stop doing garbages!
let's look at the base model pipeline itself. It seems the dev team completely forgot how to build a clean, unified diffusion model. Instead, you are just stacking monstrous multi-layered layers on top of each other.Why do we need a separate 19GB text parser model, plus a massive 17.5GB text encoder trunk inside the base model, plus the 7B DiT engine? Why is it impossible to bake the text comprehension directly into the diffusion training process or use a modern, compressed text encoder?You are forcing the local open-source community to run a cascading chain of three massive, redundant models just to get a single image. This is not efficient AI architecture; this is lazy 'monster-stacking' wrapped in пафос (hype) marketing."
dude, shut the frek up? who are you? why dont you go and create something of your own instead of being angry about something some one else is giving you for free. how pathetic can you be?
this post is true garbage
Dude, show them how to do it, what is holding you back to make better model? /s
i was thinking the same before but training a model to deeply understand language from scratch alongside image gen is computationally exorbitant and even if you do it, it will result in poorer prompt adherence
It works beautifully, but the only 'problem' is that the text encoder sometimes struggles to fully grasp the prompt's intent, requiring a more extended or detailed prompt to produce the desired image. At Q8 for the DiT model and Q6 for the text encoder, it stays well bellow
12 GB of VRAM.
agree, 垃圾玩具，生图模型AI感太严重，肢体错乱，一言难尽，不如krea2