acceptodds
Withdrawn from ICLR 2027

2-Bit Autoregressive Image Generation on Mobile: An End-to-End Framework via Complex-Valued QAT

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image generation is one of the most valuable applications of generative models. As users become increasingly sensitive to data sovereignty and privacy, migrating inference from the cloud to the edge (on-device) has become an urgent necessity. Deploying generative models on mobile hardware requires low storage footprint, low inference latency, and low energy consumption. Model quantization has emerged as the predominant technical trajectory to address these challenges. Quantization-aware training (QAT) embeds quantization operators into the training process, allowing model parameters to actively adapt to low-level representation, thus exceeding the performance of post-training quantization (PTQ) in the three dimensions of storage, latency and energy consumption. This paper proposes a complex-valued QAT framework for autoregressive image generation that addresses channel pairing and layer-wise quantization sensitivity, and supports on-device deployment. We firstly introduce complex-valued quantization methods to the field of image generation. Complex-valued QAT maps weights to the set , offering superior representation capacity compared to real-valued methods under equivalent storage budgets. And we propose Weight Channel Reorder strategy, which forces complex projection points toward the coordinate axes to significantly suppress quantization error. Extensive experiments on LlamaGen and VAR show that our 2-bit quantized models approach full-precision performance (e.g., FID 2.83 vs. 2.62 on LlamaGen-XL), achieving lower FID than all compared 2-bit QAT baselines and most 6-bit PTQ methods. We further develop a GGUF export pipeline and deploy the model through llama.cpp with a Metal GPU backend, achieving fully offline AR image generation on iPhone. Code and demo are available at https://anonymous.4open.science/r/AR_Quant/.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
2-bit quantization, Autoregressive image generation, Mobile phone deployment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rgKDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a 2-bit quantization framework for autoregressive image generation, combining existing complex-valued QAT with weight channel reordering and sensitive-layer preservation. Experiments on LlamaGen and VAR show improved generation quality, while iPhone deployment demonstrates model compression and reductions in latency and energy use.

Critical strengths
  1. WCR provides a simple and useful idea for complex-valued quantization. It connects channel reordering with the axis-aligned codebook to improve weight pairing.

  2. The paper implements a complete pipeline from complex-valued QAT to mobile inference using Metal kernels. The iPhone experiments show substantial model compression and measurable reductions in latency and energy use.

Critical weaknesses
  1. The main methodological contribution is WCR, but its advantage over simpler channel sorting methods is not sufficiently established. In Table 4, mean absolute magnitude and RMS achieve axis-alignment results close to WCR, but their final FID is not reported. Comparing these strategies under identical QAT settings, with repeated runs to assess small differences, would clarify whether the proposed scoring rule offers benefits beyond simple statistical sorting.

  2. Sensitive-layer preservation accounts for a substantial part of the gains, but its contribution needs clearer evaluation. In Table 3, preserving sensitive layers reduces FID from 4.72 to 3.54, while adding WCR alone reduces it to 3.96. The paper should clarify whether all baselines preserve the same layers and compare them under the same setting.

  3. Although the full method improves results on both VAR and LlamaGen, these results do not show how much WCR contributes on LlamaGen. A component ablation on a LlamaGen model would clarify whether WCR’s gains extend beyond VAR.

  4. The writing and figures do not explain WCR clearly. The connection between channel scoring, weight pairing, and complex mapping is unclear, making the method and its rationale difficult to follow.

Questions for authors
  1. Could the authors compare the final FID of WCR, mean absolute magnitude, and RMS sorting under the same QAT settings?

  2. Does WCR reduce weight reconstruction or layer-output error, beyond improving axis alignment?

  3. Could the authors clarify whether all baselines preserve the same sensitive layers? If the settings differ, how much improvement remains when the same layers are preserved?

Reviewer_Cx8fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets on-device deployment of autoregressive image generators, where storage, latency, and energy constraints call for ultra-low-bit weights. The authors argue that real-valued 2-bit QAT methods such as BitNet b1.58 and EfficientQAT lack the representational capacity needed for fine-grained image generation. Building on the complex-valued QAT of Fairy2i, which quantizes weights to {±1,±i}\{\pm1, \pm i\}, they propose Weight Channel Reorder (WCR). WCR sorts and interleaves output channels so that paired weights project closer to the quantization axes. A sensitivity analysis motivates keeping the embedding and output layers in full precision. On LlamaGen and VAR, the 2-bit models achieve lower FID than the compared 2-bit QAT baselines (e.g., FID 2.83 vs. 2.62 for full precision on LlamaGen-XL). The authors also deploy the models on an iPhone through a GGUF export and llama.cpp's Metal backend.

