acceptodds
Under review as a conference paper at ICLR 2027

CORD: Competitive Regional Diagrams for Fast and Compact Large Image Representation

Abstract

Explicit primitive-based methods have achieved high fidelity but remain costly to scale to large images. Recent competitive representations improve reconstruction through per-pixel competition. We observe that the same geometry-dependent competition recurs throughout the pipeline: it is repeatedly updated over millions of pixels during fitting, altered by geometry quantization, and costly to rediscover during rendering. We introduce COmpetitive Regional Diagrams (CORD), which localizes this recurring competition within independently fitted regional diagrams that serve as common units of fitting, coding, and rendering. Bounded competitive fitting limits the size of each optimization problem and completes pruning early, leaving retained primitives time to adapt. Decoded-state coding adapts appearance to decoded geometry and evaluates alternatives by their decoded distortion and actual encoded size. Image-wide rate allocation reuses these alternatives across budgets and directs bytes by reconstruction gain, without per-budget retuning or retraining. Persistent contributor mapping compiles decoder-selected contributor lists into compact GPU state, sharing IDs across neighboring pixels to avoid repeated discovery at fixed geometry and output grid. Experiments show that CORD achieves better rate–distortion trade-offs than the evaluated baselines. On FGF2, CORD High gains 4.60 dB over SGI with 19% smaller payloads, an 18.9 optimization speedup, and nearly 200 FPS for repeated rendering on an A800 GPU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large image representation, competitive diagrams, explicit 2D primitive, Gaussian splatting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vLeJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposed the bounded competitive fitting for the task of large image representation with primitives. It fit primitives to image tiles as opposed to pixels to keep the complexity (storage size) bounded. It also implemented decoded-state coding and rate allocation to compress the fitted primitives and to split an image-wide byte budget across tiles. To further increase the hardware efficiency, it then proposed persistent contributor mapping to remove the per-render contributor search. On FGF2, CORD’s High setting outperforms the SGI baseline by 4.60 dB with 19% smaller payloads and an 18.9x optimization speedup.

Critical strengths
  1. The method looks well-crafted and well-balanced between the complexity and the efficiency it can bring.
  2. The methods are paired with a native GPU implementation that boosts the hardware efficiency.
  3. The experimental numbers are supported by clear ablations to know the rough contribution of each proposed component from both quality and efficiency perspective.
Critical weaknesses
  1. One of the major motivations of this paper is storage compression. However, I don't see a comparison with modern codecs, which can also bring a lot of savings in storage. Then the question is: why go for this instead of just referring to traditional codecs?
  2. For the contribution on mitigating repeated rendering, what's the edge agains simply keeping the decoded image in memory?
  3. 9 images, and the 3 used for tuning are also in the test set ... which can be limited for evaluation
Questions for authors

Please address my concerns in the witness.

Reviewer_bYieOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CORD, a framework for compact and efficient representation of high-resolution images based on competitive spatial primitives. Building upon Soft Anisotropic Diagrams (SAD), CORD introduces three main components: (1) bounded regional fitting, which partitions large images into independently optimized regions and improves fitting efficiency through joint color initialization, early pruning, and pixel sampling; (2) decoded-state coding and rate allocation, which adapts primitive colors to quantized geometry and selects regional encoding configurations under a global bitrate budget based on measured rate–distortion performance; and (3) persistent contributor mapping, which caches pixel-to-primitive associations and uses block-wise shared dictionaries to reduce memory consumption and accelerate repeated rendering.

The method is evaluated on nine high-resolution images from three datasets, with comparisons against neural and explicit image representations. Experiments demonstrate substantial improvements in reconstruction quality, storage efficiency, and optimization speed.

Critical strengths

1. Well-designed and coherent system optimization. The paper presents a carefully designed framework that jointly addresses fitting, compression, and rendering efficiency for high-resolution image representations. Although some individual techniques are relatively straightforward, their integration is well motivated by the characteristics of competitive primitive-based representations. In particular, decoded-state appearance adaptation and persistent contributor mapping provide practical solutions to the challenges introduced by quantization and repeated contributor discovery. Overall, I consider this a technically well-executed system-level contribution.

2. Strong empirical improvements with solid experimental validation. The paper demonstrates substantial improvements in reconstruction quality, storage efficiency, and optimization speed over existing explicit image representations. For example, on FGF2, CORD achieves a 4.60 dB PSNR improvement over SGI with 19% smaller representations and an 18.9× fitting speedup. Beyond these results, the paper provides extensive ablation studies, rate–distortion comparisons, and detailed computational cost analyses to support its design choices. The magnitude of the improvements and the relatively thorough validation provide strong evidence of the method's practical effectiveness.

Critical weaknesses

I did not identify any major weaknesses that would justify rejection.

Questions for authors

I have two clarification questions regarding the experimental evidence:

1. Attribution of performance improvements. Table 2 reports a 3.03 dB improvement from bounded regional fitting, whereas the controlled experiments in Table 14 show much smaller gains (0.16–0.22 dB). Could the authors clarify which differences in experimental settings primarily account for this gap, and how much of the improvement can be attributed to regional decomposition itself?

2. End-to-end computational efficiency. The appendix provides detailed measurements of coding and mapping construction costs, but the main efficiency comparisons primarily focus on fitting and repeated rendering. Could the authors clarify the overall speedup when including the complete encoding and decoding pipeline, particularly for scenarios where each image is rendered only once? This would help assess the practical significance of the reported efficiency gains.

Reviewer_fNatOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes a new method, CORD, for representing large images. Overall CORD is based on SAD. The basic pipeline of CORD is: (1) divide the image into regions, and fit each region independently. (2) re-optimize the colors after geometry quantization and decoding, and encoding for each region under a storage budget. (3) Store the per-pixel contributor lists and reuse them for repeated rendering.

Critical strengths
  • The results are quite good. Compared with the baseline methods like SGI and SAD, CORD achieves better reconstruction quality using smaller storage and less fitting time. The color re-adaptation after geometry decoding is reasonable, as quantization changes the blending weights and the original colors are no longer be suitable.

  • The ablation studies are looks sufficient -- (1) the authors ablate the color re-adaptation method, and (2) In the the mapping-layout comparison, the output result keeps the same while reducing the GPU memory and rendering latency. These experiments help readers understand the contributions of each component.

Critical weaknesses
  • My main concern is that the number of test images is too small. There are only nine images tested in the experiments -- the diversity seems very limited.

  • The novelty is a bit limited. Basically the method follows SAD, and regional decomposition, quantization-aware refinement, and slope-based rate allocation have already been proposed. The main contribution here seems to be how these are combined to the encoding pipeline.

  • For me, some parts are not easy to follow. In Sec. 3.3, it is unclear which encoding options require a new fitting run, and which to reuse an existing state. Also, the timing definitions are scattered across the paper, which is a bit hard to follow.

Questions for authors
  • Can the authors test on more independent images?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.