>>106844219
lodestone might have figured out a good "ramtorch" way to train and/or run inference on models with more system RAM than vram
pony v7 auraflow version is disappointing (as many expected) and pony team decides to train qwen
wan 2.5 is api only [for now? some people were pessimistic about that]
qwen image edit got an updated version, it is good. as are the updated lightning lora to speed it up
huyuanimage3.0 won some benchmarks but the aesthetics suck and it's too big for now