Support OmniGen2 in ComfyUI

[neverbiasu]

Hi, guys. I made a ComfyUI custom node for OmniGen2, here is the repo

[Luciennnnnnn]

Hi, thanks for you contribution, we will update our README to include your implementation.

[Amazon90]

Hi, thanks for you contribution, we will update our README to include your implementation.

Do you have any plans to provide native support for ComfyUI? The node provided by this person is only a pipeline implementation and doesn't support loading a quantized version of the “OmniGen2” model.

[Luciennnnnnn]

@Amazon90 Hi, thanks for your interest! Quantization isn’t supported yet for OmniGen2. Native support for ComfyUI is not a priority right now, but we’ll consider it after we release the training code, dataset, benchmarks, and other key components.

[able2608]

For quantized inference, bitsandbytes 4-bit and 8-bit inference seems to be quite straightforward. OmniGen (the original one) already got support, and judging from the pull request, it only requires passing the quantization config to the loader. For others (like gguf), it might require a lot more effort.

[Luciennnnnnn]

@neverbiasu readme updated, thanks for your effort!

[Greatz08]

It would be great if we get quantized model for omnigen2, so that low VRAM users can also try and enjoy better speed

[Luciennnnnnn]

@Greatz08 Thanks for the suggestion! We'll consider adding quantization in the future.

[neverbiasu]

@Luciennnnnnn I appreciate it! Really looking forward to seeing more of your amazing work! And I will keep iterate the node to achieve better experience.

[Greatz08]

@Luciennnnnnn Thank you very much for understanding our need for a quantized model, and i hope we can get it soon :-))

[Amazon90]

For quantized inference, bitsandbytes 4-bit and 8-bit inference seems to be quite straightforward. OmniGen (the original one) already got support, and judging from the pull request, it only requires passing the quantization config to the loader. For others (like gguf), it might require a lot more effort.

That’s right — the reason we need native support for ComfyUI is to enable loading GGUF models. With GGUF, even users with 8–12GB of VRAM can run the OmniGen2 model. Compared to 4-bit and 8-bit implementations using pipe, GGUF requires less VRAM, runs faster, and the downloadable model size is also smaller.

[Amazon90]

Comfy-Org/ComfyUI#8669