Steering Qwen-Image-2.1 with Qwen3-VL's internal states, no training needed
zacharynado · x · 2026-11-03
A demo chart shows a clever no-training combo: feed Qwen3-VL-8B's internal hidden states at one of its later layers straight into Qwen-Image-2.1 to generate images. The LM is prompted to "strongly visually imagine" a concept (a giraffe, loneliness, your favorite food) without naming it; the image model reads those internal states and draws what the model 'pictured'. Comparisons across prompt styles and layer choices show that vivid-visualization prompts and later layers yield far more on-topic images, and even abstract concepts produce meaningful results—effectively visualizing a model's mind's eye.
Related event: Qwen3-VL Hidden States Drive Qwen-Image Generation Without Training(2 posts)→
More from Multimodal
- FLORA launches 4-week AI filmmaking course: $250, final 30-60s film deliverable — round · 2026-11-03
- MiniMax H3 image gen melts far-away faces' eyes, user complains — Ninndzaa · 2026-11-03
- Open-Source Video Frame Sync Fixes Dropped/Duplicate Frames in AI-Generated Video — pftq · 2026-11-03
- Creator: Seedance 2.5's control is so good that AI video is 'essentially solved' — Uncanny_Harry · 2026-11-03
- Watercolor Lord of the Rings image set via Midjourney --sref codes — sergeantsref · 2026-11-03
- Matt Wolfe Covers a Free, Open-Source World Model — A Genie 4 Alternative — Matt Wolfe · 2026-11-03