Source-linked AI summary
Rate-Adaptive One-Step Diffusion Compression for AIGC Images
Nitiz Khanal
TL;DR
At ultra-low rates, AIGC images contain text, edges, and repeated motifs that conventional codecs may erase, while generative decoding can alter source geometry or symbols. The entry combines rate-specialized one-step diffusion-codec candidates with encoder-side optimization and global allocation, ranking fifth with the second-lowest DISTS among verified entries.
Problem
At ultra-low rates, AIGC images contain text, edges, motifs, UI-like layouts, and stylized texture that conventional codecs remove, while generative decoders may alter source symbols or geometry.
Source: §1 · p. 2Method
The system combines four rate-specialized AEIC-ME checkpoints, encoder-side latent optimization, and a measured-file-size knapsack allocation with a fixed residual restorer.
Source: §1 · p. 2, §3.2 · p. 4Results
Ranked 5th, the entry had DISTS 0.038970, second-lowest among ten verified entries; reweighting its fixed candidate pool raised achievable PSNR only to 27.08 dB, a 0.06 dB improvement.
Source: §4.4 · p. 10, §5 · p. 14Takeaways & Limitations
The authors conclude that the perceptual track fits this entry better and identify fine-tuning or replacing the base codec’s fidelity objective as the most promising path forward.
Source: §6 · p. 14Global allocation requires all source/candidate pairs and the total pixel budget before solving, so it is not an online or independently rate-controlled operation.
Source: §3.5 · p. 7
Abstract
from arXiv · showhide
We describe our entry to the LoViF 2026 AIGC Image Compression Challenge, a benchmark for ultra-low-bitrate coding of AI-generated images under a strict global rate budget of 0.025 bits per pixel (BPP). Generated imagery poses a distinct challenge for compression: it frequently contains rendered typography, synthetic edges, repeated motifs, UI-like layout, and stylized micro-texture that conventional distortion-oriented codecs erase at this rate, while unconstrained generative decoders can restore plausible-looking detail that no longer matches the source geometry or symbols. We treat this as a rate-perception allocation problem. Our system fine-tunes four rate-specialized checkpoints of the AEIC one-step diffusion codec, generates per-image candidates from all four, including one latent refined through encoder-side test-time optimization (TTO) with periodic entropy-conditioning refresh, entropy-codes every candidate with practical rANS coding, and selects exactly one bitstream per image with an exact multiple-choice knapsack solved over true coded file sizes. A fixed, zero-additional-bit residual restoration network is applied at decode time. Every submitted bitstream is independently decodable by the shipped decoder, which uses no source image or external side information. We report the full pipeline, an ablation history spanning 77 logged experiments, and a set of negative results, including why PSNR could not be pushed to parity with rate-distortion-oriented competitors under this architecture, useful to future participants. This is a challenge report: our entry scored 31.527739 (PSNR 27.02 dB, MS-SSIM 0.9176, LPIPS 0.0778, DISTS 0.0390 at 0.02495 BPP), the second-best DISTS on the leaderboard, ranking 5th on the final test-phase leaderboard announced August 4, 2026.
1 Introduction
Ultra-low-bitrate AIGC compression must preserve text, edges, motifs, and stylized textures that conventional codecs may erase, without generative decoders changing source symbols or geometry. The paper addresses this as rate-perception allocation, combining rate-specialized diffusion-codec candidates with measured-size global selection.
1 Introduction
- At ultra-low rates, AIGC images combine rendered text, synthetic edges, repeated motifs, UI-like layouts, and stylized texture that conventional codecs remove while generative decoders may alter symbols or geometry.The challenge evaluates PSNR, MS-SSIM, LPIPS, and DISTS jointly.Source: §1 · pp. 1–2
- The system treats compression as rate-perception allocation, selecting one candidate per image across four rate-specialized checkpoints using a knapsack MILP on measured file sizes.The encoder has the complete fixed image set and source images before allocation; this is an offline dataset-level setting without independent per-image rate guarantees.Source: §1 · p. 2
- The pipeline combines encoder-side latent test-time optimization and a zero-additional-bit residual restorer with a practical one-step diffusion codec.The MILP enforces the global rate budget using actual file sizes rather than estimated rates or mean per-image BPP.Source: §1 · p. 2
- Across 77 development experiments, cross-family fusion with OneDC or OSCAR was non-competitive, and a differentiable PSNR objective did not close the PSNR gap.The authors suggest the gap may reflect the architecture’s fidelity-perception trade-off rather than loss weighting.Source: §1 · p. 2
2 Related Work
Related work spans distortion-oriented learned codecs, diffusion priors for extreme compression, text-aware AIGC methods, and perceptual quality metrics. The paper situates its approach among these methods and distinguishes its use of a learned text-preserving bias without transmitted OCR side-information.
