How to convert FLUX.1-Depth/Canny/Fill-dev.safetensors to Q8?

[ymzlygw]

Hi, blackforest just released their strong controlnet model. But It is too big sample as flux.1 dec fp16 (24GB).
could you please how to convert the FLUX.1-Depth/Canny/Fill-dev.safetensors to Q8 and can you support it ? Thanks!

[ymzlygw]

@city96 Could you help me about this ? thanks!

[jjiikkkk]

search youtube video, there is a colab to do so. But I do not know how to do that.

[ymzlygw]

search youtube video, there is a colab to do so. But I do not know how to do that.

can you share the video url ? I don't know search what key words

[whitespace-rebel]

There are conversions available on huggingface. Here for example is quantized FLUX.1-Fill-dev:
https://huggingface.co/YarvixPA/FLUX.1-Fill-dev-gguf/tree/main

FLUX.1-Fill-dev.gguf itself is supported by the ComfyUI-GGUF nodes

The problem I have is that FLUX.1 Dev LoRAs are not working with FLUX.1-Fill-dev.gguf. This is the error I get:
ERROR lora diffusion_model.img_in.weight shape '[3072, 384]' is invalid for input of size 196608

Are the LoRAs even supposed to work with FLUX.1-Fill-dev ?