Since this is a dedistilled version of "Flux schnell", Apache 2.0 license applies making it free to use commercially.
This fine tuned checkpoint is based on Flux schnell de-distilled thus requires a special comfyUI workflow and won't work very well with standard Flux dev workflows since it's uisng real CFG.
This checkpoint has been trained on high resolution images that have been processed to enable the fine-tune to train on every single detail of the original image, thus working around the 1024x1204 limitation, enabling the model to produce very fine details during tiled upscales that can hold up even in 32K upscales. The result, extremely detailed and realistic skin and overall realism at an unprecedented scale.
This first alpha version has been trained on male subjects only but elements like skin details will likely partically carry over though not confirmed.
Training for female subjects happening as we speak.
Highest quality slow version:
Turbo version with shamelessly good quality!!!:
Resources:
Lora:
Fast Lora: https://huggingface.co/ostris/OpenFLUX.1/blob/main/openflux1-v0.1.0-fast-lora.safetensors
Turbo Lora: https://huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha
Text encoders:
ViT-L-14-BEST-smooth-GmP-ft:
t5-v1_1-xxl:
https://huggingface.co/city96/t5-v1_1-xxl-encoder-bf16/resolve/main/model.safetensors?download=true
Clip G:
Recommended settings:
Without Turbo and Fast Lora
CFG: 2-4
Steps: 50-60
With Turbo and Fast Lora
CFG: 2-4
Steps: 8-16
Turbo Lora weight: 1
Fast Lora weight: 0.33
This Flux variant works with negative prompts so use them if you want to. Negative prompt will work with and without the Turbo Lora.
Description
Q3_K_S GGUF version
FAQ
Comments (99)
This is such a polished project! Happy to try it soon. Did the training include North Africans, Levant-Arabs, and Turks? The gens look amazing! :D
Thank you. It didn't include that specific nationalities yet. Have to up my data scraping game haha.
This model is like magic. I honestly don't understand how it can work so well. It's a leap in improvement compared to other models. You could stop the training here and I'd still find this very useful. Bravo !!!
Thank you. I was brainstorming to train it this way since quite a while but didn't have the means to create the scirpts needed to process the dataset. Thanks to chatGPT I wrote a bunch of scripts that automate it. I think this is also the missing step for others who create variants of Flux like Flex, Flux1 lite, etc to give it the extra detail.
Why is the Q2 base model listed as SD 1.5? Is it a mistake?
Fixed
Hi, your work is amazing, i am very happy that you have achieved this.
but i have a question, can you tell me how you did it?
"that have been processed to enable the fine-tune to train on every single detail of the original image"
Thank you! I've cut up high resolution images into smaller pieces so I can train it.
@tamtamx1332 what captioner do you use for the image caption?
Guys, any idea why sliders are not visible?
Delete Mixlabs custom node.
Really gr8 piece of art.. Thanks
Very very good, congratulation !
Will this work in Forge (without ComfyUI workflows)? In which case can you supply some settings for CFG, Steps, Sampler, etc. ?
I’m but sure about Forge but CFG needs to be around 3.5 and steps around 50 without the turbo and fast Lora and 13 steps with. Turbo Lora weight 1 and fast Lora weight 0.33.
did you test it with Forge?
@AshrafHamdi I haven't yet.
Excellent job, works on people just fine too. Also you made a discovery I noticed works on all Flux models: you can use triple clip loader and add Clip-G to the mix to increase quality. Nice find.
It's listed as Refiner. So I figure it's supposed to be used as a Refiner, not as a base model. Right?
Nope it's a model but it's a dedistilled version of Flux tha'ts why when we use that type of model we chose refiner since Civitai doesn't care to put a separate category for it.
This is absolutely awesome, I'm running the fp16 model together with the Turbo-workflow in ComfyUI, and getting great results! Just one question: Why did you choose to go with CFG 3 instead of the Flux-standard CFG 1, and why are you increasing the CFG for each upscale? I'm fairly new at this, so I'd love some insights into things like this to better understand them :)
Thank you. The CFG is not 1 because this is not regular Flux dev but Flux dev dedistilled which puts it on par with Flux pro. Blackforestlabs has distilled the model before releaseing which was limiting its capabilities. This model won't work well with a regular Flux dev workflow.
Excellent! The quality is amazing! I have a question: Can the upscale be used without using the image generator? I mean, can I use my own image and apply both upscalers? I would like to see the results that way.
Thank you. Yes I have an upscale version here. https://drive.google.com/file/d/14Bw_4OTDYh6LKoxT3tgTRSfEVSOd8jkq/view
20 minutes for a render.. you gotta be kidding.
In my case it takes 69 secs to give the output. RTX 3090 WIN 10
@jtsanborn Ok, I somehow got it working now... yes its taking about 1 minute .. Detail does look good
@jtsanborn what? how? I also have 3090 and 128 of ram (not that this matters i guess) but the turbo workflow takes about 20 minutes for me as well. Not sure what I am doing wrong? or is it because I use the fp16 model and I need the fp8?
