Qwen-Image-2.1 + Heretic Text Encoder (Colab / Gradio)

One-click Google Colab notebook and a standalone Gradio app that runs Qwen/Qwen-Image-2.1 with the community Heretic (abliterated) text encoder:

• Transformers bf16 TE: pottokao/Qwen-Image-2.1-Text-Encoder-Heretic
• Not the ComfyUI-only GGUF: pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF

Open in Colab

1. Open the notebook (Open-in-Colab badge / URL from this repo once published).

2. Runtime → Change runtime type → GPU (T4 is OK with CPU offload).

3. Runtime → Run all.

4. Wait for model downloads (~25–35 GB+ on first run; can take 15–40+ minutes).

5. In the last cell output, copy the line:

   Running on public URL: https://xxxx.gradio.live

   That is your temporary Gradio share link (~72 hours). It is not a permanent hosted demo.

Open-in-Colab

Direct URL:

https://colab.research.google.com/github/staed-web/qwen-image-21-heretic-colab/blob/main/Qwen_Image_2_1_Heretic_Gradio.ipynb

Files

Hugging Face Spaces

1. Create a new Space (Gradio SDK, GPU or ZeroGPU).
2. Upload app.py + requirements.txt (+ this README).
3. ZeroGPU: app.py wraps generate with @spaces.GPU when spaces is importable.
4. Expect a long cold start (large model download). Free Spaces may OOM or time out; Colab with T4 + offload is the more reliable free path.

There is no permanent live Gradio URL shipped with this repo unless you deploy it yourself.

How the Heretic TE is loaded

import torch
from transformers import Qwen3VLForConditionalGeneration  # fallbacks in notebook/app
from diffusers import QwenImage21Pipeline

text_encoder = Qwen3VLForConditionalGeneration.from_pretrained(
    "pottokao/Qwen-Image-2.1-Text-Encoder-Heretic",
    torch_dtype=torch.bfloat16,
)
pipe = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1",
    text_encoder=text_encoder,
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()  # needed for Colab free T4 (~15 GB)

UI features

• Prompt + optional reference image (edit / I2I via image= on the pipeline)

• Aspect ratios tuned for Colab (1024 default, 768, 1280×720, etc.) — not official 2048×2048

• Steps (default 25), seed / randomize

• Optional Transparent RGBA checkbox that wraps the prompt with the official template:

  This is an RGBA image with transparency. {prompt}. The image has alpha channel and the background is transparent.

Caveats (honesty)

Official free demos (stock encoder, not Heretic)

• Blog: https://qwen.ai/blog?id=qwen-image-2.1
• Mainland China free UI: https://wuli.art/explore

Install (local)

pip install "torch>=2.4.0" "transformers>=5.17" accelerate pillow gradio spaces
pip install git+https://github.com/huggingface/diffusers
python app.py
# or: GRADIO_SHARE=1 python app.py