Critical strengths
  1. The paper delivers a complete, working pipeline from quantization-aware training to fully offline generation on a phone. The authors export the 2-bit complex-quantized models to GGUF, integrate them with llama.cpp/ggml, and run them on the Apple GPU through Metal. They report model size, latency, and energy on an iPhone 17 Pro Max. To my knowledge, this is the first public demonstration of 2-bit autoregressive image generation on a mobile device. That engineering effort has practical value for the on-device generation community.

  2. The 2-bit models retain image quality surprisingly well across two autoregressive paradigms and four model sizes. On ImageNet 256×256 with 50K samples, LlamaGen-XL reaches FID 2.83 vs. 2.62 at full precision, and VAR-d20 reaches 3.28 vs. 2.69, consistently ahead of Fairy2i, BitNet b1.58, and EfficientQAT. WCR is a simple permutation with no inference cost. The ablation in Table 3 shows that it helps both with and without full-precision layers. This suggests that complex-valued QAT is a promising direction for extremely low-bit visual generators.

Critical weaknesses
  1. The technical novelty is limited, and the main comparisons do not isolate the proposed contribution. The lossless widely-linear real-to-complex mapping, the {±1,±i}\{\pm1, \pm i\} codebook, the recursive residual quantization, and the STE-based training all come from Fairy2i. The new components are therefore WCR and the decision to keep the embedding and output layers in full precision, which is common practice in low-bit quantization. In the main tables, the Fairy2i baseline quantizes all layers, while "Ours" keeps word_embed and head in full precision. Under the same layer preservation, Fairy2i reaches 3.54 vs. 3.28 for Ours. The gain attributable to WCR is therefore about 0.26 FID, not the 1.44 reported in Section 4.2. The comparisons with other baselines are also hard to interpret. It is not stated whether BitNet b1.58 and EfficientQAT kept the same full-precision layers, and BitNet uses fewer bits. The 6-bit PTQ baselines appear to quantize activations as well, while the proposed method keeps BF16 activations. Without matched settings, it is difficult to judge how much WCR improves over existing complex-valued QAT.

  2. The efficiency claims are stronger than what the measurements establish. The reported model sizes (2293.76 MiB for VAR-d20, about 2959 MB for LlamaGen-XL) correspond to FP32 weights. The 12.3× and 9.7× compression ratios are therefore relative to FP32 and they would be roughly half that relative to the FP16/BF16 models normally deployed. The measured speedup is only 1.24×, on a single model and a single device. Energy is estimated from battery drain. Latency for LlamaGen is not reported, and neither is the effective bit-width including scaling factors and the residual stage. Most importantly, the natural deployment baseline is missing: running the same models with llama.cpp's existing low-bit GGUF formats (e.g., Q4_K or Q2_K), in both quality and latency. Without it, it is unclear whether the complex-valued format offers a practical advantage over off-the-shelf mobile quantization.

Questions for authors
  1. Did all QAT baselines keep the same layers in full precision as the proposed method (word_embed/head for VAR, the output layer for LlamaGen) and use the same 25-epoch schedule? Can the authors report Fairy2i with matched layer preservation on all four models? Do the 6-bit PTQ baselines also quantize activations?

  2. What is the exact effective bit-width per weight, including scaling-factor granularity, the residual stage, and the full-precision layers? How does this compare with the budgets of BitNet b1.58, EfficientQAT, and Fairy2i? This would clarify whether the comparisons are made at matched storage.

  3. How do the 2-bit models compare with llama.cpp's standard GGUF quantization of the same models (e.g., Q4_K, Q2_K) on the same iPhone, in FID, model size, latency, and energy? What precision does the "FP" deployment baseline use? Why is the speedup limited to 1.24×?

Reviewer_Xa6oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper applies the complex-valued QAT from Fairy2i to class conditional Autoregressive Image generation models, specifically LlamaGen and VAR. In the process of this application, the authors find that a simple "Weight Channel Reordering strategy" helps the quantization in the visual domain and improves the visual quality of the generation. Unfortunately the paper fails to truthfully delineate derived application v/s their proposal honestly and appropriately and needs major introspection in its exposition. More in the Weaknesses section.

Critical strengths
  1. The paper is mostly well written. It is easy to follow, and doesn't need too much effort to digest.
  2. The experiments are sound, and demonstrate the improvements obtained using their proposed simple weight channel reordering.
Critical weaknesses
  1. Quantization is important per-se, but this paper's motivation for why Quantization is important poorly defined. Check the section below on Introduction/Motivation related flaws. There are various factual flaws in the paper's positioning and I strongly recommend fact-checking and rewriting.

  2. Contributions are overly oversold. Regarding the actual method, I detail what's been obfuscated and wrongly attributed in the following points, here I strongly recommend removing the contribution number 3 from this paper. Demonstrating "on-device deployment" engineering is a good to have, but cannot be claimed as a contribution of a research submission anymore. Especially given the ready availability of AI agents, and their capability of porting complex systems over to hardware [1] such as Apple Metal, I strongly vote against allowing submissions to claim this effort as a major contribution.

