ali-vilab
TeaCache support for Chroma model.
As the title suggest, would love to see support for the Chroma model as it seems to need something slightly different than standard FLUX teacache to work. If it's even possible.
https://huggingface.co/lodestones/Chroma
Thank you!
EDIT: I'd look at the PR stuff you guys link here and there, but honestly I wouldn't even know where to start even with all that info. I'm don't know coding and stuff.
You can follow the instructions in the README to support new models. Welcome to launch a PR if you support new models.
So, take whatever I've discovered here with a huge grain of salt. I am not a coder nor a programmer. All of this is just trial and error and a little help of Claude 4. I just would like for someone to look into adding support for this model so I'm just experimenting with my very limited coding and programming knowledge.
But, I managed to get TeaCache to load with ComfyUI-TeaCache, tho we had to counter for some attributes missing like "time_in" and "vec".
But once we made it work, at first it seemed like coefficients were completely wrong, but it seems like there is something completely different going on because no matter what we tried, even "[1.0, 0.0, 0.0, 0.0, 0.0]" there was a hard threshold between 0.264 and 0.265 that just wouldn't move or budge. On 0.265 or above, TeaCache would cache, but extremely aggressively. On 0.264 or below, it would be basically disabled, no apparent caching effects.
I have no clue if that tells anyone something, but yea. There you go.
所以，对我在这里发现的任何东西都持保留态度。我_不是_程序员，也不是程序员。所有这些都只是反复试验和 Claude 4 的一点帮助。我只是想找人考虑添加对这个模型的支持，所以我只是在试验我非常有限的编码和编程知识。
但是，我设法让 TeaCache 使用 ComfyUI-TeaCache 加载，因此我们不得不抵消一些缺失的属性，例如“time_in”和“vec”。
但是一旦我们让它起作用，起初似乎系数是完全错误的，但似乎发生了完全不同的事情，因为无论我们尝试什么，即使是“[1.0， 0.0， 0.0， 0.0， 0.0]”，在 0.264 和 0.265 之间也有一个_硬_阈值，它不会移动或移动。在 0.265 或更高版本上，TeaCache 会缓存，但非常激进。在 0.264 或更低版本下，它基本上会被禁用，没有明显的缓存效果。
我不知道这是否告诉了任何人什么，但确实如此。给你。
Try MagCache
