- This topic has 24 replies, 6 voices, and was last updated 5 hours, 36 minutes ago by
.
-
Posts
-
@Frank22
Definitely raised a smirk 🙂I took your advice and looked at StabilityMatrix.
I have installed it as the portable option on what is now my spare laptop [which is an Alienware 3080].
It looks quite impressive - all I have to do now is get it all to work and produce something nice. Quite a lot that I haven't seen before, so will start slowly with the SD options.
If I manage to make sense, and progress, with it I will try the portable option to mov things over to the 4090 laptop.
{anyone actually used this option?].I am glad to see wholesome and family friendly book covers 🙂
No worries, frank, I am also home for the holidays, so I have been experimenting a loooot with LORAs these days. I mean...I have so many, might as well test them out.
I did find something that really appeals to me...but so many combos to test out
I looked at some YouTube videos re. Flux lora training and they talk about training times (Flux Gym) of 12+ hours for mid-range GPUs like my 4070 at 1024px.
That's just too much commitment for me considering I can fully train a SD1.5 in 40 minutes.
It's going to be 41C (106F) here today so that would be a problem too with running a GPU flat out all day. Like most people here I don't have aircon.
Besides that, I don't think Flux loras work very well in the first place compared to 1.5/SDXL. I've tried a lot of Civitai ones and they're very unreliable with character likeness.Yes, FLUX training is brutal, I guess most people train their loras on civitai or other paid service.
As for likeness loras, I had best results either using img2img with the lora, or applying the Lora at the Adetailer phase which I made sure to be the last phase of the image generation, (in the Adetailer prompt I only put the LORA trigger)...which is something I was working on these days btw 🙂
Also, to fix hands with Adetailer, I found a trick..I prompt (in the Adetailer prompt) "woman hand" with [color] nail". It improves things considerably.
Here is an example with both Adetailers (the face is of the actress Jewel Staite, also note the hand). Both Adetailers at 0.5 denoise.
I like your tip about "...in the Adetailer prompt I only put the LORA trigger.." But a lot of Flux loras don't come with trigger words for some reason. I don't do many females with makeup or nail polish and I have more of a problem getting a natural nail colour.
Although Flux was trained on a huge amount of data, with humans I don't think there was enough variety in the dataset. No random faces of random people on the street, unruly beards, random patches of skin blemishes, asymmetric etc. The skin for example is almost always "perfect", over-tanned/brown which is hard to overcome when there's no negative prompting*.
* if I increase the CFG to 1.1 for example I can enter negative prompts but I don't know if it actually works. They say CFG should be left at 1.@frank22 The higher your CFG, the closer the AI will stick to your prompt. A lower number allows it to be more flexible in interpreting the prompt. I generally set the CFG around 3.5 which works for most of my images, although I've gone as high as 5 on the odd occasion.
Eve - CFG Scale or Distilled CFG Scale? Are you using Flux?
Because Flux-dev is a distilled model, it is recommended to set CFG=1 and then not use negative prompts, using “Distilled CFG Guidance” instead.@frank22 The reason many Loras have no trigger is that with Flux it's not really needed. I just add the parts that indicate an artistic style, for example.
As for CFG...it's not even in my workflows and don't bother with negatives, you might be able to up to 4 words max, but it hardly works. Use Guidance instead. I typically use 3 to make images a bit less saturated and plasticky.
As for natural nails, I haven't tried it, but I guess mentioning them in the adetailer prompt would probably work.
@frank22 I match the Distilled scale to the CFG scale. I've always got them set to the same number, usually 3.5
I'm using this Flux checkpoint: https://civitai.com/models/416205/xuer-one-cyan-ten-color
For the images below, I used this prompt:
[Count Dracula dancing an Argentine Tango with a Vampire Queen, elegant pose, dramatic expression, gothic ballroom, cinematic lighting, high contrast, detailed costumes, intricate patterns]
This seed: [4270600746]
The only thing I changed was the CFG. I didn't use any Loras, which is why that isn't Henry Cavill tripping the light fantastic, and I didn't use any negative prompts. My sampling method was [DPM++2M] and I used [Beta] for the schedule type.
Sorry for the late reply Eve, I just got up. I'd like to try that ckpt... it's a pity they're all so big. The person who made the ckpt suggested this site which looks great but it looks like you need to be a Chinese resident to join. I'm definitely going to have to try playing with the CFG scale.
That site has a lot of interesting Loras Frank, I might need to look into registering for it. I'll try that just as soon as I can rip my eyes away from Mr Cavill up there.
I also found this site Eve. Mostly Asian models but not all of them. Signing up is easy, just provide an email address and you're in.
- You must be logged in to reply to this topic.