  3. Regarding the proposed method, the authors do not properly delineate what's new and what's being taken from previous papers. For instance line 160 - 198 are borderline plagiarism from Fairy2i [2], including equation-by-equation copy. The problem isn't that the paper uses it, but that it calls it their proposal, instead of calling it a preliminary section.

  4. Same with the section on Hierarchical quantization strategy: This is basically HAWQ’s second-order perturbation analysis [3], HAWQ-V2’s HVP-based aggregate-curvature layer ranking [4], and FIT’s empirical-Fisher-trace sensitivity metric [5]. None of these are properly pointed out in the this paper, and are appropriated as the paper's own original proposals.

  5. Weight Channel Reorder appears to be the paper’s only original algorithmic contribution. The method is simple and technically motivated: output channels are ranked by the magnitude of their signed mean and interleaved so that adjacent ranked channels form complex pairs, with the objective of improving alignment with the four-axis complex quantization codebook. Although the proposed statistic is a coarse heuristic and does not directly minimize complex quantization error, the method is computationally inexpensive and appears empirically beneficial. Its novelty and effectiveness should be evaluated against random permutations, alternative channel statistics, and direct quantization-error-based pairing. Some simple, but important baselines, especially for a paper that is entirely resting on the shoulders of this weight channel reordering strategy alone.

  6. The experiments are mostly sound, and it is interesting that this simple WCR does improve what is otherwise just application of Fairy2i [2] for image generation, but as mentioned in 5, the simple permutation and pairing flavor baselines must be included to demonstrate the robustness of the WCR clearly.

Introduction/Motivation flaws:

  1. Lines 38-43 are highly speculative. I have not seen any direct need so far, for the years I have worked myself, so any references or surveys backing these statements would be helpful. Otherwise, consider rewriting the motivation or toning down the speculative claims. To my understanding, the Dhar et al 2021 survey covers old ML techniques and nothing about generative models image/video/audio or otherwise. Same with lines 44-48, they are highly speculative and under-backed. As far as I know, the highest demand in the industry at the moment is for hybrid deployments for generative models, and not for entirely edge deployments.

  2. Lines 49-53, Note that Quantization is just one technique that offers a low-effort-low-reward generic solution for on-device deployments. On the contrary, specialized model post-trainings, Efficient Attentions, FLOPs reductions, and Model pruning are still preferred over, or in conjunction with, off-the-shelf quantizers. Consider revision. This argument is further strengthened by Figure 4, where model size is drastically improved, but latency is not affected much. Thus as mentioned, Quantization is not a one-stop solution for improving all aspects of on-device deployments, and other efficiency techniques are equally important for deploying E2E complex generation systems on edge devices.

  3. Line 58, I'd recommend to stay away from power consumption for this paper. Because power consumption is a very complex phenomenon for on-device executions. For instance the authors correlate Matmul operations and power-consumption, which is very loosely true, but it turns out that power-consumption is actually most affected by memory transfers, and any gains if to be obtained lie there. Fact checking recommended. This argument is further strengthened by figure 4, where very minimal power consumption effect can be seen with Quantization.

References: [1] https://x.com/unixpickle/status/2109502228620840976?s=20 [2] Wang, F., Tan, X., Huang, B., Zhang, Y., Wang, G., Cong, P., & Yang, T. (2025). Fairy2i: Training Complex LLMs from Real LLMs with All Parameters in ${\pm 1,\pm i} $. arXiv preprint arXiv:2512.02901. [3] Dong, Z., Yao, Z., Gholami, A., Mahoney, M. W., & Keutzer, K. (2019). Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 293-302). [4] Dong, Z., Yao, Z., Arfeen, D., Gholami, A., Mahoney, M. W., & Keutzer, K. (2020). Hawq-v2: Hessian aware trace-weighted quantization of neural networks. Advances in neural information processing systems, 33, 18518-18529. [5] Zandonati, B., Pol, A. A., Pierini, M., Sirkin, O., & Kopetz, T. (2022). Fit: A metric for model sensitivity. arXiv preprint arXiv:2210.08502.

Justification: Overall, this paper is a Borderline for me, mostly inclining towards rejection actually, because of the notorious presentation and borderline plagiarism. But I am not opposed to giving the benefit of doubt to the authors here. I don't mind borderline accepting this paper, in spite of minimal incremental contribution. But the wrong claims and contributions must be fixed, which I am unsure if this is something that can be done within the scope of rebuttal. Generally interested in other reviews for this paper as well.

Questions for authors

None

1 comment

No response received from authors on this one yet.

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

Related papers

Discussion (0)

Sign in to comment.