# Backpropagating through Qwen3-VL-8B's vision stack to the input image

> Source: <https://x.com/_smitop/status/2122859608066973939>
> Published: 2026-11-18 01:09:06+00:00

**Key takeaway:** Backpropagating through Qwen3-VL-8B's vision stack all the way to the input pixels lets you optimize an image toward a target, the technique behind the released Qwen3-VL Soft-Token Inversion Hugging Face Space.

loops!@_smitopbackproping through the Qwen3-VL-8B vision stack to the input image, to optimize the input image to make it like look like a target image00:001:09 AM · Nov 18, 2026·565Views131

loops!@_smitopbackproping through the Qwen3-VL-8B vision stack to the input image, to optimize the input image to make it like look like a target image00:001:09 AM · Nov 18, 2026·565Views131

loops!@_smitop14hQwen3-VL Soft-Token Inversion - a Hugging Face Space by syvbFrom huggingface.co22103

loops!@_smitop14h(kinda inspired by x.com/voooooogel/sta…)thebes@voooooogelNov 16deep dreamed images that activate neurons and chain lens readouts in qwen 3.6 27b42