Learned image compression
- Learned codecs combine analysis/synthesis transforms, entropy models, and practical range coding, but distortion-oriented objectives tend to blur or remove fine text and repeated patterns at ultra-low rates.The cited examples include hyperpriors, context models, and rANS [4].Source: §2 · pp. 2–3
Diffusion priors for extreme compression
AIGC-specific and text-aware compression
Perceptual quality metrics
- The challenge evaluates PSNR, MS-SSIM, LPIPS, and DISTS, with LPIPS using the organizers’ AlexNet backbone and the local evaluator matching checked leaderboard values.PSNR and MS-SSIM are computed in RGB.Source: §2 · p. 3
3 Method
The method allocates rate across candidate bitstreams from four AEIC-ME checkpoints, including encoder-refined latents, then applies a fixed decoder-side restorer. Selection uses measured file sizes under a global budget, while each chosen bitstream remains independently decodable.
3.1 Overview
- The encoder evaluates candidates from four checkpoints and a TTO-refined latent, then selects one stream per image using a global knapsack over true coded file sizes.The decoder uses the checkpoint byte, reconstructs with a one-step diffusion pass, and applies a fixed restorer without source-image access.Source: §3.1 · p. 4
3.2 Backbone: AEIC-ME with a one-step diffusion prior
- AEIC-ME maps RGB content to a quantized latent scored by an autoregressive entropy model and decoded through an SD-Turbo one-step diffusion prior.Overlapping latent and VAE tiles support large-image reconstruction while bounding peak activation memory and avoiding visible seams.Source: §3.2 · p. 4
3.3 Rate-specialized checkpoint bank and bitstream signaling
- Four fine-tuned checkpoints cover complementary rate/perception regions; a leading byte selects the codec, while raw-versus-restored choice is un signaled and restoration is unconditional.Their target rates are measured natural operating regions, not imposed per-image constraints.Source: §3.3 · p. 5, Table 1 · p. 5
3.4 Encoder-side latent test-time optimization
- TTO optimizes a random latent crop through the frozen diffusion path, refreshes entropy conditioning periodically, and uses a differentiable rate estimate only as a soft proxy before actual rANS coding.The rate penalty discourages exceeding the trajectory's starting estimate; final bit counts use measured coded file sizes.Source: §3.4 · p. 6
3.5 Global MILP rate allocation
- The multiple-choice knapsack chooses one candidate per image to maximize challenge score under a total-bit constraint using measured candidate scores and coded lengths.The constraint uses pixel-weighted BPP and actual on-disk size, including each header byte.Source: §3.5 · p. 7
- At a 0.02499-BPP cap, HiGHS scored 111.53643 versus greedy's 111.53236, a +0.00407 S pool-level retrospective gain, not a new submission.The submitted default-tolerance solution was near-optimal, not certified optimal.Source: §3.5 · p. 7
3.6 Decoder-side residual restoration
- A fixed residual CNN corrects recurring color, edge, and local-texture errors without source pixels or side information, adding zero BPP.TTO optimizes against the raw decode, not through the restorer; four archived PNGs are not reproduced exactly because the decoder always restores and the bitstream lacks a restoration flag.Source: §3.6 · p. 8
3.7 Implementation and reproducibility
- The 4.76-GB decoder bundle fits the 5-GB package limit, but mixed-checkpoint decoding and full-resolution restoration can exceed 22–24-GB GPU memory on the largest images.Overlapping-tile restoration bounds activation memory and matches a full-image forward pass numerically; the reported 8.5 s/image archive field is hard-coded, not a measured benchmark.Source: §3.7 · pp. 8–9
4 Experiments
The experiments combine supplementary codec training, rate-specialized models, and validation-based selection with an evaluation spanning the challenge metrics. The submitted system ranked fifth on the final test leaderboard, with DISTS second-lowest overall, while development experiments documented both gains and unsuccessful candidates.
