I added Qwen-Image 2.1 + LoRA support to TensorSharp (GGUF, local inference)

I maintain TensorSharp, an open-source inference engine. It can now run Qwen-Image 2.1 locally for text-to-image generation and image editing, with support for its LoRA adapters.

I’ve added configs for regular style and editing LoRAs, plus accelerated adapters with their own sampling recipes. For example, Pruna 8-step runs at 8 steps, and Viggle Turbo uses 6 transformer passes. Those are fewer model passes, not a claim of a measured end-to-end speedup on particular hardware.

Model files:
Qwen-Image 2.1 GGUF
Qwen-Image 2.1 VAE
Qwen3-VL-8B-Instruct GGUF and vision projector

LoRAs you can try:
Pruna 8-step / 5-step
Viggle Turbo
Qwen-Image-2.1-Fix (DoRA)
Film Stills
Object Remover Bbox

The base config specifies the exact files to download.
To try the Pruna adapter:
TensorSharp.Cli --config config/qwen-image-2.1.json \
--lora config/lora/qwen-image-2.1-pruna-8step.json \
--prompt "A small bookstore on a rainy evening" \
--width 1024 --height 1024

Repo: https://github.com/zhongkaifu/TensorSharp

If you’re running Qwen-Image 2.1 locally, I’d be curious which LoRAs you’ve found useful and how the accelerated ones compare for your prompts.