Check my exclusive models on Mage: ParagonXL / NovaXL / NovaXL Lightning / NovaXL V2 / NovaXL Pony / NovaXL Pony Lightning / RealDreamXL / RealDreamXL Lightning
You can support me directly on Boosty
RealFlux Hugging Face Full Collection
Using the Model
Compact Version The Compact version already includes VAE, Clip-L, and T5XXL.
For ComfyUI, place the model in the following path:
ComfyUI_windows_portable\ComfyUI\models\checkpointsFor Forge, place the model in the following path:
webui_forge_cu121_torch231\webui\models\Stable-diffusion
Transformer Version (Unet) The Transformer version requires you to have Flux VAE, Clip-L, and T5XXL.
For ComfyUI:
Place the VAE in the following path:
ComfyUI_windows_portable\ComfyUI\models\vaePlace Clip-L and T5XXL in the following path:
ComfyUI_windows_portable\ComfyUI\models\clipPlace the Transformer (Unet) in the following path:
ComfyUI_windows_portable\ComfyUI\models\unet
Then you can use the Basic Workflow by moving the image into the ComfyUI workspace.
Dev Optimal Generation Settings
Sampling Method and Schedule Type: Euler Beta
Sampling Steps: 25+
Distilled CFG Scale: 3.5+
CFG Scale: 1.0
Schnell Optimal Generation Settings
Sampling Method and Schedule Type: Euler Beta
Sampling Steps: 4-6
Distilled CFG Scale: 3.5
CFG Scale: 1.0
Limitations of RealFlux 1.0b Dev
This version of the model does not have a variety of styles and is intended for realism.
Currently, there are issues with anatomy when generating NSFW content.
There may be issues with anatomy and text.
Limitations of RealFlux 1.0b Schnell
Cannot generate NSFW content.
There may be issues with anatomy, as the model is based on Flux Schnell.
There may be issues with generating images in the anime style.
Description
FAQ
Comments (45)
The Flux version of RealVisXL is highly appreciated! Can't wait to try it.
Will you also have a version based on [dev]?
Hello! Thank you!
This is currently the first version of the model, and it does have some limitations (The Dev version will not have such limitations), but I will work very hard on the 1.0b2 version to ensure the model has many styles, the ability to generate NSFW content, and improved anatomy.
The Dev version will also be released after successful testing.
@SG_161222 It's good to hear a dev version is on the way. I think I'll wait for that as your checkpoints are my goto ones and always worth it.
Just adding this here, you can follow the progress on the author's HuggingFace page: https://huggingface.co/collections/SG161222/realflux-flux-66d7ee8a0429a62c38d21df1
Hey there,
I believe the default comfy path was changed a while ago
from :
ComfyUI_windows_portable\ComfyUI\models\unet
to :
ComfyUI_windows_portable\ComfyUI\models\diffusion_models
btw, amazing to see you make Flux models! Thank you!
Hello! Thank you!
Regarding the location path: my model (Transformer Version) is specifically located in the unet folder. I just checked, and the model works equally well in both the unet and diffusion_models folders.
@SG_161222 If it still works with the unet folder, great!
They anounced to use the diffusion_models folder going forward a while ago and I only read it in the small "News" part in the top right of the comfy manager. No idea what the reason behind that was, but I thought I mention it, just in case 👌
I was waiting for that !!!. Can you also make the NF4 Version of it ? FP16 is really slow. It could be better when u compare the base model with your RealFlux side by side at the same seed.
Hi! Thank you.
I will try to make the NF4 version of the model soon.
Soon, I will make a small comparison between the base model and RealFlux Schnell, and I will share the comparison post in the comments on your post.
wow this look great! was it trained with Kohya?
Realism is only achieved on a full-fledged Dev. We will wait and after release we will test it! Thanks, good luck!
