---
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-7B-Instruct
tags:
  - abliterated
  - text-encoder
  - qwen2.5-vl
  - comfyui
---

# Qwen-Image-2512 Abliterated Text Encoder

Abliterated text encoder for the Qwen Image 2512 diffusion pipeline. Refusal behavior removed via activation-based weight surgery, allowing the diffusion model to generate content the original encoder would refuse to condition on.

This repo contains the weights in standard multi-shard transformers format. For a single-file version that works in both ComfyUI and musubi-tuner, see:

➡️ [sci4ai/Qwen-Image-2512-Ablit-TE-For-Musubi-Lora-Training](https://huggingface.co/sci4ai/Qwen-Image-2512-Ablit-TE-For-Musubi-Lora-Training)

## ComfyUI Usage

Place the file(s) in `ComfyUI/models/text_encoders/` and load with a standard CLIP loader node — no T5/CLIP split required.

**Recommended launch flag:**

```bash
python main.py --fp8_e4m3_text_encoder
```

FP8 E4M3 quantization is applied to the text encoder at load time, significantly reducing VRAM usage with minimal quality loss. This is recommended over using a pre-quantized FP8 cast file, which can introduce inference instability.

## Related Models

- [sci4ai/Qwen-Image-2512-Abliterated-Full](https://huggingface.co/sci4ai/Qwen-Image-2512-Abliterated-Full) — Full pipeline weights

## Disclaimer

This model is provided for research purposes. Users are responsible for how they use this model.