A new arXiv preprint reports an image-communication system that reached bitrates of 0.0062 to 0.0468 bits per pixel on a selected 1,000-image ImageNet-V2 subset. Those rates were reported as more than 90 percent below the range for Cheng2020-Attention, while the reconstructed images scored 0.923 to 0.973 on CLIP similarity and 0.140 to 0.595 on LPIPS. The catch is at the receiver: the implementation took 1,012 milliseconds to decode an image, compared with 6 milliseconds for SwinJSCC. The result is a clear exchange of communication efficiency for computation, which is how the authors frame the method.
The system sends the beginning and predicts the rest
The design, called Ada-TokenCom, divides the message into an informative beginning and a predicted ending. It arithmetic-codes the retained prefix, then uses the same autoregressive model at the receiver to generate the tokens that were not sent. Autoregressive means the model predicts each next token from the ones already in context. The retained prefix is the rate control: send more tokens for a fuller transmitted description, or fewer when the link is tight.
That decision is made jointly with the wireless transmission settings. The adapter considers five coding choices paired with QPSK and 16QAM, creating 10 candidate modulation-and-coding schemes. MCS is the bundle of choices that determines how data is coded and sent. A Lyapunov drift-plus-penalty controller then picks a feasible prefix-length and MCS pair while managing a long-term symbol budget.
A test built around image tokens and a simulated link
The main image setup used an ImageNet-1K-trained VQGAN tokenizer with LlamaGen-L. Each 256 by 256 image was converted into a 16 by 16 grid, or 256 visual tokens, drawn from a codebook, or library, of 16,384 entries. In the self-information analysis, the researchers sampled 1,000 images. ImageNet-V2 was not used to train that tokenizer-model setup.
One clue behind the compression strategy came from measuring self-information, the average number of bits associated with a token. Taming Transformers had the lowest reported averages: 8.09 bits per token when conditioned on prior context and 8.37 without that conditioning. LlamaGen-L recorded 11.22 and 11.33, while LlamaGen-B recorded 11.42 and 11.54. Across the sequence, information was higher early and lower after more preceding tokens had been observed, a pattern consistent with using context to predict the tail.
More transmitted context was also associated with higher reported semantic quality in the class-to-image test. At a retained prefix of 32 tokens, the mean CLIP-quality score was 0.843; at 256 tokens, it was 0.937. Here, CLIP is the study's semantic-quality measure, while PSNR is its pixel-fidelity measure.
The clearest gains appeared at low rates
On the randomly selected 1,000-image validation subset, class-conditioned Ada-TokenCom spanned 0.0062 to 0.0468 bits per pixel, or bpp. That was reported as more than 90 percent below Cheng2020-Attention's approximately 0.145 to 0.711 bpp range and below HiFiC's 0.688 to 1.003 bpp range. CLIP similarity ranged from 0.923 to 0.973, while LPIPS, another reported perceptual measure, ranged from 0.140 to 0.595. A favorable low-rate FID was also reported, but it was calculated on the selected subset and is not directly comparable with full-benchmark FID. Conventional codecs still had higher PSNR.
The wireless comparison favored meaning over pixel exactness
For the wireless comparison, all methods were evaluated on 10,000 ImageNet-V2 images. SwinJSCC was trained on 1.2 million ImageNet-1K images and validated on 50,000; the end-to-end comparison ran for 103 simulation time slots. Separate policy sweeps covered four dynamic-SNR scenarios, each reported as 104 time slots. Retransmission overhead counted in the channel bandwidth ratio, or CBR, while quality scores were calculated only for successfully reconstructed images.
The channel itself was modeled as a point-to-point, quasi-static block-fading Rayleigh link. The simulation assumed perfect channel-state information at the receiver, zero-forcing equalization, soft demodulation, channel decoding and CRC verification. Under those conditions, Ada-TokenCom reported 21 to 28 percent higher CLIP similarity than SwinJSCC around the study's constrained bandwidth point, with an approximately 47 percent gain at lower CBR. LPIPS was up to about 60 percent lower at comparable CBR, and FID up to about 77 percent lower at high signal-to-noise ratio. SwinJSCC retained higher PSNR.
Bandwidth savings come with a receiver-side bill
The adaptive controller was not always conservative. In the parameter sweep, a small adaptation weight underused the budget, while an excessively large one could violate it. The displayed setting was approximately 106 and was selected as aggressive but near-feasible. In the reported SNR process, with parameters 12 and 10, final-slot throughput came close to the long-term oracle. Under limited budgets, cumulative CLIP and perceptual reward were higher, although link utilization could temporarily exceed one before moving toward the constraint boundary.
The largest practical drawback was receiver cost. Ada-TokenCom was listed with 372.21 million encoder parameters and 385.36 million decoder parameters, compared with 18.34 million and 14.73 million for SwinJSCC. Its reported inference cost was 311.41 and 422.47 GFlops, versus 35.01 and 34.11 GFlops. Encoder latency was 19 milliseconds versus 10, but decoder latency was 1,012 milliseconds versus 6. The gap reflects the paper's central bargain: less traffic on the link in exchange for more work at the receiver.
A promising trade, still bounded by the test
The study evaluated a VQGAN and LlamaGen token system on ImageNet-derived data and modeled the wireless channel as a Rayleigh link, so the reported comparisons describe that setup. Pixel-level fidelity was not the strongest result, with conventional codecs retaining higher PSNR, and the selected-subset FID is not directly comparable with full-benchmark values. The authors identify autoregressive generation latency as a current limitation and propose faster generation and broader extensions.
The supplied document is an arXiv preprint, version 1, dated 28 Aug 2026.
Paper data and sources
Original title: Ada-TokenCom: Rate-Adaptive Token Communications via Large-Model-Driven Token Compression and Generation
Authors: Zijun Zhang, Li Qiao, Mahdi Boloursaz Mashhadi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text