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
This version of the model is unstable.
FAQ
Comments (83)
The 1.0 Dev version is currently unstable. There are issues with NSFW generation and anatomical problems in some cases.
I may have found a new approach for training the next version of the model, which should help fix many of the shortcomings of the first beta version. I will start working on it today.
I am also designing a model like you and I have the same problem. But in my model such words as “perfect hands, feet” fixes the situation. There is another option Lora Hands XL + SD 1.5 + FLUX.1-dev works well with it.
@alexoterin Thank you for the tips :)
Soon I will try my training method to avoid model collapse. If that doesn't help, we will hope for the release of the non-distilled version of the model.
Thanks. Will try this regardless and looking forward to the update.
Also, are there any specific resolution that can be considered optimal for this? or we go with the old 1024x1024. Thank you.
@makaveli_313 You can use both standard SDXL resolutions and higher resolutions (such as 1280x1280, 1536x1536, and others).
@SG_161222 there is now an undistilled version of schnell (i.e. Apache 2 licence) available.
Would that help in this case?
@StableLlama This might help :)
Thanks for your hard work
@chuckest Thanks :)
Thanks! I find dev to be much superior to schnell, if you could focus effort on dev that would be awesome!
@testuser1111 Anywhere I can read the difference between this Schnell and Dev?... or if someone can explain in short layman terms I would appreciate it : )
@SG_161222 "we will hope for the release of the non-distilled version of the model."
It already exists: https://huggingface.co/nyanko7/flux-dev-de-distill
@Yuri42069 Yes, and I am already working on my model based on the non-distilled model.
https://civitai.com/models/812892/flux1-dev2pro?modelVersionId=908995
This should help with your model training,.
Thank you! I think this could help me :)
how can i use this checkpoint on replicate.com?
Thanks for the first dev version.
Thanks! The model follows the clues very well.
Thank you :)
Thank you, looking amazing. Please add the NF4 version if you can
we definitely need more trained flux checkpoints, but you've told us what it can't do... what can it do that base flux can't?
The same as base flux with a bunch of unnecessary loras, for example
Yes! Dev version is here already baby, could you please upload a Compact Version of it as well please?
@SG_161222 thank you!
Love you bro <3
Hello everyone!
Today I am starting work on a more promising model, Verus Vision. This is a non-distilled model, so there will be far fewer training issues compared to Flux Schnell / Dev.
There will be no further updates to RealFlux for now.
You can follow the progress of Verus Vision training HERE.
If you want something non-distilled, why not going for this
https://huggingface.co/nyanko7/flux-dev-de-distill
It's flux dev but non-distilled, it works well, I think you'll have way less trouble training it than vanilla flux
@Yuri42069 Verus Vision is just based on this non-distilled model.
This looks very promising - glad to see you active with the Flux "can of worms", and am looking forward to seeing progress very much!
it's exciting to see someone biting the bullet and starting work on finetunes. we really appreciate your hard work
im confused. what exactly is verus vision is it based on FLUX?
@SG_161222 Then maybe this is also worth paying attention to https://huggingface.co/ostris/OpenFLUX.1
after all, Schnell is licensed Apache 2.0
Interesting, are you able to upload an fp16 version? Or GGUF Q8 made from fp16? thanks
So this is an actual fine-tune, right?
It won't turn every character into a woman?
Thanks a lot for sharing, but my question is: how is it made? Is it a finetune? A merge? Some information would be appreciated. After all, it’s still 11 GB to download… a few details would be helpful
Currently, I am conducting quick tests of Verus Vision on the RealFlux dataset, and it is working much better.
Right now, there is an intensive collection of images for the Verus Vision 1.0b training dataset.
You can find sample images from the test version HERE.
Are you using undistilled Flux Dev as a base? If so, would you mind sharing what trainer and settings are you're using? It seems like undistilled Flux Dev is untrainable but I'm probably using the wrong settings.
Hello, are you using Undistilled Flux as a base for Verus Vision? If so, would you mind sharing what trainer and settings you used? It seems that Undistilled Flux Dev requires different training settings than regular Flux.
almost done now! excited to see the results
Hello,
So this Verus Vision is another kind of rendering method?, like stable and flux?, or is it just a model name you've picked and is simply based on flux : )?
@makaveli_313 Hi! This is just a new name. The model is based on Flux Dev (de-distill). Most likely, there will be no updates to this model either. But I am glad to announce that RealVis based on SD3.5 will be in training soon.
@SG_161222 Oh I see..
