Model Type: LoRA
Trained On: Flux Kontext (via AI Toolkit)
Best Results With: Flux Dev (T2I / I2I modes)
Trigger Word: unflux this
Purpose: Remove unnatural anatomical exaggerations and synthetic skin effects generated by Flux
Note: Even though this LoRA was trained in Flux Kontext, the results obtained with better quality in my opinion are using it together with Flux Dev through T2I or I2I.
in I2I using Flux Dev I liked the result when using the sampler: ipndm or res_2s and most importantly the scheduler: kl_optimal or beta57, it allows you to increase the denoise and have more subtle changes in the input image, so even using a very high denoise like 0.83 (ipndm + kl_optimal) or 0.55 (res_2s + beta57) it will still change little in the reference image.
to unlock access to the latest samplers and schedulers (kl_optimal, res_2s, beta57) that will provide better quality you can install this custom node on comfyui: https://github.com/ClownsharkBatwing/RES4LYF
š§ What is UnFlux?
UnFlux is a corrective LoRA created to address a specific issue in the Flux family of models (especially Flux Kontext and Flux Dev):
These models tend to generate anatomical distortions and synthetic skin textures, leading to unrealistic or stylized outputs ā especially in portraits.
UnFlux v1 was trained to neutralize these distortions and restore a more natural appearance, while preserving the structure, style and lighting of the original generation.
Although trained on Flux Kontext, UnFlux surprisingly performs even better on Flux Dev, producing smoother, more balanced results.
šÆ What UnFlux Fixes
This LoRA focuses on correcting exaggerated or biased details frequently seen in Flux-generated images of people, such as:
Overly defined jawlines and chins
Straight and unnatural mandibles
Square noses with unnatural bridge shapes
Over-emphasized collarbones (clavicles)
Plastic-like skin without texture or tone variation
These are not traditional "sharpness" issues, but anatomical rendering artifacts that result from Flux's internal bias toward stylized body features.
š§Ŗ Training Info
Trainer: AI Toolkit
Steps: 8,000
Datasets:
Control dataset: 156 images generated by Flux with exaggerated anatomy
Target dataset: 156 human-corrected, realistic versions
LoRA Rank: 32
Optimizer: Automagic
Resolution: 1024p (v2 will target high-res realism)
The goal was to teach the model to convert Flux-style artificial anatomy into more natural, humanlike representations.
āļø Usage
šø Flux Dev (recommended):
Works best in Text-to-Image (T2I) or Image-to-Image (I2I)
Add
unflux thisto your promptApply LoRA at 1.0 weight for visible effect
šø Flux Kontext:
Also uses
unflux thisas triggerLess effective than on Flux Dev, but still works in transformation pipelines
Recommended weight: ~0.6 to avoid excessive softness
This behavior was not intended and is expected to be corrected in v2
Note: Although the LoRA wasnāt trained for skin smoothing, on Flux Dev it appears to soften the skin slightly ā an effect not intentionally included in training, and subject to further investigation.
š What's Next in v2
UnFlux v2 is already in development and will feature:
Training on high-resolution target images
Restoration of natural skin textures (including pores and tone variation)
Improved correction for stylized anatomical exaggerations
This upgrade will move beyond just āfixing distortionsā and focus on adding realism and depth back into Flux generations.
š„ Download & Use
Just drop this LoRA into your ComfyUI or A1111 pipeline.
Add unflux this in your prompt and set the LoRA weight. No extra setup needed.
š¢ Feedback Welcome
Help shape future versions! Share your feedback, results, or use cases.
Description
FAQ
Comments (27)
Awesome work done here!
An absolute machine in training. Amazing work, master Alisson! šš
What made you come to this conclusion?
@zGenMediaĀ I appreciate his work! Heās always helping the community, especially the Brazilian one, which lacks high quality technical content. Constantly sharing knowledge on Discord and creating videos to help beginners who are just starting! Everything heās doing, heās sharing with the community. I always joke around with him on discord calling a training machine, the guys never stop always producing and contributing.
I could say more, but I think you get the idea behind my conclusion.
When you created your datasets, specifically the 'Control dataset' of 156 Flux-generated images and the 'Target dataset' of 156 human-corrected versions, were these images paired directly? For example, was each image in the target dataset a corrected version of a specific image from the control dataset, essentially using the same subject or scene before and after correction? Or were these two independent sets of images?
The mention of '156 images generated by Flux with exaggerated anatomy' and '156 human-corrected, realistic versions' for your datasets is clear. But could you elaborate on how these pairs were generated (if that was your strategy in the first place)? Specifically, did you take existing natural/realistic images and transform them into the 'unnatural Flux look' via img2img (or a similar process with Flux) to create your 'control' dataset, and then use the original realistic images as your 'target' dataset? Or was the process different, perhaps starting with initial Flux generations and then manually correcting them?
