Uncensored Text Encoder for Krea 2 Released
nathandreamfast · reddit · 2026-07-16
The author applied "abliteration" to the `Qwen3-VL-4B-Instruct` text encoder used by **Krea 2**, removing its refusal tendencies, and packaged it as a checkpoint ready for ComfyUI. Key results from the post: - **HarmBench** compliance rate increased from 30.8% to 100% - In the low KL version, **GSM8K** dropped by only 1.83%, and **MMLU** remained virtually unchanged - Available in multiple quantization formats: INT8, FP8, bf16, NVFP4, MXFP8 The author also notes a caveat: since this is a text encoder rather than a chat model, "uncensored" doesn't mean the exact same thing as it does for standard generative models, but it remains fully functional within ComfyUI's visual understanding workflows.
Related event: Uncensored Qwen3-VL Text Encoder Released for Krea 2 and ComfyUI(2 posts)→
More from Multimodal
- LTX 2.3 LoRA demo changes a video’s camera angle — CQDSN · 2026-07-21
- Reddit user shares a surreal ChatGPT-generated poster — Creamy-Sundae-9991 · 2026-07-21
- A cinematic SEEDANCE 2 prompt turns an empty sunrise city into a memory-driven video — LudovicCreator · 2026-07-21
- Krea 2 Identity Edit transfers a karate pose from a line sketch without ControlNet — NatalieCrypto · 2026-07-21
- SenseTime unveils U1 Pro and open-sources a 50M-sample vision dataset at WAIC 2026 — 机器之心 · 2026-07-21
- Anatomy of Dynamic AI Images: Subject, Environment, and Camera — GPU_FieldNotes · 2026-07-21