Simple Guide To LoRAs In Stable Diffusion | Fooocus
HTML-код
- Опубликовано: 3 июн 2024
- A straight to the point guide about SDXL LoRAs in Fooocus. Where to get them. How to Install them. Best settings, combinations and more!
If you have yet to Install Fooocus, Stable Diffusions easiest AI Image Generator. Check out this video here: • Fooocus Tutorial - Sta...
All the Loras Used in this Vid
Alf - civitai.com/models/135963/alf...
Jean-Baptiste - civitai.com/models/100159?mod...
Deatil Tweaker XL - civitai.com/models/122359/det...
Dissolve Style - civitai.com/models/245889/dis...
Perfect Eyes XL - civitai.com/models/118427/per...
-----
00:00 Intro
00:15 How To Install
01:30 How To Use LoRAs
05:05 Using SD1.5 LoRAs? 
Наука
![Kehlani - Next 2 U [Official Music Video]](http://i.ytimg.com/vi/MpI7ekFG8A0/mqdefault.jpg)
If you don't have Fooocus, you can download it from here: github.com/lllyasviel/Fooocus
All the LoRAs used are in the description.
This kind of content is why I subscribed to the channel. Good information.
And I will continue to try my best. Thanks. 😁
Recently just got into fooocus and I love it
Indeed, its a great piece of software that you can easily miss.
None of the SDXL 1.0 Loras I’m adding are appearing, what might the issue be?
This is very helpful ❤
Thanks! It was useful!)
Great Video, Can u tell me what is Base Model and how to use it ?
how to add lora in foocus webui, please
Great vid! I'm curious to know how you managed to create a consistent fox character throughout the multiple prompts. Do you have a video on making prompts with consistent results for non-humanoid characters face/body? Keep up the great work!
as always very nice content indeed.
Thanks again!
Thanks a lot!
Thank you, it's interesting and concise. 👍
But there is one question: some Lora have a CLIP SKIP parameter: 1 and others have this parameter equal to 2 (CLIP SKIP: 2).
How then to use it?
It's not a setting available to change in Fooocus unfortunately. But the LoRAs I have used with that requirement have still been successful.
Great work , as always.
One question, do we really need to use the trigger words in prompt to activate Loras (using Fooocus client)?
Isn't it enough that we select lora from the list (right side Loras panels) as set up to weigh more than 0 ?
For me, using Fooocus is about simplifying some steps, and not using additional prompts to activating Loras looks like one of those quality life improvements in comparison to other clients.
No it isn't always necessary. Most will work in some form without it, but with a long text prompt or multiple LoRAs, the effect may not happen or get lost behind the others.
Hey man thanks for the tutorial. Do you by any chance know why drag and drop doesnt work for me?
Some people have had issues with that in certain browsers (like firefox) try chrome and see if that fixes it.
Thanks for the vid! Are you planning on making tutorials for ComfyUI/A1111?
Yes eventually, There are many things I want to showcase that are not possible in Fooocus.
@@JumpIntoAI Nice! Looking forward to it
Hello. As a new user to AI, I would like to ask you what Lora would you suggest is best suited for realistic vegetation (forests, plants etc). I'm looking through the list on Civiai but I don't know what I am specifically looking for. Any suggestions will be appreciated. Thank you.
Somewhat depends what you are going for, however if generating a landscape and the vegetation is giving you too much of a fuzzy paintbrush feel. Using the "detail tweaker XL" Lora and at 2 or higher can help. Also Using a 1.5 checkpoint as a refiner (at 0.4) like Realistic vision or Photon can give very realistic results.
And if in fooocus, using the "fooocus photograph" and "photo HDR" styles can help.
@@JumpIntoAI Thank you for replying and the suggestions. I'll look for those right away.
so what do we do for Loras that are a person, and a lora style that is stylized? it seems every time I make one higher, it looks more like the person but at the cost of the style being gone. They are both SDXL loras. Any tips for those situations? and does it matter what order the loras are in? does it matter where you place the trigger word in the prompt?
