You could try different setting. Mine generally don't have the bad chin. I use DPM2 Beta, guidance scale 0. I don't use swarm so I am not sure what options you have for settings there.

Thanks. I haven't used it much myself. Just so much slower I just have so much to do, so little time.

As far as I know it works as normal in Comfy as long as you have the right loader for it. Then you just have the clip, vae and text encoders set like usual. The main difference is you disable the Flux Guidance Scale and use the regular CFG setting which seems to work best between 3.5 and 8. Here is the workflow I use on the rare occasion that I ever use Comfy. https://civitai.com/models/897505/multiple-loader-basic-flux
I am sure there are other loaders, but the one I had found that worked was the Mixed Cast Flux Loader. This workflow has that and 2 other loaders for other types of models such as GGUFs and just standard models.

@andy82 I have never tried using any Loras with it myself. I just made the De-Distilled version because someone requested it. I haven't really used it beyond making the test images for the model gallery. For me it's just too slow to be useful on a daily basis. I haven't tried using the base De-Distilled model with Loras either, so I don't know how it works with them in general, so I'm afraid I can't be much help with that aspect of it.

That's great. I rarely ever use any version other than the 8 Step anymore. Especially been using the Chin fix version almost exclusively since I put it out.

Thanks a lot.

I use the DPM2 sampler with Beta scheduler mostly. Main thing that will be different between De-distilled and regular is with De-distilled you keep the Flux Guidance Scale at 0 and use the CFG instead. Seems that CFG works best at 3.5 and up. With the regular models, CFG is 1 and I also keep the guidance scale at 0 normally but sometimes use 2.5.

I should also point out that these setting are for me, on my hardware, using Forge UI. I don't know for sure what setting work best in other programs such as Comfy. I almost never use any except Forge.

@nox4892252 Yes, I don't think it is built in to the model itself. Again, in Forge at least, I leave the VAE and text encoders always selected in the software so it is always enabled no matter which model I am using.

@hades6666969 The guidance of 0 with the higher CFG setting is ONLY for the De-Distilled version. With the chin fix version of the 8 Step, CFG is always at 1 and the guidance can vary, but I usually get best results with guidance at either 0 or at 2.5 to 3.5, but I almost always use 0.

That's strange. I don't get that. I will get it once in a while if I am specifically prompting for a woman with black hair, but only rarely. I can't say for sure why you would get that, but one possibility might be which sampler you use. That can make a pretty big difference on some things.

Thanks. I do try to only merge stuff that adds realistic quality. I do love the Flux base, though I do think it needs some expanding on the types of faces. I don't want to erase the ones it has though.

It's the regular Flux vae, just called ae.safetensors. Here are the ones I use: https://imgur.com/eI4j1ps

I recommend the Hyper 8 Step versions. I can't really see any quality difference and they are just a lot faster. Either the V1.1 Hyper 8 Step with Chin Fix or V1.1 Hyper 8-Step. All of mine except for the De-Distilled one are FP8 versions as well.

@2018cfh  I hope you like it. It is mildly NSFW but not really a porn type checkpoint. The porn type example on there are mostly a combination of Img2Img with my Flux checkpoints along with some inpainting using my SDXL Acorn Is Boning checkpoint, so they are not straight porn generations. I get bored sometimes with the same old generations, so I often find cartoony Pony outputs and turn them into realistic looking images with my checkpoints.

https://civitai.com/models/897505/multiple-loader-basic-flux

I can't say for sure since I never use Comfy (I have used it, but in day to day use, I really don't). In general, I don't use euler either. It was the first working sampler for Flux I think, but I have found that DPM2+beta is way better. If Comfy has DPM2 I would try that. As for the FluxGuidance, I am not sure if it works exactly the same in Comfy as in Forge, but I almost always have the FluxGuidance at 0. If not at 0, I have found it has to be up to at least 2.5 as it looks awful at 1 or 1.5. Other than that, I can't help much with Comfy. Overall I find Comfy to be a frustrating nightmare to use, so I really only use it for merging and once in a while just to see if something works in it.

