- This topic has 38 replies, 8 voices, and was last updated 1 month ago by
.
-
Posts
-
With Flux hands are about 90% correct now.
Great! 90% is good enough. Next step is to acquire a supercomputer with 24 GB VRAM. Does NZXT have any rental machines available with a 4090? 😀
".... acquire a supercomputer with 24 GB VRAM"
That's the recommended VRAM for Flux lora training. For using Flux my 12GB 4070 performs surprisingly well. I've saved a lot of money and could buy a 4090 if I wanted, but I can't justify that to myself. It's just not worth it for AI rendering and I have no plan to train ATM. Rental?... I know you're kidding 🙂
@legolas18 So which upscaler do you use for hirez?@frank22 FLUX. 🙂 I mean I re-feed the image to itself, and it just adds extra details.
The McBoaty upscaler is a great way to add additional details on a high res upscale - definitely recommend that
I've come to the conclusion that character consistency, even using a lora and specific prompting, is more difficult than with SD1.5. Everything else follows prompts very accurately except my character which is all over the shop. Maybe if I trained my own character it would be better but I'm not sure I've got the VRAM to do that.
Maybe if I trained my own character it would be better but I'm not sure I've got the VRAM to do that.
Go on, get that shiny new 4090. Make Christmas merrier for Nvidia. 🤑 😀
But surely there must be some online service for lora training that's way cheaper? Maybe a rental cloud GPU?
Certainly not at Xmas time. It's an expensive month for me... car insurance, 3rd party insurance, registration, strata fees... I'm not broke but it's a bad time of year. Stop mentioning rental or I'll put a hex on you.
Alright, I'll stop. 😀 But I wasn't joking when I mentioned cloud GPU providers.
Also, somehow I find that image scary to look at, lol. Take it down ASAP.
🙂
Hi, you probably already know that I still consider DS superior to AI for illustration needs BUT I am always experimenting and willing to listen to differing opinions. So after reading much hype about Flux (and in spite of my unhappy previous experimenting with SD) I decided to try again with an open mind. Following available advices and extensive documentation I installed Forge (with Stability) and downloaded huge packs of checkpoints and LORAs to test the thing. I supposed my 24 VRAM GPU would be great (it was very easy and fast with SD1.5), I even compared full brutal F1D, nf4 and fp8 versions. The results were poor and really not much better than SD1.5 AND it was extremely slow (with constant poor handling of memory, nf4 was faster maybe because it has VAE included and I may have not correctly used the right ones for pure F1 and fp8. About the LORAs, i tested a beautiful girl (HAfu for flux) and I had excellent hyper realistic results as long as I strictly used the same prompts as the example image (with no visible hands...). as soon as I tried to become creative, giving her a sword and a full body position, nightmare happened (three hands with demonic fingers, sword flying while I had used long descriptive sentences). I get the impression that I did not understand something or that there is really overhype, I thought that Flux was slow but efficient (no need to have 1000 images for only one acceptable). But really the hands question is not solved.... that was the main reason for my return to AI, I used a "detailed hands LORA" in the same composition and... it was worse (as if LORAs were conflicting).
That was just to express my disappointment, not to prove a point or start a flame war. I really would like to know if there is for Flux a good way to improve hands, (backgrounds and expressions are OKish). I would like to be convinced that AI is a true thing and not a 32 million dollar pic randomizer burning GPUs... I am still willing to learn. Thanks for any advice or help (notably an opinion about the best VAE for the various flux iterations).@corbeau I have never used Flux with anything but ComfyUI. There it works pretty well.
To improve hands, I just use Adetailer for hands at about 50%. I am also using one for faces, but at 25% denoise (or 0.25).
As for LORAs in general, I find that many are hit or miss, often require to be set at a lower strength than indicated too. Also, the best checkpoint, so far for me, is the base one.
For best results, though, I am using an LLM where I just put in a little description, and the LLM generates the descriptive text making the whole thing look better (all in one workflow, using "only" 50 GB of RAM or so).
Also, it seems that LORAs triggers are kinda useless.
Finally, try using deis (sampler) simple (scheduler) at 30 steps with Flux-dev (I use 16fp only, btw).
For VAE, just use the original one ae.sft
Hope this helps. 🙂
@frank22 It's Christmas time...everyone wants their "gifts". I see that end of year payments are not limited to any country.
Also, I am waiting for the RTX6090 at this point. I have the 3090, I see no point to spend close to 2k to upgrade to the upcoming 5090, even though it'll have 32GB of VRAM.
@legolas18
Thanks a lot for these precious pieces of information. I will launch a batch of tests with Comfy as you suggest. All your hints are very useful and could help me avoiding certain traps.
Have a nice Christmas time !@legolas18 - "... using "only" 50 GB of RAM or so.." ? I don't know what that means. Comfy is more customisable but does it produce better images? The UI looks like a dog's breakfast. IMO Adetailer makes the face skin look shiny with too much contrast and fake-looking. There seems to be a issue with skin in general, and not just with Flux.... far too tanned, rosy cheeks and getting pale/fair skin is virtually impossible no matter how I prompt for it. I can fix most of the colour/contrast issue in postwork but I shouldn't have to.
And don't give me cleft chins unless I ask for one damnit. Were there 5000 images of Kirk Douglas in the Flux dataset?
@3dnoob - "..cloud GPU providers" I could use Civitai but I don't have many points. All the other ones I have to pay I think, which is against my principles.@frank22 It means that loading an LLM + Flux + 4 LORAs + Adetailer models takes up some RAM (it stays loaded into it unless you close comfy or you un-load them with a button).
As for skin, try using deis (sampler) simple (scheduler) at 30 steps and avoiding using LORAS with obviously overcooked skins. 🙂
As for Adetailer, you can actually adjust the de-noise...set it to something lower like 0.25. Also, the model you are using to fix things in Adetailer is also important. I am using the Anzhc Face seg 1024 v2 y8n (No Dill) from Civitai
https://civitai.com/models/293448?modelVersionId=577987
The "No Dill" it's because Comfy doesn't auto download it, and so it usually gives error. Also, Segmentation models seem to be working than the bbox ones for me (their selection is more precise, but I haven't found a No Dill hand segmentation model).
- You must be logged in to reply to this topic.