4.1 Training data
- About 15,700 supplementary generated images supported codec fine-tuning, while the restorer used 3,200 official-training-split pairs; validation and test images did not update model parameters.The 800 official training images were supplemented with images generated from public prompt collections; test sources were used for permitted encoder-side optimization and allocation.Source: §4.1 · p. 9
4.2 Training protocol
- Four codec specialists shared an architecture but varied in loss weights, initialization, and corpus; validation selection considered challenge metrics, BPP, stream lengths, and decoder reproducibility.The restorer used an L1:LPIPS:DISTS loss-weight ratio of 1:1:1.5; tested edge-aware and adversarial variants were not adopted.Source: §4.2 · pp. 9–10
4.3 Evaluation protocol
- The evaluation score combines PSNR, MS-SSIM, LPIPS, and DISTS; the reported objective leaderboard is distinct from the organizers’ human-assessed perceptual perspective.The system was optimized against the stated score, not a human-preference signal; official leaderboard values use B = S −80.Source: §4.3 · p. 10
4.4 Main results
- Adding three fine-tuned checkpoints raised the early validation mix from 109.55 to 109.89 (+0.34) under the same, later-superseded mean-BPP solver.This comparison used the same early methodology and is not comparable to the later Table 6 rows.Source: Table 3 · p. 11
- Ranked 5th, the entry recorded DISTS 0.038970, second-lowest overall, while the top four reported 28.6–29.8 dB PSNR versus its 27.0 dB.The official final-test results were recalculated by the organizers; the entry’s DISTS was lowest among entries ranked above it.Source: §4.4 · p. 10, Table 5 · p. 11
4.5 Qualitative results
- On a hand-selected text-heavy poster, 500-step TTO sharpened typography and repeated motifs as rate rose from 0.01774 to 0.01885 BPP; restoration made smaller corrections at zero additional bits.The figure illustrates stage effects for one image, not aggregate performance.Source: §4.5 · p. 12, Figure 2 · p. 12
4.6 Submission progression and development history
- Matched validation rows isolate TTO at +0.3310 S under the same MILP; v20→v21 adds refreshed 500-step candidates, but cannot establish test-set transfer.Progression rows are legal-rate mixed candidates; v20 was not evaluated on the test set.Source: §4.6 · pp. 12–13, Table 6 · p. 13
- The fixed restorer gained +0.22 S on average across candidate types, rising to +0.27 with longer training; screening found lower MSE weight monotonically hurt score.A 16-block, 2×-capacity restorer scored slightly worse than the smaller model at matched convergence (111.25 vs. 111.27 S). Other screened candidates included wider TTO crops, DISTS-surrogate weights, multiple seeds, weight-souping, and cross-family fusion.Source: §4.6 · p. 13
5 Discussion: the PSNR gap
The discussion tests whether candidate reweighting or adding a PSNR term to TTO can reduce the PSNR gap, reporting two experiments with different evidential strength.
5 Discussion: the PSNR gap
- 27.08 dB was the highest legal-rate PSNR from reweighting the fixed candidate pool, improving only 0.06 dB while combined score fell from B=31.53 to 31.48.The authors conclude that selection reweighting cannot materially close the gap for the existing candidates; this fixed-pool analysis does not establish unseen-data generalization.Source: §5 · p. 14
- 26.05 dB mean PSNR with an added TTO PSNR term was below the existing candidate’s 26.16 dB on the same four images, alongside perceptual-metric regressions.The n=4 screen is preliminary and cannot rule out gains from a different learning rate, weight schedule, or longer trajectory.Source: §5 · p. 14
- The authors judge it likely, but not proven, that the residual PSNR gap reflects the one-step diffusion synthesis path rather than a loss-weighting artifact correctable within the pool.They note that generic distortion-oriented codecs improved PSNR on selected images at substantially worse LPIPS/DISTS, and identify DISTS—not PSNR or LPIPS—as the system’s confirmed strength.Source: §5 · p. 14
6 Conclusion
The submission combines rate-specialized AEIC-ME checkpoints, encoder-side latent optimization, measured-size rate allocation, and fixed restoration. It ranked 5th, with second-best DISTS, while trailing the top four on PSNR.
6 Conclusion
- Rank 5 with score 31.527739 and second-best DISTS, while trailing the top four on PSNR, summarizes the submission’s final standing.The authors say the perceptual award track fits better than the split-track award alternatives.Source: §6 · p. 14
- The submission used four fine-tuned AEIC-ME checkpoints, encoder-side latent TTO with entropy-conditioning refresh, measured-size knapsack allocation, and a fixed zero-additional-bit restoration network..Source: §6 · p. 14
- The authors identify the diffusion backbone’s fidelity/perception trade-off as the likely source of the PSNR gap and propose changing the base codec’s fidelity objective.TTO does not modify decoder weights, and the described codec/checkpoints were those submitted for the Code Submission Phase.Source: §6 · p. 14