OrbitQuant quantizes FLUX.1 and Wan 2.1 with no calibration data

A new preprint on arXiv makes a practical-sounding claim for anyone deploying open image and video generators: you can push diffusion transformers to very aggressive bit-widths without collecting a calibration dataset at all.

The method, called OrbitQuant, is aimed at a real headache in post-training quantization for diffusion transformers. Activations in these models shift across timesteps, prompts, and guidance branches, which forces existing PTQ approaches to re-fit calibration data for every new checkpoint or modality. The authors' route around this is to quantize in what they call a normalized, rotated basis. A randomized permuted block-Hadamard rotation reshapes each coordinate so it concentrates around one fixed, known marginal regardless of the input, which lets a single Lloyd-Max codebook cover every timestep, prompt, and layer of a given input dimension. The rotation on the weight side is absorbed offline, so at runtime only a forward rotation on the activations is left.

The headline number in the abstract is W2A4, two-bit weights and four-bit activations, on image diffusion transformers with what the authors describe as usable generation quality. They report state-of-the-art PTQ at several low-bit settings across FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX, and say the same recipe transfers from image to video with no per-modality tuning.

Take the specifics as reported, not settled. This is a v1 arXiv preprint from early July 2026, 'usable generation quality' is the authors' own framing, and the abstract itself does not give wall-clock speedups, memory footprints, or head-to-head sample quality scores against prior low-bit PTQ. Independent reproductions on real FLUX.1 and Wan 2.1 workloads, and on the inference kernels people actually run, will decide whether the rotated-basis trick holds up outside the paper's evals.

If it does hold up, the group that benefits most is the practitioners running open image and video models on constrained hardware. Skipping calibration means no per-checkpoint dataset to curate, and one recipe that carries from image models to video is the kind of ergonomic win that small teams doing self-hosted or on-device generation actually notice.