I don't see anywhere on his page where there is any Acorn Is Spinning Flux models listed. They have a few Flux models but I can see no info on them as to a name. It would never surprise me if someone puts out my model as their own though. It has happened before on SeaArt and I had to go through some steps to show the dates I uploaded my original models as someone on there had created a username to match mine on here and upload 5 or 6 of my models as their own. Luckily all the dates are documented on the site here. I always upload here with new models so it's pretty easy to prove when someone else uploads it a week later claiming it as their own.

@isv I have never uploaded any of my Flux models to tensor art. I have never cared for the site much. I have a couple of my SDXL models up over there, but I have never been able to get them fully active there. I have had them come up as duplicate too which just makes me wonder how they even determine the difference in models. Their system seems to suck. I just gave up even messing with them. I have been on Civitai since I first started with models and have always uploaded them here first. Never had a problem here.

Yeah, I think the makers of Flux must have built in some kind of poison pill type thing with regards to nipples and pussies. I am not sure how to solve that, but keep looking for solutions. My Acorn Is Spinning Flux models are not trained, but merged with a number of different Lora models to nudge things closer to what I want. It's not a perfect solution, but so far it's the best I have. I am always testing new things to see what works. I have been wanting to attempt actual training, but my computer is not really up to the task and I don't have money to buy online compute power to try it.

@rzzg I always check out those checkpoints whenever someone puts one out that claims to be NSFW, but I have yet to see one that is any good. Whatever they are doing to Flux to get their results seems to me to make the output quality less than it was with SDXL. For instance, the one you linked here, just looking at the gallery output, doesn't look nearly as good as my Acorn Is Boning XL model for SDXL. For me, until a Flux version can be done without losing the realism of Flux, it is not really worth doing. I get better results using my Flux model and then in-painting it with my Acorn Is Boning SDXL model. Or even simpler sometimes is using the good outputs from Acorn Is Boning and do what I call reverse in-painting with my Flux model basically in an img2img way.

As for training, I am not really very familiar with how to do the training for Flux. Since I have never had a PC that powerful, I have never taken the time to go through the tutorials to do actual training. I see it seems to involve a lot of command line type scripts and stuff, which I have never had any luck getting to work for anything so far. I was trying to convert some of my models to GGUF because people were requesting them. Finally I just had some people who are better at that stuff convert them for me. I would like to learn the training stuff at some point.

Yeah, Flux won't do those. When I generate images of that type I use a combination of different methods and models. The main way for me is do the image in my Acorn Is Boning XL model and then put the output through the in-painting tab as an img2img basically. Mask off the parts that Flux will destroy, then basically run the whole image through Fluxifying everything except the masked parts. Then the whole image looks pretty much like a Flux image. Takes a few steps, but works pretty well.

That is strange. I don't really use Comfy to generate images, so unfortunately I couldn't say what might cause that. If I were just guessing, I would think maybe they updated something with the samplers or schedulers, but it would just be a guess and I don't know why it would affect one model but not another.

Which version of the model are you using?

@97Buckeye Only other thing I can think of is to make sure you are having the distilled cfg off and using the regular CFG instead. First image I generated in the De-Distilled version in quite a while I forgot to change those setting and it looked pretty bad. I tried it out after you mentioned the issue, and other than forgetting on the first one to change those settings, it works perfectly for me. I use Forge though rather than Comfy so it is hard to say what Comfy might have changed that might affect it.

@97Buckeye Oh, that's good then. Not sure if I have ever used a negative prompt with it or not.

I have been working on an updated version. I have been testing one that drastically improves nipples. I have been creating some Loras myself to merge with the model. It's promising, but not sure when I might put it out. I am considering whether I want to put out an experimental version so other people can test it and give feedback. So far I have only done one quick merge that is promising, which makes me want to do a better one before I consider it a real version.