DemonFlux is my first FLUX model.
The checkpoint is based on FLUX 0.1 schnell and FLUX 0.1 DEV. It also got the clip_l from DEMONCORE (SDXL). I used FP8-versions to save space and RAM.
How to use it?
This model doesn't need tons of loaders. Its all inside the model (including Unet, Clips and VAE)
All you have to do is to download the model and update Comfy. Put the checkpoint in the checkpoint folder.
This model works from 2 steps.. I recommend 3 steps with the settings from the workflow.
Here is the basic workflow for it: https://civarchive.com/articles/6508
I also highly recommend the DemonFlux+Colossus 12 workflow: https://civarchive.com/articles/6668
Feel free to test other settings and samplers of course :-)
Credits:
FP8-checkpoints I used: https://huggingface.co/Kijai/flux-fp8 or https://civarchive.com/models/618486/flux1?modelVersionId=694756 from Kijai
Thanks to Black Forest Labs for this amazing base model https://huggingface.co/black-forest-labs (check out their base models)
License:
Make sure to use this model for science or personal use only!
https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
Description
Important: you don't have to download anything else than this checkpoint.. VAE and the clips are included. Use this workflow for it.. https://civitai.com/articles/6508
Have fun creating :-)
FAQ
Comments (42)
This can be awesome, any plans to upload it on tensor art as well?
I will upload it to tensor when they support it
@Afroman4peace it does already, You can use flux models like any other models
Is this an actual finetune using SimpleTuner or is it just Demoncore's CLIP merged with Flux's CLIP?
Because from my and other's testing the CLIP_L model doesn't have much of an impact when changed, as was shown with the Nepotism Flux upload, so I really don't see much point in downloading a 15 gig file for a CLIP model that doesn't change much of anything if that is the case.
Hi, I've wrote this in the description. It's a merge between the flux schnell and flux dev. And yes the impact of the clip_l is not very strong but there is certainly a difference. If you like do a comparison between the normal schnell checkpoint. The goal of this checkpoint here was more to create something for mid range computers that is easier to install. And no.. this is not just a clip model here.. it's everything you need to download.. It's unet + clips + VAE..
Don't worry I will create trained Flux models ;-) This is something I can't do at one weekend.. Stay tuned..
@Afroman4peace interesting. When you say it's a merge between Schnell and Dev, do you mean that the Unet weights were essentially added together and then divided, like your typical 0.5 simple merge?
If so, I'm curious what the effects of such a merge would be, and what the step counts now look like during generation.
is it possible to remove the NSFW filter?
I think that Flux isn't that censored like SD3. It's more a training thing. Correct me if I am wrong here..
@Afroman4peace the piece of cloth is always put over the crotch area i don't know if they labeled that area with cloth while training but removing that would be great
@Tuve I will look into it soon. not sure if I can train Flux with my 4070ti.. Well Flux is also very new to me ;-)
@Afroman4peace From what I heard we can't train Flux yet. That might be wrong. Also, I heard that since it is a 12B parameter model, we would need to rent the really good GPUs from something like run pod once we could train it, because even something like a RTX4090 won't likely be able to train it in a timely manner. With that said, I also have tested the cloth crotch thing, and see the same things. If you want rear shots, no problem, but crotches always seem to be covered except in weird cases where it just does plain skin with 0 detail like as if the material is arm skin or something. For giggles I tried super increasing the references referring to that area to see if I could get it to show and when it would show something other than plain cloth or skin... it was horrifying... lol
I have a 4090, though, so if it turns out training is something that can be done on it, then I would be able to, but I would have to learn to do such first. lol If you wanted to provide assistance, I could potentially aid by running the stuff from my PC to ensure you would not be limited from moving forward. I look forward to see what you could do with Flux. Your other models before the newest version of SD were awesome. Not saying your models on the new version are bad, I just have not messed with AI art for a while cause life got busy, and have not messed with it for a while. I think they trained their model on images larger than 1024x1024 because I have tested and gotten great results for images that were 512x2160, 1024x2048, and 1920x1080.
@SuoixDs Thanks for the offer. I think I can try it with runpod. Also I may have to invest into a 4090 some day. I am was waiting for the 5x series but it seems that the release is a bit delayed. May getting a used one. Let's see. Compared with SD3 Flux is much better in my opinion. SDXL isn't dead yet. I have a created a Colossus version lately that will even challenge Flux regarding graphics. Yesterday I combined it in a workflow for DemonFlux in the second stage. The results were amazing. I am gonna post some images in the gallery.
@SuoixDs The model has to contain the data you want to see, as in, they had to have trained it on porn for it to make porn. It is not physically capable of imagining what genitals look like if the dataset they used didn't contain tagged photos of that concept.
It's not like there's porn hiding in there and someone coded in "when there's a vagina put cloth over it," it's that the model doesn't know what is supposed to be there so it fills in the blanks.
