Midjourney web experience is now open to everyone

Post by Der_Einzige » Thu, Aug 22, 2024, 5:23 AM UTC

ethanpil wrote: post #103

smusamashah wrote: post #38

Ideogram 2.0 was also released today, and it's nerfed (anatomy is lot worse than 1.0 now) just like StableDiffusion versions after 1.5 which is very disappointing. Well good thing we have Flux out in the open now, both midjourney releasing web version or ideogram releasing there 2.0 on the same day after 2 weeks of flux won't redeem them as much. Flux Dev is amazing, check what SD community is doing with it on https:…

Why are they getting worse?

Post by ffsm8 » Thu, Aug 22, 2024, 5:44 AM UTC

stemlord wrote: post #98

ffsm8 wrote: post #22

Earlier quoted context omitted.

You honestly think giving full, no rate limited free access is viable in the LLM space, where each execution is actually pretty expensive? Because that's what you're asking for, as it'd be trivial to reset the rate without accounts.

Chatgpt does it

Post by mrweasel » Thu, Aug 22, 2024, 6:02 AM UTC

ffsm8 wrote: post #22

mrweasel wrote: post #14

Earlier quoted context omitted.

While I understand that they might need some account or token to stop abuse, just having my sign in is a big no for anything I just want to try out. After the whole social media trend more or less collapsed into a privacy invading data collection nightmare, I've been extremely reluctant to sign up for anything.

You honestly think giving full, no rate limited free access is viable in the LLM space, where each execution is actually pretty expensive? Because that's what you're asking for, as it'd be trivial to reset the rate without accounts.

Post by Loranubi » Thu, Aug 22, 2024, 6:15 AM UTC

Der_Einzige wrote: post #130

liuliu wrote: post #92

Earlier quoted context omitted.

Quality issue should be mainly due to using FP16 accmulators for GEMM in M1-M2, A14-A16 devices (it is not a problem for SD v1 / SDXL models due to smaller channel count). This is changed to FP32 accmulators for GEMM in these devices with 1.20240820.1 release. q8p should have comparable quality to non-quantized models (in Draw Things, it is called FLUX.1 [dev] (Exact)).

Claims that quantization doesn’t hurt models are made all the time but rely on the fact that almost all evaluations today of LLMs hardly scratch their surface. If we evaluated LLMs properly, even large quants would be detectably worse, and by a significant amount.

Post by bilekas » Thu, Aug 22, 2024, 7:07 AM UTC

chenglou wrote: post #129

Website contributor here! Answering some of the technical questions on Twitter: https://x.com/_chenglou/status/1826425720522080480