The images you create are absolutely outstanding but I have to (slightly) disagree on your realism point. I actually don't think either Dev or Schnell do realism all that well. Dev is definitely better but I think it still requires a further upscale/refinement operation in order to be truly photorealistic and I've found Schnell to be just as capable at photorealism if you include that extra operation. I also think Schnell models do a better job than Dev at color saturation, image composition and facial expressions. All my opinion of course and this is using ComfyUI locally. If generating online then Dev is definitely the better option.
@Grumblebutt The best way is to use it and use ultimate detailer... then LUTS... you need at least two... one for bg and one for foreground... and I also use SDXL loras for lighting... perfect photo copy...
@Grumblebutt I agree with you, I meant compared to shell, dev is better in quality. At the moment there is no model where very high realism, it can be achieved only with the help of different Lora such as for example iPhonePhoto.
@zGenMedia What is LUTS? I'm still a bit of a noob and not familiar that that term.
@Grumblebutt https://www.studiobinder.com/blog/what-is-lut/
@roberto_baggio Ahh, thanks!
@zGenMedia Do you have the workflow link that you use?
Hello everyone and thank you for your feedback and images!
The Dev version of the model will be completed by the end of the week. This model surpasses the Schnell version in terms of detail and anatomy, supports basic NSFW, and has improved text generation. The model does not yet support a wide range of styles and concepts, but all of that will be included in future versions.
Thank you all again and have a great day!
Awesome, can't wait to try it. Does the Dev version use the same training dataset as Schnell? If I have one criticism of the Schnell version, it's that I tend to see the same or very similar faces even after changing seeds and/or the prompt. More facial diversity would be good.
I'm surprised that people are training on the schnell version. It seems that it always get neglected in attention for the dev version which has a more restrictive license but ironically have more resources devoted towards it. Thanks for releasing the model.
Did the dev version move somewhere? The progress page on HF isn't there any more.
@testuser1111 I encountered a problem with uploading the model to HF (my internet connection has been unstable lately). Once I resolve it, the model will be available on HF again. I am also updating the model page. I hope there won't be any problems with CivitAI, although last time it took me 5 hours to upload the model there.
Could you remove the plastic skin feeling of flux base model?
@jeriff The best way to make more realistic skin is to upscale and refine with a SD 1.5 or SDXL realistic model. If you use ComfyUI you can look at my realistic images and copy the workflow.
Just got word of this LOL. let me download.
Great model, our we getting .GGUF versions?
I need to figure out how to do the .GGUF version myself! Ha I have 16 vram and for some damn reason I still struggle w/ these flux models. hate/love w/ flux ha.
@SG_161222 YASS another nice thing I might do is use the 4 k m one to be my lil guy I use for Open WebUI, so when I use my chatbot I have there I can generate images from there also really fast. DANKA
Wow thank you so much! Whats the difference between the Transformer and the Compact Version?
Thank you! This is mentioned in the model description. Otherwise, they are absolutely identical models in terms of image generation.
please think about mixing the two "1.d and schnell" as unchained hybrid model. It works in lower steps and supports everything 1.d does.
I think there is no pointing in doing this. If he is training shnell is probably because of license. If not, then he should focus on dev as it would be the highest quality. People would make hybrid if they want based on that.
hes training schnell for a reason. he wants the commercial license. personally schnell sucks so fucking bad im not even going to bother downloading this
Dev cannot be used for commercial stuff, Schnell can.
But by all means make your own merge and post it! More is better!
@Thatrandomjew I understand all of you guys' answers, if it was as simple as adding a lora in comfy I would do that , don't have the gpu vram or power or knowledge to do this hybrid stuff .
Well done SG, looking forward to give it a try!
Thank you! :)
it makes the women look asian even in your "european woman" prompts
Hello, did something happened to the dev version? Its placeholder page is gone from huggingface
He has problems with the internet, he can't upload the model. He wrote about it.
@daroth @minimishka The problem will be resolved soon and the model will be uploaded to HF and CivitAI.