Recently, I heard about a project called AnyNode. It claims to achieve any function using a single general-purpose node based on LLMs. How is it implemented? Looking at the code, the author uses predefined prompts to write our desired "new node" functionality as a complete requirement, sends it to the large language model to generate the corresponding Python code, and then executes it as the node's default function. It's quite interesting.
After installing and trying it out, I think this plugin still has several issues that need to be addressed in practical use:
1. It cannot provide direct feedback. For example, when I expected to output only the Red Channel, the resulting image was just a line, clearly incorrect. Since ComfyUI didn't report an error, the large model wouldn't update the code, missing the opportunity to gradually achieve a correct method.
2. It cannot save nodes. Even if a perfectly functional node is created, it cannot be reused the next time ComfyUI is opened.
3. The prompts are too difficult to write, requiring a certain level of coding knowledge to achieve the desired outcome.
4. The quality of the nodes is limited by the intelligence of the large model. There is a significant difference between accessing the highly intelligent cloud-based GPT-4 and the local Phi3-Mini.
As for the limited number of supported cloud-based models, that's purely a matter of workload, and I believe it's just a matter of time. However, addressing the earlier points, especially the first one, may require changes to the node's interaction logic.
Overall, this "universal plugin" is worth trying out but isn't very practical yet. Kudos to the author for the idea, and I look forward to future updates.
Plugin URL: [github.com/lks...](github.com/lks...)
22 авг 2024