Anonymous
6/21/2025, 4:16:46 AM
No.105657573
>>105657513
Explain with a straight face how all of Flux Dev, Lumina 2, Cogview 4, HiDream, and now Cosmos 2 are actually even very different from each other at all in terms of output. Then address why HiDream was for some reason treated like the second coming in comparison to the rest of the post-Flux ones, despite it also being the only one with a very significant and noticeable dropoff in prompt adherence after exactly 128 tokens.
SD 3.5 Medium also had its strengths (good support or up to 2 megapixel gens, a distinctly non-plastic photographic dataset, etc) but despite that also being receieved well in comprison to the original SD 3.0 everyone just sort of randomly stopped talking about it.
TLDR no one wants to do fucking anything to make anything into anything better than it is, that's the core of the problem probably. It's not a lack of viable options to use as baselines.
Explain with a straight face how all of Flux Dev, Lumina 2, Cogview 4, HiDream, and now Cosmos 2 are actually even very different from each other at all in terms of output. Then address why HiDream was for some reason treated like the second coming in comparison to the rest of the post-Flux ones, despite it also being the only one with a very significant and noticeable dropoff in prompt adherence after exactly 128 tokens.
SD 3.5 Medium also had its strengths (good support or up to 2 megapixel gens, a distinctly non-plastic photographic dataset, etc) but despite that also being receieved well in comprison to the original SD 3.0 everyone just sort of randomly stopped talking about it.
TLDR no one wants to do fucking anything to make anything into anything better than it is, that's the core of the problem probably. It's not a lack of viable options to use as baselines.