@SG_161222 sad for tensor but if skin looks already good in base 3.5, i'm pretty sure we can have an amazing version with your works, very excited about Realvis!

@SG_161222 thanks!!

@SG_161222 Thank you for quick replay

@SG_161222 yes

Ok so this is more an experiment than a full finetune :-) I get you didn't use the merging process but the finetune one then ? Will try your checkpoint soon

@SG_161222 is there a reason why you do it through Lora and not direct fine-tuning?

@SG_161222 It turned out to be a very realistic model but so far I've tried with Lora and am getting artifacts.

Hopefully the 5090 will be a big leap as 'speculative' articles are stating it will be.
Similarly the more accessible 5080 hopefully a noticeable jump in it/s compared to 4090 : )

Thats same for all de-distilled models. Simple solution is to increase steps a lot, if it doesnt work, then it doesnt work.

I think it might be possible to use some of acceleration LORAs, but Im not sure how well they play with FLUX control nets, if they work at all.

@Mescalamba Ok, I will try to double the steps and see if it works better. That would be excruciatingly slow, but hey.

The Shakker Labs Union Pro tile controlnet works fine with Alimama's FLUX.1-Turbo-Alpha 8-step LoRA with distilled Flux models. I have not tried it with de-distilled ones like Verus Vision yet. I will post my findings when I do.

@mmdd2543 Well, from my experience some Turbo/Hyper/Schnell loras work with individual de-distill models, but I havent try any control net yet. Mostly cause Im usually happy if those models work at all.

@SG_161222 awesome ty so much now i can test

They don't really work well together at this moment according to my tests.

@AugmentedRealityCat maybe you know, is it working with flux controlnets?

@manpotent I get bad results with the Shakker-Labs ControlNet Union Pro controlnet in tile and depth modes. The images come out fuzzy and lacking detail. :(

Try to set Diffusion in Low Bits to Automatic (FP16 LoRA). It works for me for this and other models too.

@wregiszter632 I assume you are talking about ForgeUI? Do you know what would be the equivalent of this in ComfyUI?

@mmdd2543 Yes, I use Forge UI. I'm sorry, but I don't know how to do this in ComfyUI.

@wregiszter632 No worries!

i know, now you know you need 4090 :)

Ok, now I know what my problem was. I did't notice the part where it says this is based on the de-distilled version of flux. I was using the distilled CFG like on most flux models, but after I read the de-distilled thing, I used the normal CFG instead and it's working fine.
If someone else is having this problem, just put the 3.5 value in CFG, instead of distilled CFG.

@PepitoPalotes Hello, I too use Forge but had issues with another Flux on Forge as well.
Could you share the full settings you used please?, Sampler, Scheduler, Sampling steps, Hires.fix if any?

Thank you : )

@makaveli_313 just follow the instructions in the description and it should work fine. I have tried other options, and some work and some don't. For example, don't use karras scheduler, or SDE samplers. DPM++ 2M (the non SDE one) works fine though, and SGM uniform scheduler is giving me great results. The [forge] flux realistic samplers also generates good realistic images, but it's slower.

Workflow?

@zerocool22 just snag it from any of my pictures here or on the creators hugging face there is a workflow shared there, mine is a little complex but i share my workflows for all my pictures

can we take a look at workflow?

In my case it does look better than pixelwave for my prompt, OP did a great job

@offmybach you can copy and paste full workflows from imagese from this site (as long as the user did not disable metadata) by either 1. downloading any image here made with comfy (it will say workflow or it is tagged as comfy) and loading that image into your comfyui or 2. the nodes button on civit copies the full workflow to your clipboard and you just paste into a blank worklow in comfyui, best to look for that "nodes" button by any shared image here @zerocool22 i hope that helps

are you using the Adaptive Guider? Try setting the real CFG to 5.0. This model blows you away once you figure out the correct workflow!

@adfsgrv yeah I am my other post literally mentions about it. And I’ve tried the workflows here and others. You have to use adaptive guider also for colossus 2.1 also as its a similarly made model, setting it between 3.5-6 depending on the tokens and whatever luck but colossus comes out mostly noiseless whereas noise is very hard to get rid of with this one. I did some experiments you’ll see my images vary in quality I posted below some come out just fine others are “artistically noisy”, it could be the data it could be the training it could just my video card as the biggest problem with flux is just how much damn tuning is available and importantly different prompts react differently with flux settings.

