ComfyUI · NSFW · 18+
Abliterated Qwen3-VL: uncensored prompt reading
Both packs read your prompt through a Qwen3-VL language model. The Image Pack ships Huihui-Qwen3-VL-4B-Instruct-abliterated. The Video Pack ships the Qwen3-VL 32B variant of MiniMax-H3. Abliteration removed the refusal behavior from these models, so your prompt reaches the sampler as written. Nothing leaves your machine.
What abliteration means
Qwen3-VL is the open vision-language model from Alibaba. Huihui AI publishes abliterated variants. They modified the weights directly to remove the safety filtering, and released the result on Hugging Face under the Apache 2.0 license.
For you this means the text encoder passes adult prompts through unchanged. There is no server in the loop that could refuse, because there is no server at all.
Where it sits in the pack
| Pack | Text encoder | Quantization |
|---|---|---|
| Image Pack | Huihui-Qwen3-VL-4B-Instruct-abliterated | fp8_scaled, 5.2 GB |
| Video Pack | Qwen3-VL 32B (MiniMax-H3) | nvfp4 awq, runs on any RTX 30+ |
The files install into models/text_encoders inside your ahegao-workspace. The Qwen Image VAE handles latent decoding in the Image Pack.
Why a 4B model is enough
The 4B encoder reads dense, block-structured prompts well. It keeps the Image Pack under a 16 GB VRAM budget, so it runs on an RTX 4070 Ti Super or an RTX 4080. The 32B encoder in the Video Pack reads long temporal descriptions. That is what scene-to-soundtrack matching needs.
Adults only (18+). All sample output is synthetic, computer-generated fiction. Last verified October 2026.