Comfy WaveSpeed on Windows
Comfy WaveSpeed
Guidance
This is the program that has made the biggest impact on me in recent days: "Comfy-WaveSpeed".
The speed improvement ratio is comparable to the former TensorRT, and it supports ControlNet and LoRA, which were impossible with TensorRT.
I haven't tested it for video, but it seems to support both HunyuanVideo and LTXV.
The author, chengzeyi, is the same person who created "Stable-Fast," the legendary acceleration node that disappeared without ever seeing the light of day. He is someone who developed it with the goal of "defeating TensorRT."
As for "Stable-Fast"...as far as I know, I am the only one in Japan who properly and correctly explained how to install this technology and released it for free. (The article I uploaded to a certain place has already been deleted.)
It was highly difficult... that thing... I managed to install it by crushing the error messages one by one... but in conclusion, it was a technology with many problems for practical use. I thought the theory itself was amazing, though.
By the way, the experience of investigating the cause of every error and completing the installation of Stable-Fast on my own through trial and error, after being on the verge of giving up several times... is the origin of my current confidence in ComfyUI... or rather, the reason I have come to believe thateven if you are not a professionally trained engineer, if you work on it seriously, you can manage to a certain extent.
On the other hand, Nvidia's highly anticipated TensorRT was usable to some extent with SD1.5, but in the end, it could not support ControlNet or LoRA with SDXL, was completely ignored by the Forge series, and I honestly thought that acceleration in this form was impossible, so I forgot about it for a while, but
"Comfy-WaveSpeed" has appeared with amazing perfection and versatility. It perfectly supports ControlNet and LoRA, and also speeds up various Detailers. I offer my heartfelt respect and gratitude to chengzeyi for his great achievement.
People say it's 3x or 4x faster... and in Flux1 and SDXL, where I created the workflows, it did indeed achieve that level of speedup overall, including the Detailer. In a word: Amazing!
About torch.compile
There is a rumor circulating in some places that it does not support Windows, but that is not the case at all.
In fact, I have successfully completed the Compile Model+ node in Flux1.Also, on February 25, 2025, I succeeded in loading the settings that maximize the functionality of torch.compile.
With that premise, please refer to the following for how to operate the torch.compile function part.
Dealing with new errors
Please refer to the article below and take action if applicable.
chengzeyi
Looking at chengzeyi's profile... he seems to be quite an impressive person. He is an engineer at the Chinese AI development group SiliconFlow... seeing this, I can see that he is indeed someone capable of achieving his goal of "defeating TensorRT."
Technical tips
The information on the effects for each sampler that shiba*2 has compiled into an article is extremely useful. It is true that euler a has a low effect.
One might tend to think that all Euler-based samplers are the same, but there are indeed differences. When using Euler-based samplers, I tend to use 'a' rather than the plain version, so I found that interesting.
Also, since they have described the reasons why differences in effects occur, this information is also very valuable.
Furthermore, since Kimama / Easygoing has verified the settings for the Dynamic Caching part in considerable detail over two parts, the following articles are also very valuable. I would like to start using the recommended settings calculated by ChatGPT right away.
Fixed jsons
I have not applied WaveSpeed to all the jsons I have created, but basically, for SDXL and above, I am applying it including derivative versions.
How to install nodes of WaveSpeed
I am continuously updating the WaveSpeed-added versions of the workflows I have created so far, but here I will explain how to add WaveSpeed nodes to existing jsons.
The following is written on the premise that the preparation for installing triton and introducing torch.compile, as well as the installation of global_config.py and the addition of environment variables, which are explained in the article below, have been completed.
On SDXL
The figure below is a very common workflow for SDXL (with LoRA), and I will use this workflow as an example to explain.
First, double-click in an empty space as shown in the figure below, enter "apply first" in the search word, and select the Apply First Block Cache node.
Next, as shown in the figure below, insert the node between Load LoRA and Ksampler and wire it, and set rdt to 0.2(*The figure below shows 0.12, but this is incorrect; the official recommendation for SDXL is 0.2). This is the official recommended value for SDXL. For further optimization, please try trial and error by referring to the articles by shiba*2 and Kimama / Easygoing mentioned above.
The node insertion position also works immediately after the Base Model Loader as shown in the figure below. Honestly, I don't know which is better.
On Flux1
In the case of Flux1, the basics are the same, but change the rd of the Apply First Block Cache node to 0.12 as shown in the figure below, and add a torch.compile node by KJ Nodes immediately after it. Then, select max-autotune-no-cudagraphs in the Mode field as shown in the figure below.
Note that even in a state where only the Apply First Block Cache node is added without the torch compile node, a certain amount of speedup will occur.