TL;DR for operators
Reducing numerical precision can make an image generator less computationally demanding, but where precision is retained can matter as much as the average bit-width. HGA-Quant1 addresses two decisions: which weight-reconstruction errors deserve protection, and which generation stages should receive higher-precision activations.
On Infinity-2B, protecting early generation stages delivers an ImageReward score of 0.83523 at 5.5711 effective activation bits, compared with 0.44577 at 5.5543 bits when precision is concentrated late. Yet the same early-stage scheduling approach substantially worsens FlexVAR image quality. Aggressive weight-tail protection also fails under some extreme low-bit configurations.
For teams compressing existing visual generators, the research supports treating quantization as a model-specific allocation problem. Benchmark the weight-calibration objective and stage schedule independently, enforce explicit image-quality thresholds, and measure actual hardware performance. The paper establishes improvements in generation quality under selected numerical budgets, not verified latency or energy savings.
Two Similar Precision Budgets, Two Very Different Images
An engineering team needs to run an image generator with fewer computational resources without compromising the quality customers expect. Reducing numerical precision is one option. But choosing an average number of bits does not determine which calculations should retain those bits.
This distinction becomes consequential when an image is generated progressively. Errors introduced while establishing its overall structure may affect everything produced afterward. Errors introduced during later refinement may have more limited consequences.
The HGA-Quant study makes this distinction measurable through a controlled activation-allocation comparison on Infinity-2B.
| Infinity-2B configuration | High-precision activation stages | Effective activation bits | ImageReward ↑ |
|---|---|---|---|
| Early-high allocation | 1–11 | 5.5711 | 0.83523 |
| Late-high allocation | 13 | 5.5543 | 0.44577 |
| Uniform A6 | All stages at A6 | 6.0000 | 0.82882 |
Source: Appendix D, Table 4, and Section 4.2, Table 1. The mixed schedules use A8 for listed stages and A4 elsewhere, with W4 weights. Higher ImageReward is better.
The early-high and late-high schedules have nearly identical effective activation precision. Their ImageReward scores differ by approximately 0.3895. The early-high schedule even slightly exceeds uniform A6 despite using a lower effective precision.
The effective bit-width is calculated by weighting each stage’s assigned precision according to its estimated linear-layer computational work. It is not a native 5.6-bit format.
The appendix comparison is a placement test rather than proof that the early-high schedule is globally optimal. Nevertheless, it establishes a consequential point: an average precision budget does not sufficiently characterize generation quality.
Protecting Early Calculations Before Later Refinement
Autoregressive visual generators construct images through successive coarse-to-fine stages. Earlier stages establish broad structure and conditioning context; later stages refine details, generally processing increasingly dense image-token grids.
HGA-Quant’s Coarse-to-Fine Activation Precision Scheduling (CAPS) exploits this ordering. It assigns 8-bit activations to a consecutive prefix of early stages and 4-bit activations to the remaining stages, subject to a FLOPs-weighted precision budget. The prefill stage retains high precision separately.
The method uses a greedy allocation rule: retain high precision from the beginning of generation until the next stage would exceed the budget. This is computationally straightforward, although it does not optimize arbitrary combinations of stages.
A separate sensitivity experiment on Infinity-2B helps explain the choice. Replacing A8 with A4 for the earliest consecutive stage pair reduces ImageReward by 0.17448. Applying the same precision reduction to stages 11–12 reduces it by only 0.00442.
This controlled perturbation supports the hypothesis that early numerical errors can propagate through subsequent generation. It also explains why the budgeted early-high schedule outperforms the late-high alternative.
For inference engineers, the contribution is a scheduling principle worth testing: allocate precision according to observed stage sensitivity and computational cost, rather than applying one activation bit-width everywhere.
The qualification is that the measured sensitivity belongs to the tested generator. It cannot automatically determine the appropriate schedule for another architecture.
Weight Calibration Has Its Own Allocation Problem
Stage scheduling addresses activation errors during generation. HGA-Quant’s other contribution concerns how pretrained model weights are represented at low precision.
A quantizer replaces continuous weight values with approximations drawn from a limited set of representable numbers. Conventional reconstruction objectives often minimize squared error across all coefficients. Because many weights cluster near the center of their distributions, errors among those numerous central coefficients can dominate calibration.
But numerical frequency does not necessarily reflect a coefficient’s influence on generated images.
HGA-Quant introduces Tail-Aware Bounded-init Grid Refinement (T-BGR). For each output channel, it identifies tail coefficients whose distance from the channel mean exceeds three standard deviations in the primary configuration. Their reconstruction errors receive a higher weight—eight times the central-coefficient weight under the reported default setting.
The algorithm then refines clipping intervals and quantizer parameters against this weighted reconstruction objective.
Crucially, T-BGR changes the calibration objective, not the final weight representation. It does not preserve tails through an additional full-precision computation branch.
The component ablation provides evidence for this design. With 4-bit weights and full-precision activations on Infinity-2B:
- Ordinary bounded grid refinement achieves ImageReward 0.79494.
- Restoring all detected tail coefficients to full precision increases the score to 0.83709.
- T-BGR with tail weighting factor eight achieves 0.84753, without that separate full-precision tail branch.
