mesmertech/Mesmer-Image-21-Nunchaku

Mesmer Image 21 Nunchaku

Built with Qwen. Independent rank128 Nunchaku quantizations of the Qwen Image 2.1 transformer for text-to-image generation and reference-image editing. Research and evaluation only under the Qwen Research License. See NOTICE for attribution and modifications.

File	Contents	Hardware
qwen-image-2.1-int4-r128.safetensors	Signed INT4 weights/activations, BF16 rank128 branches and boundary layers	Tested on RTX 4070 Ti SUPER with the custom runtime
qwen-image-2.1-fp4-r128.safetensors	Native NVFP4 weights/activations, BF16 rank128 branches and boundary layers	Tested on RTX 5090 (Blackwell SM12.0) with the custom runtime

Each file includes its transformer configuration and runtime manifest in safetensors metadata. This repository contains transformer weights only. A compatible custom Qwen Image 2.1 Nunchaku runtime and separately sourced text encoder, VAE, processor and scheduler are required. Stock Diffusers loading alone does not understand the packed custom linear format.

The FP4 file was converted separately from the original BF16 transformer and tested with generation, single-reference editing and two-reference composition. It is not a relabeled INT4 checkpoint.

The INT4 file consolidates the previously sharded calibrated transformer without changing any tensor value. It is not a lossless substitute for BF16: faces, hands, counts and layouts can drift. Fixed seeds are not guaranteed to produce identical pixels. Benchmarks and full-size comparisons include the measured local limitations; they are not cloud latency promises.

The older complete pipeline and reproduction files remain available at immutable revision 2b1c7e324511916189b1635b6622b40d54ae73f5 for existing pinned builds. New downloads should select the desired transformer file explicitly.