@OlliFlysTheRocketĀ the process was exactly that, I took images of real humans in the second dataset and passed them through flux in an I2I with a denoise of 0.5. Another thing I did was in the first dataset, because there were 2, in this case the 156 images were because there were 2 datasets of 78, in the first the images were generated in Flux and refined through I2I using Stable Diffusion 3.5, it may have been my mistake, I don't know, after this process I still passed the target images through Topaz for quality refinements, this was the first dataset, the second was the one where I took images of humans and passed through Flux to add the imperfections, as the training was already in progress, I took the opportunity and took a save of these Loras during the training and used it as Lora in Flux to add the imperfections, only instead of putting a positive force I put a negative force to activate the effect of my Lora in the opposite way.
@AlissonerdxĀ Science Bi***. i like it, had a similar thought the other day of reverse engineering the training. you just took it to the next level.
@Alissonerdx Bro, amazing. Did you heard of the LoRA SameFace. I think it works the same as you say. I am using that LoRA with a negative value too. Around minus 0.3. Works great. How much rendertime went into this? I feel at some point there will be new Checkpoints that have that capability, maybe because they will merge with your LoRA (although i dont know how that really work to be honest). Anyway. Keep up the good work.
That is the SameFace LoRA i was talking about:
@OlliFlysTheRocketĀ This SameFace is very interesting, so the training took about 8 hours, excluding the time I spent preparing the dataset, but now I'm preparing a dataset with 200 realistic images and generating the control images with flux + NAG.
example images does not look that much different from flux girl images. what did you change?
I didn't understand your question, can you elaborate?
So you made an "UnFlux" LoRA without getting rid of the FluxChinā¢? š
I appreciate your efforts to reduce the plastic skin gloss, but keeping the Flux chin seems like a bit of an oversight here
What do you mean? The chin marking is reduced, of course it's not perfect but that's why I made it clear that there will be a V2.
This was meant as a joke more than anything as a noticeable amount of your example showcase images show some degree of FluxChin which seemed ironic to me because of the name "UnFlux" not remedying one of Flux' most glaring biases.
But I'm glad you're making a v2 that possibly factors that in! š
@RedPinkRetroĀ I don't understand how this is a joke? I honestly don't understand your point here, have you even tested the model? Did you read the descriptions or are you just a mere commentator? If you don't want chin just increase the strength of the Lora and that's it, you can clearly see that the chin marking is reduced through the comparative images, the way to use it depends on the person, one of the main problems of the Flux is the marked chin but the plastic skin continues to be a greatly relevant problem.
U can stet the Model Blocks Buster to
## 2 = 0.98
to reduce the Flux Butt Chin
@denrakeiwĀ Thanks man, but don't worry, I wasn't struggling in getting rid of the FluxChin. I was only amused by its presence in the first place. The creator just appears blind to my sense of humor it seems and chose to take offense instead.
Can't explain colors to the blind and all šš¤¦āāļø
@RedPinkRetroĀ What exactly was the joke? Was it the laughing emoji alone that made it funny to you? Or was it some play on words that we don't notice? or a clever piece of irony? Was it that you were saying something that would be absurd to say? Ohh! I just got the joke! It's a complaint about something you authentically see as a problem in something someone is working hard to make, giving that work away for free, and saying they were still working on further improving the shortcomings, and that makes it funny, right?
I think it's perfectly clear which one of you doesn't have a sense of humor, if you think you made a funny. Some failed lora results are indeed hilarious, but this isn't one of them. You mentioned something that was actually improved, but not enough for your demands. That doesn't make it funny to point that out.
@JellaiĀ That's exactly the point, sometimes people don't even test things and send me a comment saying that I forgot something, which is one of Flux's biggest problems. It doesn't cost much to download Lora and test it. If you want the Lora effect to be more evident, increase its strength. It's that simple. Yes, in the images I put as a showcase, the marked chin appears, etc., but that doesn't mean I forgot about it. Of course I know about it. Just read the giant description I left talking about the effects and the improvements that can be made in this V1. My type of conversation is direct and based on evidence. If it's not like that, it's better not to talk or comment at all. In this case, it could have been a simple feedback: "Look, I think the chin marks are still evident" and not a "So you made an 'UnFlux' LoRA without getting rid of the FluxChinā¢?" You see? Anyway, I don't like to extend this type of conversation. Those who want to use it, use it. Those who don't, train their own model or use the "thousands" of other models that don't even come close to solving this problem, apart from those that use giant workflows with sigma and several upscales applied to try to solve this problem. This is just a lora.
Great Lora, looking forward to UnFlux v2!
Great job
Please share your results, it helps to improve this lora.
Life saving lora