The restoration experiment is a diagnostic: it tests whether quantization errors in statistically rare weights materially affect output quality. T-BGR demonstrates that changing reconstruction priorities can recover quality without changing the low-bit storage format.
The main benchmarks reinforce the value of this calibration choice. At W4A4 on Infinity-2B, uniform T-BGR achieves ImageReward 0.44772 versus 0.25269 for QuaRot. At W4A6, it reaches 0.82882, exceeding the evaluated quantized competitors in that configuration.
These are comparative benchmark results, not guarantees that tail-aware calibration will improve every precision setting.
The Same Rules Can Fail on Another Generator
The strongest reason to avoid a universal HGA-Quant configuration comes from the paper’s own counterexamples.
On Infinity-2B, CAPS improves ImageReward from 0.82882 under uniform W4A6 to 0.83523 at 5.5711 effective activation bits. GenEval also rises from 0.68003 to 0.69043.
However, the corresponding mixed-precision schedules reduce quality on HART, VAR-d30, and FlexVAR.
The FlexVAR result is particularly consequential. Relative to uniform W4A6, CAPS at 5.5000 effective activation bits increases Fréchet Inception Distance (FID) from 2.9150 to 8.6066, where lower is better. Inception Score falls from 290.874 to 201.977.
This is a main cross-backbone comparison, not a peripheral sensitivity check. It shows that a lower average activation precision and an intuitively sensible early-prefix schedule can produce a substantial quality penalty.
Weight calibration exhibits a related tradeoff.
At W4A4 on Infinity-2B, setting the tail-weighting factor to zero yields ImageReward 0.50311, exceeding the default factor of eight at 0.44772. Under W3 weight-only quantization, ordinary bounded grid refinement scores 0.58618, whereas the stronger tail-weighted configuration falls to 0.32632.
These ablations reveal why preserving unusual coefficients cannot be the entire objective. Allocating more reconstruction capacity to tails can leave insufficient numerical resolution for the central mass of weights. The preferred balance changes with quantization severity.
Together, the experiments challenge two broad assumptions: that early-stage precision protection always helps, and that stronger tail protection always improves low-bit generation.
Turning Quantization Results Into Deployment Decisions
For teams operating image-generation systems, HGA-Quant suggests separating decisions that are often bundled into a single compression configuration.
The first decision is weight calibration. Compare ordinary reconstruction against tail-aware calibration at each intended weight and activation precision. The relevant question is whether protecting rare coefficients improves generation quality under the actual numerical constraints.
The second decision is activation scheduling. Establish a uniform-precision baseline, measure sensitivity across generation stages, and compare candidate allocations under approximately matched computational budgets. An early-high schedule is a reasonable candidate, not a universally established optimum.
The third decision is release acceptance. Evaluate every target backbone against the image-quality measures relevant to its intended workload. For prompt-conditioned generation, these may include alignment and preference-oriented scores. For distributional image-generation tasks, FID and Inception Score may be appropriate. A configuration should pass the product’s quality requirements before its numerical efficiency becomes a reason for adoption.
The official HGA-Quant implementation provides separate workspaces for Infinity, HART, VAR, and FlexVAR, along with controls for calibration and mixed-precision scheduling. This makes architecture-specific comparisons feasible, subject to the necessary checkpoints, environments, and evaluation assets.
The Boundary Between Computational Proxies and Hardware Savings
The experiments cover four pretrained visual autoregressive generators and report performance on established image-generation benchmarks. They use an NVIDIA RTX A6000 GPU, but the activation-efficiency analysis relies on estimated linear-layer FLOPs weighted by bit-width.
That measure does not directly establish end-to-end latency, throughput, energy consumption, or edge-device performance. It also does not account for every hardware-level cost associated with changing precision between stages.
The study does not provide comprehensive repeated-run uncertainty estimates for its principal benchmark comparisons. Its conclusions therefore concern the reported configurations and measurements, rather than guaranteed improvements across arbitrary architectures or operating environments.
For an engineering team making a deployment decision, numerical-budget results should motivate candidate configurations. Actual acceptance requires measuring inference performance and memory consumption on the target hardware while preserving required output quality.
Precision Should Follow Demonstrated Sensitivity
HGA-Quant’s contribution extends beyond demonstrating that 4-bit weights can support image generation. It shows that quantization quality depends on two allocation choices: which weight errors influence calibration, and which generation stages retain expensive numerical precision.
The Infinity-2B evidence makes precision placement difficult to ignore. The FlexVAR and extreme low-bit results make a universal allocation rule equally difficult to defend.
The resulting engineering principle is conditional but actionable: protect the numerical information that experiments show to be consequential, revisit that allocation for each model and bit-width, and distinguish estimated computational savings from measured deployment gains.
That is a more reliable foundation for image-generator compression than optimizing average precision alone.
Cognaptus: Automate the Present, Incubate the Future.
-
Xinmeng Liu and Xianglong Yan and Kaicheng Yang and Weisheng Dong and Yulun Zhang (2026). HGA-Quant: Heterogeneity-Aware Quantization for AutoRegressive Visual Generation Models. arXiv:2610.15297. https://arxiv.org/abs/2610.15297 ↩︎