jayleaton/localrouter-qwen-image-2.1

LocalRouter packs derived from Qwen-Image 2.1

Target: jayleaton/localrouter-qwen-image-2.1. For LocalRouter, not a general pipeline checkpoint. These packs are for non-commercial research or evaluation only. Free access does not grant commercial rights.

Licence and attribution

The weights, including the extracted text encoder and VAE, remain subject to the Qwen Research License. Section 2 permits derivatives and redistribution only for research or evaluation. Section 3 requires the licence, modification notices and upstream attribution. Commercial use needs a separate grant from Qwen. The upstream publisher is Hangzhou Tongyi Laboratory Technology Co., Ltd.; conversion is by the LocalRouter project. See the accompanying LICENSE and Notice for the upstream agreement and required attribution. LocalRouter's Apache-2.0 software licence does not relicense these weights.

Conversion

• nvfp4/: DiT linears converted with stk_twin to packed e2m1 codes, e4m3 block scales, global scales and static input scales calibrated with the committed prompts and scale tables. Side weights retain BF16 or FP32 precision.
• fp8s/: DiT converted to the TensorFold A8 path with static calibration scales; this is the shipped FP8 pack, not the older per-call scaling experiment.
• te/: extracted BF16 text encoder, renamed to the engine's operation names, with tokenizer and prompt metadata.
• vae/: extracted BF16 decoder weights, latent statistics and convolution metadata. No image encoder is included.

Format version: stk-pack/1. Converter: stk_twin from LocalRouter; source commit: d07c73c362270d52a37e93dade5ef943bb928d4f. Upstream checkpoint revision: d26bb61231c349cf6b7896fa83353113880e1ba3. These values must identify the actual conversion, not the upload. Each modified pack carries MODIFICATIONS.md. The upload preparation checks actual tensor bytes against every manifest SHA-256, checks both DiT manifests against committed reference digests, and writes SHA256SUMS for all files. Text encoder and VAE tensor digests are recorded in their manifests; there are no committed independent reference manifests for those two components. Static calibration affects output quality; NVFP4 is faster and can differ from BF16.

No hardware-specific Triton binaries are included. Generate those locally for the target GPU using the existing installer/converter path. This repository is an optional download source; LocalRouter still converts official checkpoints locally by default. Image edit is not supported by the current engine.