Hugging Face Space

Qwen-Image 2.1, uncensored.

A generate-and-edit demo for the quantized checkpoint. Text in, image out — or an image plus an instruction.

Not generating yet. This account was created today, so Hugging Face will not start a Gradio or ZeroGPU Space (HTTP 402) until it is 30 days old, on PRO, or a community GPU grant is approved. A grant request is on this Space’s Community tab. The Gradio app is already in app.py.

What will run

ZeroGPU is PyTorch. The GGUF, FP8, NVFP4, and MLX files in the repo are for ComfyUI or Apple Silicon, so the app loads official Diffusers weights and the 32 MB uncensored LoRA shipped in the same repo.

  • Base: Qwen/Qwen-Image-2.1
  • LoRA: qwen-image-2.1-uncensored-lora.safetensors
  • Modes: text-to-image, and edit with up to 10 references
  • About 1 megapixel, 28 steps by default, no safety checker

18+ only. You are responsible for prompts and outputs. Examples in the app stay ordinary (signage, a fox, a sticker).

Model repo Similar live demo

From the model card

Benchmark collage published with the quantized model

Published benchmark image from the model card, not a fresh generation from this Space.

Files in the repo

UseFileSize
This Space (Diffusers LoRA)qwen-image-2.1-uncensored-lora.safetensors32 MB
ComfyUI, recommendedqwen-image-2.1-UC-Q4_K_M.gguf4.60 GB
ComfyUI safetensorsqwen-image-2.1-UC-fp8.safetensors6.63 GB
NVIDIA FP4qwen-image-2.1-UC-NVFP4.safetensors4.20 GB
Apple MLXqwen-image-2.1-UC-MLX-4bit.safetensors4.00 GB

Local GGUF needs ComfyUI plus leejet/ComfyUI-GGUF, the INT8 text encoder, and the BF16 VAE from the model card.