TensorSharp Supports Qwen Image Editing

fuzhongkai · reddit · 2026-07-13

TensorSharp demonstrates multi-image editing and virtual try-on based on Unsloth Qwen Image Edit 2511, listing required components: main DiT model, VAE, Qwen2.5-VL text/vision encoders, and optional Lightning LoRA.

The post also provides example launch parameters for TensorSharp.Server, noting it offers OpenAI/Ollama-compatible API and Web UI, runs on Windows, macOS, Linux, and uses CUDA/Metal/Vulkan acceleration.

Performance comparison: on Qwen-Image-Edit 2511 CUDA tests, warm total time 40.44s, outperforming stable-diffusion.cpp's 48.16s; faster in sampling, VAE encode/decode, but slower in text encode. The author adds it can compete with llama.cpp on some autoregressive LLM models.

Original post →

More from Infra

Infra channel →