Lucas Image 2.1 Turbo Q2_K + Sakura encoder

Built with Qwen. Non-commercial research/evaluation use only.

A repackaged pipeline based on Qwen-Image-2.1-Turbo, containing Unsloth's existing Q2_K diffusion GGUF, Sakura's 4-bit modified text encoder, and the upstream VAE. This publisher has not fine-tuned or abliterated these files. All three weight files are copied byte-for-byte from the pinned sources in provenance.json.

The image generator is 7B-class; Q2_K describes quantization, not parameter count. The separate text encoder is 8B-class and Q4_K_M, not Q2_K. This is a three-component pipeline, not one combined GGUF.

Sakura reports changing the encoder and quantizing the diffusion weights. Its label “uncensored” does not establish a measured image-content effect. No such result is claimed for this package. Two-bit diffusion quantization can affect image quality.

Components

File	Role
`qwen-image-2.1-turbo-Q2_K.gguf`	Unsloth Q2_K diffusion weights
`Sakura-TextEncoder-Qwen3VL-8B-Uncensored-4bit-4.68GiB.gguf`	Sakura modified Q4_K_M text encoder
`qwen_image_2.1_vae_bf16.safetensors`	VAE

Use all three files with a compatible stable-diffusion.cpp build. Sampling starts with 8 steps, CFG 1.0, Euler, and the upstream Turbo sigma schedule. smoke-test.json, when present, records only a basic neutral-prompt load/generation check, not an uncensoring or quality benchmark.

Sources and license

The Qwen Research License in LICENSE applies. Original notices are retained in NOTICE and licenses/. Qwen is used descriptively; this is an independent Lucasgitguy-co package, not a Qwen/Unsloth/Sakura-endorsed release. Changes by this publisher: packaging, checksums, provenance, and documentation only.