How 1.58‑bit Quantization Cuts FLUX Parameters by 99.5% While Matching Full‑Precision Quality

Highlights

1.58‑bit FLUX : First model that compresses the 11.9 B‑parameter FLUX visual transformer to 1.58‑bit, cutting parameters by 99.5% without using image data.

Efficient linear kernel optimized for 1.58‑bit arithmetic, delivering notable memory savings and inference speed‑up.

Benchmark results on GenEval and T2I CompBench show generation quality comparable to full‑precision FLUX.

Problem Statement

Current text‑to‑image (T2I) models such as DALL·E 3 and Stable Diffusion 3 contain billions of parameters, leading to high memory consumption during inference and making deployment on resource‑constrained devices (e.g., mobile phones) impractical.

The paper investigates whether extreme low‑bit quantization (1.58‑bit) can reduce storage and memory footprints while maintaining inference efficiency and visual quality.

Proposed Solution

Target model: FLUX.1‑dev . Apply post‑training quantization to compress all linear‑layer weights to 1.58‑bit (value set {‑1, 0, +1}) without accessing any image data.

Develop a dedicated low‑bit operation kernel that accelerates 1.58‑bit computations.

Technical Details

1.58‑bit weight quantization : Inspired by BitNet b1.58, weights are stored as 2‑bit signed integers, achieving extreme bit‑reduction.

Unsupervised quantization : Relies solely on FLUX.1‑dev’s self‑supervised mechanism; no mixed‑precision or extra training data required.

Custom kernel : Optimized for low‑bit arithmetic, reducing memory usage and inference latency.

Experimental Setup

Quantization data : A calibration set of 7,232 prompts composed from the Parti‑1k dataset and T2I CompBench training prompts. No images are used.

Evaluation datasets :

GenEval – 553 prompts, 4 images per prompt.

T2I CompBench validation – 8 categories, 300 prompts each, 10 images per prompt (24,000 images total).

All images generated at 1024 × 1024 resolution for both full‑precision and 1.58‑bit FLUX.

Results

Performance : On both benchmarks, 1.58‑bit FLUX matches full‑precision FLUX; tables (not reproduced) show negligible differences before and after applying the custom kernel.

Efficiency :

Storage: Model size reduced by 7.7× (from 16‑bit to 2‑bit representation).

Inference memory: Reduced by 5.1×.

Latency: Significant improvements on lower‑end GPUs (e.g., NVIDIA L20, A10), as illustrated in the accompanying figures.

Visual quality : On GenEval and T2I CompBench, generated images retain high fidelity and alignment with prompts, comparable to the original FLUX, though ultra‑high‑resolution detail remains slightly behind.

Conclusion and Discussion

The 1.58‑bit FLUX quantizes 99.5% of transformer parameters, achieving a 7.7× storage reduction and >5.1× inference memory savings while preserving benchmark‑level generation quality. The work demonstrates that extreme low‑bit quantization is feasible for large T2I models and encourages the community to develop mobile‑friendly kernels.

Current Limitations

Speed limits : Without activation quantization and more advanced kernel optimizations, latency gains are modest.

Visual detail : High‑resolution rendering still lags behind the full‑precision model; future research will aim to close this gap.

References

[1] 1.58‑bit FLUX