Home

Awesome

WORK IN PROGRESS

Installation

Note: Sampling is slow without flash_attn !

For Linux users this doesn't mean anything but pip install flash_attn.

However doing same on Windows currently will most likely fail if you do not have a build environment setup, and even if you do it can take an hour to build. Alternative for Windows can be pre-built wheels from here, has to match your python environment: https://github.com/bdashore3/flash-attention/releases

If flash_attn is not installed, attention code will fallback to torch SDP attention, which is at least twice as slow and memory hungry.

Text encoder setup

Lumina-next uses Google's Gemma-2b -LLM: https://huggingface.co/google/gemma-2b To download it you need to consent to their terms. This means having Hugginface account and requesting access (it's instant once you do it).

Either download it yourself to ComfyUI/models/LLM/gemma-2b (don't need the gguf -file) or let the node autodownload it.

Lumina models

The nodes support the Lumina-next text to image models:

https://huggingface.co/Alpha-VLLM/Lumina-Next-SFT

https://huggingface.co/Alpha-VLLM/Lumina-Next-T2I

They are automatically downloaded to ComfyUI/models/lumina

Examples

The workflows are including in the examples -folder image

lumina_composition_example

lumina_i2i_example

Original repo:

https://github.com/Alpha-VLLM/Lumina-T2X