@Belshazzar I had to redo the workflow. Apparently, some models were not set up correctly, and it was causing this take forever and produce bad results. So I hate to say it, but you probably still have something wrong in your workflow and you may want to start over. That's what I had to do.
@trashkollector175 yes, I switched to the fp8 and it's better. thanks!
Depends on your hardware configuration and which workflow you are using. Are you using the turbo worklfow?
I can't run this on comfyUI with nordy.ai
nordy.ai?
@tamtamx1332 https://nordy.ai/comfyui
Great model. I'm getting to smooth image in upscaler, what settings should be changed to retain all skin pores and imperfections? Thanks
Where are you using it? If you are using comfyui then I would recommend to use my workflow in the description.
Can anyone confirm the compatibility with Forge UI please?
Testing it now. It is compatible but you have to use sd mode not flux because it’s using real cfg.
Tested Q8 with forge. It works, but not with the recommended text encoders, I always get errors with these. I use regular clip_l and t5_xxl_fp16 instead. Also you can use it in flux-mode, you only have to put the "Distilled CFG" slider to zero and the "CFG" silder to 2. Is really slow tho (despite this model being fllux-schnell based), but with the fast LORA you can (almost) gen at the speed of flux-dev.
where can in find ? t5xxl_1.1_bf16.safetensors
Check the description I've linked it there.
Just a tip for who wants to try: using ModelMergeFlux SigmaVision and PixelWave makes the details even better.
Pls send the links🙏
Yes, please send the links!
@BlairBrooksBB I’ll try to upload an image to the gallery with the comfyui wf
@Primaveri Thank you for uploading the workflow as I was able to get it up, but I'm missing the ModelMergeFlux SigmaVision model. Where is it located? Is it on Civitai or another site? I see PixelWave on Civitai.
@BlairBrooksBB there's a disabled node that will merge Flux Sigma, Project0 and Pixel Wave, The new merged model will be saved to folder the "checkpoints" in ComfyUI output folder. Then you just paste the new model in "models/unet" folder
@Primaveri I don't see a disabled node in the workflow that does that function. I uploaded a screenshot of how the workflow looked under one of your images, so you can see how it looks on my end. Am I supposed to download this node as well?
https://ibb.co/23c5gmxw
@BlairBrooksBB try this one: Image posted by Primaveri
@Primaveri Thank you so much, that worked for me, very well. I was able to get that checkpoint and things are working on the regular workflow that you gave me. Everything's good to go, if I have any further questions, I'll let you know here, but so far so good as these images are coming out really good. Again, thank you very for your patience in answering all of my questions, as I hope this will help others as well. Take care! 😀
This is really amazing....but the disillusion does not wait. This is great for head portrait only. As soon as you go to close-up / full body portraits I noticed that it has massive issues with creating proper hands / feet..maybe a heritage from Flux schnell?? At least I have problems using the original turbo workflow. Anybody without hand / finger issues?
Yes since it's an alpha version.
Yes, I'm having the same issues with hands, I haven't had any feet shots yet. It seems to be more perfect for headshots for now. It does its thing with upscaling, but definitely not the perfect final product yet, or close to it. It needs a lot of work before I can really use it as a standard workflow for my generations.
Thanks for your reply @tamtamx1332 , makes sense. Anyway looking forward for your women trained release , your approach / project is fantastic! Thanks for confirming @BlairBrooksBB. Wonder if training the flux-dev-de-distill would have been a more solid base, but then you'll have another licensing.
@nirodha07211 definitely a more solid base and much more realistic results. And since the architecture is based on Flux schnell it’s free to use commercially.
@nirodha07211 You're welcome, glad to see somebody else having those problems with the hands and feet. Hopefully, it gets better as time goes on.
Amazing quality! Well done! Really really good! I can't wait for the next update! Do you think it's possible to train the Dev model to have this kind of quality as well? Or it's only possible through this distilled version of Flux Schnell? Because for a full body photo there are body disfigurements and extra fingers. Anyway you are amazing for creating such a great model! Thank you so much once more!
This is pretty great as the results truly are very close to what you get with Flux Pro. The main con is that it seems to only generate portraits well, anything beyond that like showing hands and anything below the waist it gets a little wonky. It still needs some work on that end, but this is the best progress made as an alternative to using Flux Pro which doesn't use regular LORAs at the moment.
Right now, I find that the upscaler for uploading images works really well, so I might generate those images in my original workflow then bring them over to the upscaler to get sharpen up with the details on the skin which it seems to be doing amazing at thus far, in my honest opinion.
I would love it once we can just use this to generate full body shot images and I don't have to use anything else.
How does this compare to the de-distilled flux-dev model that has been around for a while?
Skin details are much more accurate at any tiled upscale level. Skin tone and color as well.
First of all, I truly don’t know how to thank you for sharing this amazing model. I’m blown away by the quality and i can't believe it has only 2K downloads? really?
and the workflow is exceptional too!
I’d like to ask if you could kindly point me to the files used in the workflow, particularly a LoRA called Detailed V2 Flux NTC.
