chroma-radiance-x0 result a true pixel space model combining JiT-x0 loss and pixnerd give it a go! github.com/comfyanonymous/Co… huggingface.co/lodestones/Ch…

Today I am releasing Ostris Cloud. If you want to help support the open source development of AI Toolkit while renting compute, this is a great way to do that. Link in 🧵

this is pretty cool conditional protein diffusion model! biorxiv.org/content/10.1101/…

it's alive!

🎉 🎉 🎉 We're open-sourcing Chronicles-OCR, a visual perception benchmark evaluating VLLMs on ancient Chinese characters. The dataset spans 3,000 years of evolution. It covers 7 historical scripts from Oracle Bone to Cursive, featuring 2,800 balanced images across highly diverse physical media. We assess models on 4 core tasks: • Character Spotting • Fine-grained Recognition • Ancient Text Parsing • Script Classification The evaluation reveals how visual distribution shifts affect model perception over time. Explore the dataset and paper below. 👇 📄 Paper: arxiv.org/abs/2605.11960 🔗 GitHub: github.com/VirtualLUOUCAS/Ch…

brand new 2 x pro6000 thanks @ComfyUI

long boi

timestep shift 2 vs 6 (5000 steps fit, both using softclamp instead of RMSnorm)

Babe, wake up, new GemLite update. Up to 1.7x faster FP8 block quantization on the RTX PRO 6000 end-2-end in vLLM!

was debugging my x0 trainer using 1 image with lots of high frequency in it and got this interesting pattern. it seems to converge on high freq first before everything else. seems like x0 is doing frequency weighting implicitly based on your data distribution.

recursive model distillation be like

The US grid, mapped. Public data for public use. 16,819 power plants and 36,872 generators from EIA. 750,000 transmission circuit-miles from HIFLD. 1,000 data centers from EPA and other open sources.

yes that's the reason why you need to scale it near origin if you're using low precision but this not practical to do. float is not continuous it's just a really weird integer this also applies to model training you can see here the low precision struggle to reach the minima

you can just tweak hyperparam mid training and it will just works every spot where the LR resets is where i changed something mid training. just make sure you flush your optimizer state and re-spool its accumulator state and everything is golden

I was a huge fan of @Bose but now not anymore the product I already paid for, you quietly made it worse. No warning. No consent. No opt-out. Just a forced firmware update that turned a seamless, premium experience into a bureaucratic nightmare that requires me to babysit a smartphone app just to switch between devices. I paid premium money for the privilege of pressing one button. That was the promise. That was the product. And you had the audacity to silently delete that experience from hardware sitting in my hands, hardware I OWN. Here's the thing about ownership that apparently needs to be spelled out: when I buy your product, the transaction is complete. You don't get to keep editing what I purchased. An over-the-air update that removes functionality isn't a feature, it's a violation. At best it's a suggestion. You do not get to treat my hardware like a subscription you can renegotiate at will. Roll it back. Make it optional. youtube.com/watch?v=CWPrjGVN…

Experimenting training a FLUX.2 klein 4B pixel space version. I am just doing patch 16 directly from the pixel space into the model. I trained just the in/out layers for 24 hrs. I just unfroze the img stream of the first block and the linear layers of the last. Converging fast.