«

Running FLUX.2-dev on a Strix Halo

{,,}
 December 3rd, 2025

FLUX.2-dev is the newest text-to-image generator introduced by the German AI company Black Forest Labs (BFL). FLUX.2.dev is the open source variant and can be downloaded and use local on your PC. It is a huge model with 32 billion parameter, 64GB in size. So you need a beefy computer with a lot of memory

https://huggingface.co/black-forest-labs/FLUX.2-dev

Luckily the Strix Halos has up to 128GB of unified memory, so it’s ideal for this huge AI models.

Easiest way to try out image, movie and audio generation is with ComfyUI.

There are several ways of installing ComfyUI, or using a toolbox, but here we follow the instructions on the ComfyUI Github repository.

Installing ComfyUI on a Strix Halo

create a virtual environment to isolate your Python install.

python3 -m venv .comfyui
source .comfyui/bin/activate
git clone https://github.com/comfyanonymous/ComfyUI.git

cd ComfyUI

pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx1151/

pip install -r requirements.txt

Running ComfyUI

In short:

TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention

Workflow for FLUX.2-dev on ComfyUI

Here is a collection of examples how to run a specific model.

https://comfyanonymous.github.io/ComfyUI_examples/

Of course sometimes you can save some VRAM by using a quantized model

https://huggingface.co/city96/FLUX.2-dev-gguf

https://huggingface.co/city96/FLUX.2-dev-gguf/blob/main/flux2-dev-Q4_K_M.gguf

But for the best quality you wanna stick to the bf16 or fp8 models. We just downloaded all the files from the official example;

https://comfyanonymous.github.io/ComfyUI_examples/flux2/

Running FLUX.2-dev with ComfyUI on Strix Halo

The best way to start ComfyUI on a Strix Halo is by adding some extra parameters for better performance:

TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py  --use-pytorch-cross-attention --disable-mmap

Don’t forget to add –disable-mmap, otherwise ComfyUI will hang.

It will take about 6 to 7 minutes to generate a 20 step 1200×600 picture. It’s not that fast, but the image quality is excellent.

 

Tags: , ,

This entry was posted on Wednesday, December 3rd, 2025 at 4:02 pm and is filed under Webtechnology. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.

Leave a Reply