x0 velocity tested: 64x latent compression model trains to clean samples in one GPU-day
ostrisai · x · 2026-10-22
Following his x0 velocity proposal, ostris shows a first real test: converting the MOSS image model into a 64x latent compression model, applying patch 8 on the latent instead of patch 2.
- Both models trained just one day on a single GPU
- Comparison samples show the flow-matching version still produces noisy painterly images, while x0 velocity yields clean portraits and animals at the same deep compression
- Training is ongoing; early but直观 results supporting his claim of better high-frequency tolerance and deeper compression viability
More from Multimodal
- LiDAR scan + LLM workflow produces dimensionally accurate 3D room models — OpenAIDevs · 2026-10-22
- Gradio distills FLUX.3 Klein's 9B prompt upsampler into 0.8B and 2B models that run on CPU — Gradio · 2026-10-22
- Pocket upsampler Arena: Gradio's distilled 0.8B planner goes head-to-head with the 9B default — Gradio · 2026-10-22
- Gradio shows off gr.Workflow by distilling FLUX.3 Klein's 9B upsampler into 0.8B and 2B planners — Gradio · 2026-10-22
- MERZraum: A Poetic 3D Project Built With AI Gaussian Splatting — Merzmensch · 2026-10-22
- ComfyUI Tutorial: Model Downloads, Folder Structure and Loaders Explained — pixaromadesign · 2026-10-22