I have found the opposite of this myself. Prompting things like 35mm, Imax and what not seem to help greatly (ive been appending terms like this to the end of my normal prompts

I agree that not only does it have no issues, but it's also much better than SDXL. However, you must forget all the "techniques" and "tricks" you used when creating prompts for SDXL. That doesn't work here.

Describe your scene in natural language, including what you want it to have, where, and how. Don't fall into the temptation of using old tricks

@colinw2292823 @javiergcim  if you have any samples of what you think is superior photorealism please provide some examples

It is S, (Schnell) means fast... Relatively speaking...

This version is meant for those of us with lower vram. I have only 12Gb vram, so I downloaded this version (from huggingface) and it works just fine. It takes around 25s to generate 4step image with my rtx 3060.

Is this file also going into the Unet folder like the original Schnell?

@olternaut That's what I'm testing now, cause a lot of workflows are using the checkpoint loader but I only have Unet versions of FLUX.

I have 24GB VRAM but this is faster and gives the exact same results as the official 22GB one. So... I see no reason not to use this FP8 version instead ^^

the same mistake. Just updated Comfyui

ahhh its working now.. it seems i dont have the correct path, because i share my models with A111, thanks.

@mindTurbulence What is the correct folder to install this file? Isn't it the Unet folder like the original Schnell?

@olternaut put it on unet folder.. make sure load the checkpoint on unet Loader node, not regular checkpoint loader

Do you have a link for that? I need to install these separately. I keep getting the error message for "assertion error CLIP state dist!"

@AntUnderboot https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

It's FP8. The original version is FP32. Basically it requires less VRAM but is a little less accurate.

@netsu Thank you i actually figured it out later, just one quick question which one use less VRAM this FP8 version or the one FP8 version released by the ComfyUI team, which also can use the checkpoint loader node?

@netsu The FP8 model released by ComfyUI is 17GB in size, while the FP8 model released by Kijai is only 11GB. Just curious, why there is a big difference in size for the same FP8 model.

@LHKJACKY The 17GB version includes the t5xxl_fp8 text encoder and the 11GB version is without it (so you need to download it separately).

The advantage of the 11GB version is that you can then pair it with a different version of T5 if you have enough system RAM to run the ft16 version. It also save you disk space and download time so that you don't have to repeatedly download the T5 that is embedded in the larger version.

The advantage of the single file version is that you can use the standard simple loader in ComfyUI.

sucks for you

Sounds like a skill issue ngl.

I seem to be able to generate nude women just fine. Maybe it's SFW if you work in a strip club?

As long as Flux will be for comfy UI and not A111, I won't use it.

Be patient ;)

@ProvenFlawless ComfyUI is infinitely superior, even when you don't know how to code.

@ProvenFlawless  Works in Forge.

@Ion6 I am fine, thanks. So many other lovely checkpoints....

A brief scroll through the image collection on Civitai proves that FLUX is very far from SFW. With creators making more loras and checkpoints by the day, there are many NSFW abilities. It is certainly less SFW than any SD checkpoint.

@MagicalErotica Not so sure about that. If your task is simply to produce a great image , the ComfyUI "murderboard" seems like taking the long way to get to the same result as Auto1111 or forge. The only benefit I see is that you can use Flux on it, but that seems more like a defect for Flux more than a Comfy UI benefit.

No. For a bunch of reasons. But the main one being that Flux does not work with Automatic1111. This may change in the future. But not right now.

i use this with loras constantly. it works fine for me. what checkpoint do you use typically?

On my RTX2070 (8Gb) graphics card, it takes an average of 90 seconds to generate a single image.

Hello, I'm new as well, but the education on flux guide here: https://education.civitai.com/quickstart-guide-to-flux-1/ helped me understand and find the clip and text encoder files. You probably need the VAE (ae file) file as well if you haven't got that yet. I was also using the wrong text encoder until I re-read this, one of them requires 32 GB of ram which I do not have. I hope this helps you!

@DangerousCat thank you for being so specific and explaining where to get the VAE. Already dowloaded few and still not the right one

@varchyvas Okay, I followed the guide and downloaded the three needed files from the hugging face website. After putting them in stability matrix, which is what I'm using locally. Along with the flux main checkpoint, one version of which is on this page directly, you will need: 1) the VAE file (ae.safetensors) https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main which goes in the models > VAE folder. 2) the clip L file (clip_l.safetensors) and 3) either t5xxl_fp16.safetensors (if you have over 32G ram) OR t5xxl_fp8_e4m3fn.safetensors (if you have less) both at https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main and both of those go into models > CLIP. then at least for webui, you need to add all those files in the VAE/Text encoder field before generating. if you are using something else it would be part of the workflow (I don't really know anything about this though, I don't use this process)

go to this page on huggingface: https://huggingface.co/camenduru/FLUX.1-dev/tree/main and download the t5xxl fp8 text encoder not the fp16 one or you'll still suffer bad runtimes locally, and this is an fp8 flux checkpoint. Then download the ae.safetensors file, and the clip_l.safetensors file. All 3 are on the page. The ae.safetensors file should be placed in you VAE folder that's inside your models folder. The clip and t5xxl files should be place in your text encoder folder, which is inside your models folder as well. In the UI after you've selected your checkpoint, there's a VAE/text encoder selection block next to checkpoint selection, in the dropdown select all 3 files you downloaded