can be treated as reliable.
minor comments (5)
  1. [passim] Model naming is inconsistent throughout (Protogen v3.4 vs. Protogen x3.4 / Protogen v3.4). Standardise on one identifier.
  2. [§3.1 / Figure 4] Figure 4 caption and surrounding text refer to both Inception Score and Fréchet Inception Distance trajectories, yet the y-axis scales and exact epoch values are hard to read; larger fonts or tabulated final values would help.
  3. [Abstract / §4] KID is mentioned in the abstract and conclusion but never defined or tabulated in the results section; either report the numbers or remove the claim.
  4. [§2.2 / Table 2] Table 2 lists identical hyper-parameter ranges for every scenario; a compact description of the three distinct prompt templates would make the experimental design clearer.
  5. [References] Several references appear incomplete or contain typographic artefacts (e.g., missing page ranges, duplicated author lists). A final bibliography pass is needed.

Circularity Check

0 steps flagged

No circular derivation: empirical fine-tuning comparison with external FID/IS and independent human raters; main claims do not reduce to inputs by construction.

full rationale

This is an empirical systems paper that fine-tunes two pretrained latent diffusion backbones on a curated Ulos motif set and compares them with standard external metrics (FID, IS, KID against the motif reference distribution) plus independent weaver interviews and a public survey. There is no first-principles derivation, uniqueness theorem, or fitted parameter that is later renamed a prediction. The optional nearest-neighbour Ulos motif loss in Eq. (1) regularizes generations toward the 299 training images, which is ordinary training-time self-reference and does not make the reported FID/IS gap or the guidance-scale recommendation equal to the inputs by construction; both models are evaluated the same way after training. Self-citations to prior DiTenun/StyleGAN work ([15]) supply background on earlier approaches and are not load-bearing for the Protogen-vs-SD comparative claim. Dataset size and lack of held-out external Ulos corpora are validity/generalization concerns, not circularity. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard LDM machinery, a small curated cultural dataset, and a handful of hand-chosen training and sampling hyper-parameters. No new physical or mathematical entities are postulated; the optional Ulos motif loss is an ad-hoc regulariser whose necessity is not demonstrated. Free parameters are the usual optimisation and sampling knobs plus the un-ablated loss weight λ.

free parameters (5)
  • learning rate = 1.0e-5
    Fixed at 1.0×10^{-5} for 50 epochs; no sensitivity study reported.
  • strength (denoising strength) = 0.65–0.85
    Swept over {0.65, 0.75, 0.85}; lower values preferred for fidelity. Chosen by experiment rather than derived.
  • guidance scale s = 5–9 (recommended)
    Swept 5–19; authors recommend 5–9 as operating range. Empirical, not theoretically fixed.
  • λ (Ulos motif loss weight)
    Appears in Eq. (1) to balance novelty vs. fidelity; numerical value never stated or ablated.
  • augmentation parameters (±30° rotation, ±15% brightness, flips) = ±30°, ±15%
    Hand-chosen transforms used to expand 88 images to 299; no ablation of their effect on final FID.
axioms (4)
  • domain assumption Latent diffusion models pretrained on large general image corpora can be successfully adapted to a small, culturally specific motif domain via continued training with text conditioning.
    Invoked throughout §2.2–2.3; standard transfer-learning premise for LDMs.
  • domain assumption FID and IS computed with an ImageNet-pretrained Inception network are meaningful proxies for visual fidelity and diversity of geometric textile motifs.
    Used as primary quantitative metrics in §3.2.1; known limitation for non-natural-image domains is not discussed.
  • ad hoc to paper The 88 curated fragments plus geometric/photometric augmentation adequately represent the structural and symbolic diversity of Toba, Karo and Simalungun Ulos.
    Stated in §2.1; load-bearing for all generalisation claims.
  • domain assumption Classifier-free guidance (Eq. 3) with positive/negative prompts is sufficient to enforce cultural constraints such as symmetry and traditional colour palette.
    Core conditioning mechanism described in §2.3.
invented entities (1)
  • Ulos motif loss L_Ulos(299) no independent evidence
    purpose: Additional perceptual nearest-neighbour regulariser intended to keep generated latents close to the 299 real training motifs.
    Introduced in Eq. (1) and Figure 3; no ablation or independent validation is provided, so its necessity remains unproven.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AI for Cultural Heritage Textiles: Fine-Tuned Latent Diffusion for Novel Ulos Motif Synthesis." pith.science (2026). https://pith.science/paper/ECNNNXGD

