Removing a single learned "refusal" direction in the text encoder separates chat-style refusal behavior from the image generation pipeline. That means you can run the encoder as a chat-capable LLM that rarely refuses, while the image denoiser — which actually determines what gets drawn — remains unchanged and keeps image outputs consistent.
What Sets It Apart
- Refusal removal, not a new denoiser: the release edits the Qwen3‑VL‑8B‑Instruct encoder to project out a refusal direction; the 7B Qwen-Image denoiser is unchanged, so the model does not gain the ability to draw content the denoiser never learned.
- Measured effects: English chat refusals drop from 88.9% to 1.2%, Russian refusals from 38% to 0%, while MMLU accuracy remains 77.35% (no meaningful change). On a 45-prompt visual probe, 0 of 45 sensitive prompts flipped versus the stock encoder; LPIPS change on neutral prompts ≈ 0.084–0.090 (small, general shift).
- Multi-format builds and quantization trade-offs: BF16 and several AD/quantized GGUF builds are provided (sizes span ~16.4 GB down to ~3.3 GB); lighter quant types move images more, AD-Q4_K is offered as a common 4-bit compromise with smaller visual drift than plain quantization.
Who it's for & trade-offs
Great fit if you need a local text encoder that behaves like a chat model without frequent refusals and that plugs into stable-diffusion.cpp or similar GGUF pipelines for experiment or red-team corpus generation. It is also useful when you want consistent image outputs while exploring encoder-level behavior. Look elsewhere if you expect the encoder alone to enable new visual capabilities or to overcome denoiser training limits — image content is governed primarily by the denoiser and its pretraining data. Also avoid deploying this as a safety mechanism: removing refusal direction reduces refusal tendencies but does not ensure outputs are safe, lawful, or ethically acceptable.
Where it fits
Use this encoder when you control the local generation stack (stable-diffusion.cpp, compatible VAE and denoiser) and need an encoder that will not steer prompts toward refusal. For different goals — e.g., enforcing safety filters, running hosted constrained services, or seeking denoiser-level changes — choose other components (a moderated service, a safety checker, or a differently trained denoiser).
