ttul

40 ms·

541.
by ttul 3y ago
Convolution filters attend to a region around each pixel; not to every other pixel (or patch in the case of DiT). In that way, they are not good at establishing long range dependencies. The U-Net in Stable Diffusion does add self-attention
542.
by ttul 3y ago
This implies that their production environment is mutable. As in, a command can run and change the production environment. That’s a no no. But I give them a pass because they are a young company. My company was similarly reckless early on,
543.
by ttul 3y ago
I’m the wrong person to ask, but it seems Stability intends to offer models from 800M to 8B parameters in size, which offers something for everyone.
544.
by ttul 3y ago
Oh they’ll never let you pay for porn generation. But they will happily entertain having you pay for quality commercial images that are basically a replacement for the entire graphic design industry.
545.
by ttul 3y ago
I suspect Stable Cascade may incorporate a DiT at some point. The UNet is easily swapped out. SC’s main innovation is the training of a semantic compressor model and a VQGAN that translates the latent output from the diffusion model back to
546.
by ttul 3y ago
If you renoise the output of the first diffusion stage to halfway and then denoise forward again, you can eliminate the bad output. This approach is called “replay” or “iterative mixing” and there are a few open source nodes for ComfyUI you
547.
by ttul 3y ago
I suppose you could train a model from scratch if you have enough money to blow…
548.
by ttul 3y ago
Not too out of date. Just replace the magic UNet with a DiT and squint. It’s doing the same thing - removing noise.
549.
by ttul 3y ago
It’s the transformer making the difference. Original stable diffusion uses convolutions, which are bad at capturing long range spatial dependencies. The diffusion transformer chops the image into patches, mixes them with a positional embedd
550.
by ttul 3y ago
Nvidia will not build that any time soon. RAM is the dividing line between charging $40,000 vs $2500…
551.
by ttul 3y ago
MPS is promising and the memory bandwidth is definitely there, but stable diffusion performance on Apple Silicon remains terribly poor compared with consumer Nvidia cards (in my humble opinion). Perhaps this is partly because so many bits o
552.
by ttul 3y ago
Stability has to make money somehow. By releasing an 8B parameter model, they’re encouraging people to use their paid API for inference. It’s not a terrible business decision. And hobbyists can play with the smaller models, which with some
553.
by ttul 3y ago
Diffusuon models will be as much a killer app as LLMs. A picture is worth 1,000 words. A video…
554.
by ttul 3y ago
Have you experimented with running diffusion models on Groq hardware?
555.
by ttul 3y ago
Who the hell are they? /s
556.
by ttul 3y ago
I’ll often ask GPT to explain a concept to me, but giving it a hint as to my existing level of knowledge in the subject area. With Google, you’re likely to get the most popular page on a topic; not the page that is appropriate for people w
557.
by ttul 3y ago
Surely within days. ComfyUI’s maintainer said he is readying the node for release perhaps by this weekend. The Stable Cascade model is otherwise known as Würschten v3 and has been floating around the open source generative image space since
558.
by ttul 3y ago
I find it interesting that they are releasing the code and lovely instructions for training, but no model. They are almost begging anonymous folks to hook the data loader up to an Apple Music account and go nuts. Not that I am suggesting an
559.
by ttul 3y ago
Oh, I'm just making an analogy; of course you would never use the EHT to image something on the moon.
560.
by ttul 3y ago
SDXL is a more sophisticated model architecture. It has more layers. The CLIP model is bigger.
561.
by ttul 3y ago
Particularly considering the rich world of SD1.5 fine-tunes, SDXL leaves so much to be desired. I'm sure it will all be sorted out eventually, but right now, the momentum in the community just isn't there with SDXL the way it is w
562.
by ttul 3y ago
ComfyUI is insanely amazing. The learning curve is well worth the effort.
563.
by ttul 3y ago
SDXL actually takes conditioning from either the old or the new CLIP, or both. The malleability of SDXL is not just down to the choice of the new CLIP; the UNet itself is more opinionated.
564.
by ttul 3y ago
I’ve been playing with diffusion a ton for the past few months, writing a new sampler that implements an iterative blending technique described in a recent paper. The latent space is rich in semantic information, so it can be a great place
565.
by ttul 3y ago
If my math is right, the angular resolution of the EHT is such that you could image an object just 99m across on the surface of the moon. Or a 2in object could be resolved from a low earth orbit satellite. Not bad. They are imaging an objec
566.
by ttul 3y ago
Here in British Columbia, the car insurance system recently received a massive overhaul. Car insurance has long been a public monopoly that many complained led to higher prices than in other provinces. That much may be true. But the recent
567.
by ttul 3y ago
If you want to employ the top ML researchers, you have to give them what they want, which is often the ability to share their discoveries with the world. Making Llama-N open may not be Zuckerberg‘s preference. It’s possible the researchers
568.
by ttul 3y ago
Yeah, but if your company wants to rent an H100, you can deploy this for your developers for much less than the cost of a developer…
569.
by ttul 3y ago
A 70B model is quite accessible; just rent a data center GPU hourly. There are easy deployment services that are getting better all the time. Smaller models can be derived from the big ones to run on a MacBook running Apple Silicon. While t
570.
by ttul 3y ago
This is the kind of basement sorcery that makes the world go round. Nicely done.

More ›