@misc{pith2026260706590,
  author       = {Pith},
  title        = {Pith review of: AI for Cultural Heritage Textiles: Fine-Tuned Latent Diffusion for Novel Ulos Motif Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECNNNXGD}},
  note         = {Machine review of arXiv:2607.06590}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Preserving and revitalising traditional textiles such as Ulos, a cultural heritage of the Batak ethnic group in North Sumatra, Indonesia, requires balancing fidelity to tradition with innovative approaches that meet contemporary design demands. Traditional Ulos weaving faces two key limitations: a narrow range of motifs and a time-intensive design process. This study presents a generative AI framework that fine-tunes two pretrained latent diffusion models: Protogen v3.4 and Stable Diffusion v1.4, on a curated, annotated dataset of high-resolution Ulos motifs to generate culturally consistent yet novel designs. Model performance is evaluated quantitatively using Frechet Inception Distance (FID), Inception Score (IS), and qualitatively through assessments by traditional weavers and members of the public. Protogen v3.4 consistently outperforms Stable Diffusion v1.4, achieving substantially lower FID (~10.5x) and higher IS (2.0x), indicating superior visual fidelity, diversity, and closer alignment with the real Ulos motif distribution. We further examine the effects of strength and guidance scale on generation quality across both models. Lower strength values consistently yield higher fidelity (lower FID), while higher strength values increase generative diversity at the cost of realism, revealing a clear fidelity-diversity tradeoff for both models. Across all tested configurations, a guidance scale of 5-9 provides the most effective balance between fidelity and diversity, stabilising FID, KID, and IS, and is recommended as the operating range for high-quality, diverse Ulos motif generation. These findings demonstrate that carefully fine-tuned generative AI can support the creative renewal of intangible cultural heritage while preserving its stylistic and symbolic integrity.

Figures

Figures reproduced from arXiv: 2607.06590 by Arlinta Barus, Daniel Oranova Siahaan, Humasak Tommy Argo Simanjuntak, Jesika Purba, Samuel Situmeang, Sitogab Girsang, Widya Manurung.

Figure 1
Figure 1. Figure 1: The example of ulos harungguan (a), traditional weaving with Gedogan (b) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 4 linked inside Pith

  1. [1]

    king of ulos

    The Title of the Paper: ACM Conference Proceedings Manuscript Submission Template: This is the subtitle of the paper, this document both explains and embodies the submission format for authors using Word. In Woodstock ’18: ACM Symposium on Neural Gaze Detection, June 03–05, 2018, Woodstock, NY. ACM, New York, NY, USA, 10 pages. NOTE: This block will be au...

    2018

  2. [2]

    jumlah lidi

    Figure 7: Top-rated LDM-generated motif images chosen by the weavers. Structural elements such as horizontal alignment, visual symmetry, and proportional repetition are core characteristics in traditional Ulos motifs. According to the weavers, the motifs generated by Protogen x3.4 at moderate strength settings (0.65–0.75) demonstrated visual balance, clea...

    2025

  3. [3]

    Available: https://tfr.news/articles/2022/8/31/beyond-batik-exploring-the-diversity-of-traditional-indonesian-textiles?

    [Online]. Available: https://tfr.news/articles/2022/8/31/beyond-batik-exploring-the-diversity-of-traditional-indonesian-textiles?. [Accessed August 2025]

    2022

  4. [4]

    DIffusion Models Already Have a Semantic Latent Space,

    M. Kwon, J. Jeong and Y. Uh, "DIffusion Models Already Have a Semantic Latent Space," in arXiv preprint arXiv:2210.10960,

    Pith/arXiv arXiv

  5. [5]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis,

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna and R. Rombach, "SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis," in arXiv preprint arXiv:2307.01952,

    Pith/arXiv arXiv

  6. [6]

    Image Data Augmentation for Deep Learning: A Survey,

    S. Yang, W. Xiao, M. Zhang, S. Guo, J. Zhao and F. Shen, "Image Data Augmentation for Deep Learning: A Survey," arXiv preprint arXiv:2204.08610, 19 April

    Pith/arXiv arXiv

  7. [7]

    Prompt-to-prompt image editing with cross attention control,

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y. Pritch and D. Cohen-Or, "Prompt-to-prompt image editing with cross attention control," arXiv preprint arXiv:2208.01626, 02 August

    Pith/arXiv arXiv

This paper was first reviewed by grok-4.5 on July 11, 2026.