nyanko7

/

flux-dev-de-distill

like 159

Hello! First of all, thank you for sharing! It's a great experimental project. After I actually tested your model, I found that the quality of the image improved significantly, especially in terms of the detail of the images.
Since my GPU only 24GB, I can't complete the inference script of you provided, but I read your script carefully, and I found that it may be a reference to the code of the xlabs x-flux-comfyui project, and then refactored the inference script, so I also did the following experiment:

1. Referring to the x-flux-comfyui project, design a workflow based on ComfyUI, load your model, to try some tests, the workflow is as belows, I first used the standard node to run, and found that it takes about 50 steps, or higher steps, to get better, better results than the dev model, here due to the limitations of my graphics card, I used FP8 quantization, to load the model and CLIP. My friend MZ also helped convert the quantization model of FP8(https://huggingface.co/MinusZoneAI/flux-dev-de-distill-fp8). Of course, I also know that based on your inference code, this workflow shouldn't be able to render the performance of the model perfectly.
   my firend MZ's workflow:

my workflow:

2. So, I made a second attempt, mapping your inference code to the corresponding module of the "x-flux-comfyui" project, and modifying the: layers.py, model_init.py, sampling.py, .\xflux\src\flux\math.py, and .\xflux\src\flux\modules\layers.py etc. The corresponding module in the script of the project "x-flux-comfyui" code according to your inference code.
   Then use the modified Node to reconfigure the workflow, the result is different from before, but not very big, and it also requires a lot of steps, such as 50-80 steps, to get the output closer to the Flux pro model.
   new workflow as below, remove the PAG node:

Here, I have a few questions ask for you, and I look forward to your guidance and help:

1. Can your inference code use fp8 to load flux and T5-xxl models? In this case, my 24GB graphics card should be able to run inference scripts.
2. Is it possible to use standard workflow nodes directly? What parameters do I need to adjust to reduce the number of steps and get a better output? With a standard node, it can plug in some Lora and it works, but controlnet cannot work fine.
3. I attach the code I have modified the node, please help me review whether this method is correct, if I use the modified code to run the workflow, can I connect to the Lora and Controlnet nodes? My previous tests were not ideal, either the effect was not achieved, or the pictures were messy.

I can't upload the script package which I had modified, if you need, I can mail to you for check and modify. thanks a lot!