license: other
license_name: qwen-research
license_link: LICENSE
base_model: huihui-ai/Qwen2.5-VL-3B-Instruct-abliterated
library_name: diffusion-single-file
tags:

• comfyui
• text-encoder
• omnigen2
• abliterated
• uncensored
• experiment

Qwen2.5-VL-3B abliterated — fp16 text encoder for OmniGen2 in ComfyUI

huihui-ai/Qwen2.5-VL-3B-Instruct-abliterated repackaged with the tensor names, shapes and dtype (fp16) of ComfyUI's
qwen_2.5_vl_fp16.safetensors, so it drops into the OmniGen2 workflow in place of the original text encoder. The
abliteration is huihui-ai's work; I only converted bf16 to fp16 and checked the names.

License: Qwen RESEARCH LICENSE (non-commercial: research or evaluation only). Modified files; see LICENSE and NOTICE.
This is an uncensored model: refusals were removed from the text part, not the image part. You are responsible for what you do with it.

What changed in the file

• 824 tensors, same names and shapes as ComfyUI's fp16 encoder; 71 of them differ in value: embed_tokens, 35 o_proj and 35 down_proj
  (the layers abliteration touches). Largest absolute difference 0.146. No inf/nan after the bf16 → fp16 cast.
• Put it in models/text_encoders, load it with CLIPLoader type omnigen2.

What it does to OmniGen2

The conditioning changes for every prompt, not only the refused ones, and by different amounts. The same seed with the
original and the abliterated encoder (three neutral prompts, DiT fp16, 20 steps):

PSNR between the two encoders on the same prompt and seed: 12.0–19.0 dB over six images. Some images change a lot, others little, and I
have not found out why; in my review it changes the result in a way I had not seen before. Treat it as an experiment. Time per step is unchanged (1.44 s against 1.42 s).

Credit: abliteration by huihui-ai; model by the Qwen team (Alibaba Cloud).
Tools and the measurement log: https://github.com/JoaoZaokk/comfy-quant-bench.