lodestone-rock
@LodestoneRock
Followers
1K
Following
2K
Media
142
Statuses
516
Researching pixels @freepik. Independently created the Chroma & Radiance models as personal projects. https://t.co/d8h2mIi9zS https://t.co/RrPsPku3y6
In your heart
Joined April 2023
chroma-radiance-x0 result a true pixel space model combining JiT-x0 loss and pixnerd give it a go! https://t.co/SGyR1BFivX
https://t.co/5bo6TPfa0Y
10
57
423
crunchy converting z-image into pixel space rn using dino v3 as auxiliary loss makes pixel space convergence ridiculously fast. also it's 3-4x faster than z-image-base because it only use 1/4th of the token to generate 1MP images if you want to mess around with this WIP model
2
11
106
PSA keep your master weights in FP32 training on bf16 will eventually plateau tho this result is a bit inconclusive because i ditched one of the loss element to make room during training but yeah BF16 master weights only good during first descent.
2
1
29
What are you fucking MORONS on about you already have my face and I’ve talked about working!!!
40
6
93
anyone know discord alternatives? i already canceled my nitro because of this dishonesty
We’ve seen some questions about our age assurance update and we want to share more clarity. We know how important these changes are to our community. Here’s what we want you to know: ‣‣‣ 𝗗𝗶𝘀𝗰𝗼𝗿𝗱 𝗶𝘀 𝗻𝗼𝘁 𝗿𝗲𝗾𝘂𝗶𝗿𝗶𝗻𝗴 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘁𝗼 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗲 𝗮
6
1
15
I always felt CFG was a patch to fix a training problem we didn't yet understand. Training with only normal distributed noise teaches the model that each step will have a perfectly normalized error from the previous step, which is not the case. Therefore, it is incapable of
22
28
416
hmm isn't it obvious this would work better than QAT? I’ve been doing this already for HQQ+, the DeepSeek re-distill work, and FP4 weight quant quality recovery. More comments below 🧵
We just launched an ultra-efficient NVFP4 precision version of Nemotron 3 Nano that delivers up to 4x higher throughput on Blackwell B200. Using our new Quantization Aware Distillation method, the NVFP4 version achieves up to 99.4% accuracy of BF16. Nemotron 3 Nano NVFP4:
6
4
20
Pretty sure my model is subtly telling me to stop changing the architecture mid-training 😅 Chroma Radiance, patch-32 pixel-space model, ~3–4× faster than latent space. Still training atm so expect rough edges. Hourly checkpoints: https://t.co/MWILprAsuq Works in ComfyUI.
4
7
66
yes you can see the manifold hypothesis in action here, the model is modeling the lowest frequency band first then gradually construct higher frequency it's somehow similar to DCT where the lowest frequency is being the most important. https://t.co/8xeHTv349D
I think that JiT ( https://t.co/JBr5lx1odv) might have been my favorite paper of 2025. From the discussions with my friends, it got quite some controversy with many people dismissing it as some trivial reinvention of x-prediction, so I would like to put my perspective on it here
0
1
20
now the private storage quota is 1TB only! check your repo guys! this unannounced changes is bad @huggingface @ClementDelangue the f is going on ? you gonna silently charge people a lot of money without them noticing!
seems like @huggingface is reducing the grandfathered free storage gradually @ClementDelangue is this intentional?
8
4
36
@huggingface @ClementDelangue the private storage one, is gradually decreasing i guess all i need to do is to make it public / migrate it right?
0
0
3
seems like @huggingface is reducing the grandfathered free storage gradually @ClementDelangue is this intentional?
3
0
24
math bug is the worst bug ever! check your math before doing calibration bilinear interpolation collapses noise variance (1.0 -> ~0.25). Use NEAREST neighbor instead! 2nd pic is nearest 3rd is bilinear it's much more in distribution now even without calibration training
Experimental chroma-radiance pixel-space model calibration results. This 1024×1024 image only needs 1024 tokens! not 4096 tokens like latent space models. 3-4× speedup! latent space is officially dead! And this is with just 50 param updates in so far.
1
2
49
the detail still crunchy but proof of concept works just need more time in the GPU oven this is 300 training steps in Prompt executed in 6.03 seconds using comfyui 30 euler steps on h100
Experimental chroma-radiance pixel-space model calibration results. This 1024×1024 image only needs 1024 tokens! not 4096 tokens like latent space models. 3-4× speedup! latent space is officially dead! And this is with just 50 param updates in so far.
3
0
47
i think the LR is a little bit too high for this calibration run. things are slightly wonky but it's learning so gonna let it run overnight to see if it stabilize.
0
0
8
Experimental chroma-radiance pixel-space model calibration results. This 1024×1024 image only needs 1024 tokens! not 4096 tokens like latent space models. 3-4× speedup! latent space is officially dead! And this is with just 50 param updates in so far.
4
13
76
toy data visualization of this feedback algorithm in action https://t.co/KvT9ksU4TH
GANs might be making a comeback. I’m treating the generator discriminator pair as a dynamical system and stabilizing it with a control theory feedback loop. No R3 regularization, no double backward. Transformer based, bijective mapping to noise latents. https://t.co/QIyfiPdXkj
2
8
129
GANs might be making a comeback. I’m treating the generator discriminator pair as a dynamical system and stabilizing it with a control theory feedback loop. No R3 regularization, no double backward. Transformer based, bijective mapping to noise latents. https://t.co/QIyfiPdXkj
16
17
247
experimenting with self stabilizing feedback loop GAN training. more details soon no need double backward (R3GAN) just pure relativistic GAN with feedback loop to control the instability.
1
0
28