is based on scnell ?
it can do 3~4 step like schnell ?
is the step speed is heavy and slow like any flux model ?
It's based on Schnell, but it has been de-distilled, so it requires about 20 steps.
For inference speed, note that since it has been de-distilled, it re-introduces CFG (and hence real negative prompts), and so each step takes twice as long. That said, if you have a 30XX or 40XX series NVIDIA GPU with at least 16GB VRAM, then speed won't be an issue for you (soon). You can get it down to 4 seconds per 1024x1024 image with First Block Cache and SVDQuant on a 4090, or 8 seconds on a 3090. It's possible to double that speed if/when someone trains a hyper/dmd2/etc lora, similar to this one.
It'll probably take at least few weeks before the community gets FBCache/SVDQuant optimizations implemented for Chroma in a way that's easy to use, and then hopefully we get some hyper-type lora experiments soon after.
It will be good to see this working on SVDQuant. Will be helpful for low VRAM users.
It will be good to see this working on SVDQuant. Will be helpful for low VRAM users.
i don't even have a GPU .
is it possible ti run it on mac?