So you feel SD 3.5 will be the future then, any reason specifically you are heading towards that direction instead of just updating Flux based model?
@makaveli_313 Flux is quite challenging to train, especially when it comes to a large model with many concepts. I’ve tried several methods and ended up with one that I used for Verus Vision, but the downside of this method is that it leads to a dead end rather quickly.
SD3.5 has fewer parameters (8.1B for Large and 2.5B for Medium) compared to Flux (12B), which will allow me to conduct local training. Another major advantage of SD3.5 is the absence of censorship and its good license. Yes, SD3.5 may fall short in some aspects compared to Flux, but all of this can be improved through training.
@SG_161222 Thanks for the detailed answer, I understand much better now.
Looking forward to RealVisXL 3.5 v1 Ultra Turbo lol : P
Which or where is the 'compact' model, as I'm getting a "no clip" error? Thanks
I've really enjoyed past RealVis models, but this one still, unfortunately, needs a lot of work. It is not capable of producing male anatomy and struggles with non-white males. I hope to see this model refined until it meets the standard of the creator's past models.
would you like train flux inpaint model?
This Model is Bananas! B-A-N-A-N-A-S
This great model is available online here.
https://app.giz.ai/assistant?mode=image-generation&baseModel=flux1&checkpoint=realflux10b_10bTransformer.safetensors
Pls make a gguf q8 from the original fp16 if possible.
Less consistently good than regular dev but I've got some of my best generations from this model. Lower floor but higher ceiling I guess. Looking forward to updates.
Hi, great work with this model. I was wondering if we could get a lora extract of the Dev model? I tried extracting myself with Kohya with network rank 640 and Clamp Quantile at 1, but I received an error "safetensors_rust.SafetensorError: File does not contain tensor double_blocks.0.img_attn.proj.weight". I am unsure how to remedy this. Thank you!
CLIPTextEncode
'NoneType' object has no attribute 'tokenize'
https://i.imgur.com/GKEDJEz.png
Transformer Dev
The Transformer version of the model does not include Clip. Connect the DualClipLoader to the Clip inputs where needed.
@SG_161222 Thx.
tried using this in comfyui - placed in check points and would not work. I am using the flux 1 work flow.
What version are you using?
Normally, the Flux models are supposed to go in the "diffusion_models" folder in ComfyUI, and you need to use the appropriate node. I was very surprised that the creator of this model said it should be placed under "checkpoints".
Normally, you're supposed to place Flux models in the "diffusion_models" folder in ComfyUI. I was very surprised that the creator of this particular model mentioned the "checkpoints" folder instead.
@Alienfodder The Compact version should be in the "checkpoints" folder, the Transformer version in the "diffusion_models" or "unet" folder.
@Alienfodder That's why I asked about the model, compact is checkpoint, transformer is unet, everything is written in the description of the model. I'll add that dev compact is here https://huggingface.co/SG161222/RealFlux_1.0b_Dev/tree/main/1%20-%20Compact%20Version
@minimishka Thank you very much for the info! I'm gonna have to pay more attention to model descriptions from now on. :)
@minimishka Sorry for late reply I was using the transformer one - newbie so I didn't notice the other options etc.
@SG_161222I will try that! Thank you!
It's okay when it works, has nice variation in facial characteristics, but Flux's chin is still prominent. Often gives very blurry or unrefined images on basic text-to-image generation with Euler Beta (for example)
my theory is that the flux team knows what they're doing and what their models are capable of, so they've done something to prevent people from turning their product NSFW. But at least you can do some thing with boobs (.) (.) but not cats (.) :(
Out of curiosity, you wrote that version 1.0b is unstable. Mind to elaborate what we can expect as unstable?
any shot you make real sd 3.5
Hi everyone! Verus Vision is now available on Hugging Face and CivitAI.
Best checkpoint! I use it with forge. Does anyone know how can i use a consistent character?
Can you upload FP16 versions please? I am training Verus Vision to compare :D
Thanks for an update here. Can you confirm that the latest Schnell is fully Apache 2 and you didn't integrate Dev into it?I intend to use it for commercial purposes.
Are you kidding? The workflow you linked is a PNG file! Can you please inform us how to use the model?
Have you tried reading what is written under the author's sample images? It starts with something like "Using the Model". No? Or dragging png files onto the ComfyUI workspace, for example?
no one want to know what is limmited!!!
what you try too train ???
That is brilliant job done! I m impressed!
Good thanks really appreciated the clear instructions.
this s is 4step ?
upload version with FP16
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.