Hey thank you! I think people being put off that it's only trained on male subjects but it's doing women as well as far as I can tell. https://civitai.com/models/666950/extremely-detailed-flux-slidersntcaixyz
@tamtamx1332 exactly, that was my guess too.
well i can see kind of bias sometime but overrall..hat down! wow.
I'm still trying to understand how that first upscale works even if is set at 0 denoise. what wizardy is this? 🤷🤣 ty for the link
@LatentDream hey thank you! Really at denoise 0 it's still doing something? Maybe there is a hidden cap how low it can go haha.
Has anyone managed to get it work with controlnets?
Controlnets only works with dual clip, unless you use a pixtopix method, which is not good. Just change the Triple Clip Loader to Dual Clip loader and your'e good to go.
Amazing and really well done, like some others I seem to not be able to do anything involving hands or feet which is a shame as it's probably my favourite checkpoint so far for Flux. Seems even with hand Loras It still won't change if anyone has any ideas on how to fix?
First of all amazing work! thank you for sharing this. but I am not sure why but both the slow workflow and the turbo workflow are extremely slow on a rtx3090 and 128ram about 15 minutes
even on the Turbo workflow.
The downloaded workflow, both turbo and slow has sigmaVisionAlpha1.safetensors inputted by default, but both the fp16 and fp8 are named differently when downloaded.
For the turbo to work as it should, should I use the fp8 model? not sure if I am doing something wrong?
Also in the dualClipLoader the longclip_L, not sure where to get it? I placed instead the clip_l encoder.
is that the same or not?
Thanks again for truly amazing work!
Anyone have a success with i2i with this?
@tamtamx1332 I’ve been playing around with this workflow for a week. It’s amazing for upscaling, but I couldn’t get it to work for style transfers (e.g., converting a 3D-rendered portrait to a realistic one). Any suggestions on the settings I could try?
@tamtamx1332 link not working
I would love so much to give this model a top review but unforunately is difficult to impossible to use for anything beyond head shots. You can scale up a face with unbelievable details. But it sucks big time at doing hands. I hope this will be improved in future versions.
Congratulations, this model is really fantastic!!!! The best I have ever tried. With an Asus Rog 3090 in 10 minutes a superlative image comes out! Hi. When will the female version of this model be released?
Thank you! I'm working on a dataset for females. If you have some nice images let me know.
@tamtamx1332 CIAO SI HO UNA BELLA COLLEZIONE DI FOTO SE TI PUO' SERVIRE.
Hi. Yes if you need I have a nice collection of high resolution photos
Hi. Yes if you need I have a nice collection of high resolution photos
Hi. Yes if you need I have a nice collection of high resolution photos TO USE TO TRAIN YOUR FEMALE MODEL WHICH I HOPE YOU WILL DO SOON
@tamtamx1332 tengo un conjunto de imagenes de mujeres en 2k y 4k. muy buen calidad. tambien de hombres is te hiciera falta y texturas de piel en 2k y 4k tambien. 200 imagenes de mujeres y 200 de hombres. entre 20 y 40 años la mayoria. tambien tengo un dataset de headshoot de imagenes de studio. si necesitas algo dimelo
Can this model be used in Forge Webui? or only comfyUI? Thanks.
Will never know
This looks very good. Which de-destilled model did you use as the base checkpoint?
Thank you. I've used the regular Flux dev de-distilled.
@tamtamx1332 Oh, but then wouldn't your fine tune be a dev model and not a Schell model? I am curious because dev models are not apachi 2.0 license. but you fine tuned it to be a apachh 2.0 schnell model somehow?
need new workflow, all mxtookit nodes are broken, change it no result
This model is amazing, I appreciate your work.
Any timeline on the model you are training on women?
This model does tend to put more masculine traits on the women in my images.
Something is off with the workflow. All the slider nodes are not responsive, some are disconnected at all. Running the workflow produces a mess, not an upscale. I believe with current ComfyUI version this is no more compatible. Please advise.
Also, instead of doing super sphagetti mode, i would suggest getnode, setnode from kijai, workflows will come much more cleaner :)
Some updates are breaking the sliders but the latest should work. Sorry I hate the set and get nodes. They are adding extra work and make things confusing. In the new comfyui you can just press the eye icon at the very bottom right of the screen to hide the noodles.
check if you installed "mixlab" nodes, this one breaks sliders and also hide the slider !
there is a fix, if you search google, by changing the python file. but i smply deinstalled "mixlab" and all is fine.
You are listing Flux.1-Schnell as the Base Model above, but in the comments you've stated that it's a tune of nyanko7/flux-dev-de-distill – which is it?
It's based on flux schnell architecture I'm guessing since when I check the architecture it shows as "Flux Schnell"
@tamtamx1332 Wait, are you unsure what checkpoint this is based on??
looking at the checkpoint file size and the quality of the output it seems like its the dev de distilled right. the workflow is also very very similar to the dev de-distilled ultra Realistic + Detailed Portraits one on openart
is there any way to train lora model for this model? i have my own flux dev trained model by fluxgym but it has bad image generate on this model.
Never heard from "Ladies first!"?
is it possible to use it on 3060 12vram and 80 gf ram?
if yes how would I do that and how to use it in Image to Image if anybody can help please let me know thanks