So simple tuners says:
Flux.1 [dev, schnell]
A100-40G (LoRA, rank-16 or lower)
A100-80G (LoRA, up to rank-256)
3x A100-80G (Full tuning, DeepSpeed ZeRO 1)
1x A100-80G (Full tuning, DeepSpeed ZeRO 3)
Flux prefers being trained with multiple GPUs.
@Afroman4peace Sorry for a likely stupid question. I have been out of AI art for about the last 6 months and last time was using A1111. First time using ComfyUI. What do you mean you have your SDXL model in a combined workflow with DemonFlux in the second stage? Are you talking about generating stuff with your Colossus model then using Img2Img on it in Flux?
@Up_punchington Flux do know how genitals look like, so it must be something else. I've tried having it do male genitals and during the denoise phase you can clearly see male genitals, but it scrubs them away later and exchange it to something else.
@Birnir I have seen a few come through for women by accident when not asking for it, but when trying to force it, it gets replaced with stuff or horrific stuff. Lots of pubic hair came through by accident when I was not actually wanting that. But if asking for it, it doesn't happen or horror often happens. I think maybe they labeled garbage in place of what you ask for and use some other tags for those parts instead. I might be totally wrong, though.
In my testing references to the groin seem to indicate the space directly under the belly button. On lots of images trying to get the woman put her hand over her groin or cover her groin, I ended up with a pregnant woman holding her hand on her belly under her belly button. This is why I say that.
I'm using this model and I'm able to do full NSFW. You just gotta make sure to process text encoding by another model before passing it to Flux.
@Birnir What workflow are you using?
@Birnir can you share a workflow how that is done please?
is it possible to use it with Forge UI ?
Forge and A1111 don't support flux (yet)
@tvange365 Thanks for the reply. I was asking because I've seen updates on Forge mentioning Flux but nothing to confirm it.
@fabricemag I will let you know when it's available. Also if the combined format is working
Can you please give me the comfyui workflow for this model
https://civitai.com/articles/6508
It's the basic workflow. I will update it soon though. Already made a better one with a SDXL upscale.
I suggest trying different ComfyUI flows from the images below, too. While the model's associated flow generated an image, it was a noisy one of the red head. I downloaded @moxie1776 's image, dragged into ComfyUI, and disconnected the upscaling and it worked. Like 50 minutes later on my 8GB RTX 2070 super. Plus one blue-screen on the first attempt. But it... worked.
I've been testing it for a few days and it works great. I use it on an i7 64GB Ram laptop and an "NVIDIA GeForce RTX 3060 Laptop" 6GB VRAM and an M.2 disk drive. I've tested it in ComfyUI and in FORGE version: f2.0.1v1.10.1-previous-223-g86ee2d94. It takes about a minute to generate the images. Thanks for this model, it's great.
Thank you for your feedback. I am glad it works for you. I have to test it myself with forge :-)
I have just added a new workflow that combines DemonFlux with SDXL (Colossus 12C Experimental) https://civitai.com/articles/6668
Just a couple tests literally, I use it on an old Xeon with DDR3 32GB RAM and nVIDIA GeForce RTX 3060 12GB VRAM using ComfyUI and recommended workflow. I can feel a sublte flavour of SDXL, what is good, I like it. I didn't expect it be fast, but considering nice upscaler, it gives me really good experience. In the future I will try to use this model in Forge as well as will try to follow with more advanced workflows, including one with refiner. Thanks for great model.
Hi can anyone help I keep getting this error, I have the model in my checkpoint folder like all the rest
"Error occurred when executing CheckpointLoaderSimple: Error while deserializing header: InvalidHeaderDeserialization"
hi. Are you using the latest Comfy UI version? If it's the case I am going to look into it when I am back home .
@Afroman4peace thank you! I have updated Comfy as of 8/21/24 not sure what else I should update?
DemonFlux+Colossus 12, works amazing!! I have deleted few other checkpoints to download this one. This model really deserves it. Thank you for great job :)
This has been a lifesaver in Forge. I couldn't get Flux to work in Forge properly and then this rode in to save the day. Works great and has allowed me to use Flux. Thank You 👍😁
Hi, thank you very much. Currently I have two other Flux projects one is a dev version (Colossus Project Flux), the other one is very similar to DemonFlux.
@Afroman4peace Thanks, I'll give it a try 👍
This model is one of my favorites, works sometimes from one step and cfg= 0.9( if you use triple upscaler). First stage(often but not always) does not need to have second step because it is denoised later but it needs to have cfg >1. More steps in stages more plastic look but very detailed. So to have photorealistic results with this model I used 1-2 steps and always tried to find sweet spot.
Tested on 4090 laptop, 16G VRAM.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.



