@joehorse I'm getting shockingly consistent and robust results with it. Every single run.
Here are some finetunes from my workflow:

I use Euler Beta with 30 steps.

So far so usual.

I never use negative prompts.

For the positive clip I use the Triple Clip Loader:

Clip G

Clip L from Colossus

T5xxl FP16

For the (empty) negative prompt that does not work.

There I use Dual Clip Loader with

Clip L standard

T5xxl FP16

Don't ask me why but still when leaving the negative empty and applying the Triple, I get broken results.

I wish you best of luck with your workflow!

@adfsgrv ill give it a try with a fresh workflow no copy and paste, i was using the GGUF Q8 T5, as it loads a bit faster, but i can run the FP16

@joehorse also I never go above 896x1152 or 1024x1024 for the initial generation. the upscaling of the initial i do with this workflow:

Amazing Flux_SD3.5_8Mega_Pixel_image_upscale_process - Pixel Version | Flux Workflows | Civitai

@adfsgrv so thank you man, i created a fresh workflow doing things as you said even the clip thing and it worked. but yeah i suspect it was my pixel size actually, its just some models don't have that problem but other models probably dont have this much realistic training data with real backgrounds. pictures came out sharp, ill post in a bit but yeah even thought this is stuff i did already know, sometimes with comfy when really want one part of your workflow to work you get confirmation bias

@adfsgrv 
before: https://civitai.com/images/39058684
after: https://civitai.com/images/39116236

before: https://civitai.com/images/38904925
after: https://civitai.com/posts/8926269

thanks for the positive post to try again, happy to admit when i made a mistake :)

@joehorse looks great! What I'm also really digging lately is to do img2img with a high denoise instead of text2img. Like a poor man's controlnet but works like a charm with VerusVision.

Image posted by adfsgrv

Image posted by adfsgrv

Load image -> Vae encode -> as latent into the sampler.

I'm just searching for great photos on the internet and use them as input.

Denoise ~0.75.

Definitely look for real high quality photos as input!

Have fun!

@adfsgrv I've been testing something similar with creating images with the dev version and then running them through img2img with the same promt but with Verus to get the more realistic look. You are right 0.75 is the magic number in my experience. But even going 0.85 can yield interesting results without breaking the image, though of course the image then changes quite a bit.

try flipping "dedistilled model" switch in your workflow

does work I've produced very realistic images with it.

@LahotiLahore if you're talking that noisy shitpic produced by this model is realistic image... then nothing to talk here.

It's not working! And this is not only for me!

@aeronliru which noisy shitpic? let me show you some of my generations maybe then you'd agree, look at my posts if you want to take a look at what this model can produce.

@LahotiLahore dude, I've chewcked fp8 model severak times - it's quality is crap with settings author mentioned in overview.

Just waste of time

@aeronliru In forge set both Distilled CFG Scale to 3.5 and - this is important - CFG Scale to 3.5.

It'll be slow but much better results than with forge defaults for flux (Distilled CFG Scale 3.5, CFG Scale 1.0) - default setting results are really bad, so looks like this model was de-distilled during training.

I've just generated one example image with these settings: https://civitai.com/posts/11119566

@wo6163 i figured out the same indeed. when you run 3.5 distilled you get this : Distilled CFG Scale will be ignored for Schnell

so yea normal cfg scale is the way to go in forge

You're not the one. author says it works but even with settings proviided in overview picture gets shitquality at all.

cfg to 3.5 not only guidance

Can you be a bit more specific? What do you mean by Flux UI? I've never heard of such a thing. Did you mean ComfyUI?

@mmdd2543 when I say: "I want to use stable diffusion" everyone says "automatic1111" or "fooocus" I guess these two are user interface but system is the same. so my question is "how can I use flux" should I download comfyuı? thx.

That's strange. All my images look great when using your prompts and seeds. So looks like your settings might be incorrect.

Are you using ComfyUI? I tried to import workflows from the images you posted, but it seems like you either didn't use ComfyUI, or you decided not to save the workflow in the images. If I have a look at your workflow, I might be able to tell you what's wrong.

In any case, if you use the settings as suggested by the author - CFG 3.5 (most Flux models use CFG 1), sampler: Euler, scheduler: Beta, steps: 25-50, empty negative prompt etc, you should get pretty good images. I use these settings and I don't have any issues.

@mmdd2543 Are we talking about the same model? My images were crappy looking. I don't use comfyui or any other out of the box tool, but my own. I have generated a lot of images and when they come out bad it isn't my workflow.

Não eu também não estou conseguindo