sci4ai/Qwen-Image-2512-Ablit-TE-For-Musubi-Lora-Training

Qwen-Image-2512 Abliterated Text Encoder (Single File)

Abliterated text encoder for the Qwen Image 2512 diffusion pipeline in a single merged safetensors file. Works for both ComfyUI inference and musubi-tuner LoRA training.

ComfyUI Usage

Place ablit_qwen2.5_vl_7b.safetensors in ComfyUI/models/text_encoders/ and load with a standard CLIP loader node — no T5/CLIP split required.

Recommended launch flag:

python main.py --fp8_e4m3_text_encoder

FP8 E4M3 quantization is applied 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.

Musubi-Tuner Usage

Point your --text_encoder argument at ablit_qwen2.5_vl_7b.safetensors. The text encoder is frozen during training; swap it in at inference time to condition generation on the abliterated encoder.

Related Models

• sci4ai/Qwen-Image-2512-Abliterated-TextEncoder — Multi-shard version
• 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.