🖼️ Autoregressive Image Generation on a low-cost GPU

Janus-Pro-1B generating images as autoregressive tokens on a single low-cost NVIDIA T4 (~$0.6/hr) — about about 25 seconds per image. VIDRAFT's efficiency line brings capable image AI to modest, affordable hardware — no datacenter GPU needed.

📊 VKUE Leaderboard 🔵 Text: 34.7B, GPU-less 🆓 Text: 34.7B on FREE CPU (~6-7 tok/s) 🤗 Model
Checking model status…
Your image will appear here — generated in ~25s on a single T4 GPU.
Autoregressive image tokens (576) → VQ decoder → 384×384 image, on one low-cost GPU. Honest scope: this uses a small T4 GPU (not GPU-less) — image generation needs a GPU today, and a cheap one makes it accessible. Janus-Pro-1B is a dense model. The GPU-less VKUE demos are for text (see the 34.7B-on-CPU badge). Model: deepseek-ai/Janus-Pro-1B.