vlens: turn Qwen3-VL hidden activations into images via FLUX.2 klein, read back by the same model
Sauers_ · x · 2026-11-02
Open-source vlens builds a "vision autoencoder": it turns one hidden activation of Qwen3-VL-4B-Instruct (residual stream at layers 9/18/27, last position, 7,680 numbers) into a picture with no text in the image path, then has the same model look at the picture and compares the new activation to the original via squared error.
Key details:
- The generator is FLUX.2 [klein] 4B, a 4-step distilled rectified-flow model; it gets no prompt or word embeddings — a small trained adapter computes all its conditioning from the activation
- Both large models stay frozen; only the adapter is trained on the readback-matches-original objective
- Effectively visualizes the "mental image" inside a language model's hidden state with a closed-loop fidelity check
More from Multimodal
- Game dev swaps 3D models for AI-generated sprites to keep animation frames consistent — AIFlow_ML · 2026-11-02
- Midjourney textures get a retro Dreamcast makeover via custom workflow — AIandDesign · 2026-11-02
- Video gen on AMD 9070 XT: 5 seconds of 480p takes 8 minutes with MiniMax H3 — darthrehad · 2026-11-02
- Redditor shares cinematic AI-generated original music video — Hellish_NDE · 2026-11-02
- I had a dream, then I generated it: the Backrooms as a waterslide — IamIfOnly · 2026-11-02
- Moonlit museum video demo made with Seedance 2.5 on Dreamina — LudovicCreator · 2026-11-02