license: other
license_name: qwen-research-license
base_model: Qwen/Qwen-Image-2.1
tags:

• text-to-image
• image-editing
• qwen-image
• diffusers
• uncensored

Qwen-Image-2.1-Uncensored-step32000

Merged LoRA checkpoint of Qwen-Image-2.1, step 32000.

Merge info

{
  "base": "Qwen-Image-2.1",
  "lora": "qwen21_full1024_r128-step00032000.safetensors",
  "scale": 1.0,
  "modules": 224,
  "network_dim": "128",
  "network_alpha": "128.0"
}

Contents

• transformer/ — diffusers QwenImage21Transformer2DModel (bf16, sharded)
• text_encoder/ — Qwen3VLForConditionalGeneration (bf16)
• vae/ — AutoencoderKLQwenImage21
• processor/, scheduler/, model_index.json
• merged/ — single-file merged transformer (qwen_image_2.1_step32000_merged_bf16.safetensors)

Usage

from diffusers import QwenImage21Pipeline
import torch

pipe = QwenImage21Pipeline.from_pretrained(
    "chatrlabs/Qwen-Image-2.1-Uncensored-step32000",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")
image = pipe(prompt="a cinematic film still", num_inference_steps=25, guidance_scale=1.0).images[0]

For ComfyUI GGUF, see chatrlabs/Qwen-Image-2.1-Uncensored-step32000-GGUF.

Verification

• 297 tensors, 32 blocks, bf16; img_in, txt_in, attn.norm_q/k preserved.
• Quantized versions: see the GGUF repo (Q8_0 ~0.565% relative RMSE, Q4_0 ~8.3%).
• s32000 is the recommended step; s90000 is known to be less stable on text rendering.

License

Released under the Qwen Research License Agreement. See LICENSE. You are responsible for complying with its terms, including any restrictions on commercial use and redistribution.