For this, order might be subjective but I would go with (background lora) (Styled Lora) (Person Lora) in both lora order and trigger. But really, the weights will be the ultimate balancer.
Depending what Loras you are using, I would go in with whatever the recommended starter settings are on the Lora webpage and then start with a simple text prompt.
Then either add more text or weight to lora to direct it where I want it to go. Believe me It can be endless trial and error if you are going for something specific. Plus it helps to lock the Seed and if the first images are a good start, keep going, otherwise try a new seed.
@@JumpIntoAI Earned a sub
Thanks for the next batch of very useful information 😄
One question.
I've been struggling with this for a while. How to force Fooocus to see subfolders in some directory, or more paths than just one for models for example? Because I see that Loras you have in different directories. Of course I'm talking about the "config.txt" file.
I have my config paths aimed at my Automatic1111 Model and Lora folders. I DID have issues with my subfolders in those not showing up at first but not long after it worked (not sure if it was an issue on my end or theirs to be honest) but they "should" work.
Multiple paths for models wont work in the config file but I did see this was a request being looked at.
@@JumpIntoAI - Thanks for the info, I also keep all the models in a special directory, but it doesn't see the subfolders. I guess I have to be patient.
Could you create a video demonstrating how to properly depict a warrior holding a sword or any other weapon? I've attempted it, but I'm struggling to get the hands just right.
Hands and hand positions are difficult. Depending on your issue, using pyracanny with an image of a pose you like helps. or trying to inpaint the area (entire hand and sword and trying to regenerate) all depends on what you are going for. And depending on your picture I have had success using layers (cutting out part of another image and putting that ontop of an image then running it through fooocus to mesh it together.) I will eventually make a vid on that too.
None of the SDXL 1.0 Loras I’m downloading and adding to the Loras folder in Fooocus are appearing as options, what might the issue be?
a few options. The folder is the fooocus/models/loras folder? and you haven't altered your config.txt to point somewhere else?
Are you adding them while fooocus is already open? If so you can hit the "refresh all files" at the bottom of the models tab.
Sorry - totally random question since I can't find an answer elsewhere. Is it possible to disable automatically blurring the background in Fooocus? Thanks in advance
Can try removing the default styles, I know "Fooocus Sharp" tends to have a subject in focus and background blurry, Also putting "blurry, blur, bokeh" in the negative prompt can help.
@@JumpIntoAI Put only those tags into the negative prompt, and removed all styles and still every background is blurry. Shrug. My guess it's the base model forcing it
@@JumpIntoAI Just to update - yeah, JuggernautXL forces a blur. I added the epicrealism refiner with the anti-blur prompts now the background is no longer blurry. Thank you
@@delorne2776 This is intriguing and I would like to try it out, but I did quite a bit of searching and could not find an epicrealism refiner. I'm thinking this must refer to one of epinikion's base models or LoRAs, all but two of which are for SD1.5.
Is possible to insert an AI character in a real photo or image with fooocus?
You can inpaint additional people into a photo and or use face swap on a picture. How natural it looks is another issue, especially if you are going for something specific. If you need someone just standing there not interreacting with anything in the background, possible. If you want to add someone sitting between 2 people or having them physically interacting with others. Probably wont work well.
@JumpIntoAI can you inpaint a picture character's nose, mouth or any anatomical peculiarities to another picture character?
Спасибо чувак, смотрю тебя из России. Очень направится Фукус, тем что он хорошо и прост. Жаль с пересадкой лица еще не очень, хотя Фукус лучше чем Controlnet.
Какая у Вас видеокарта?
Hello and welcome! I have an Nvidia 4070 Ti.
@@JumpIntoAI How fast does it go on this GPU?
great intro vid but you didn't really say why you inexplicably and suddenly add a 1.5 refiner before using 1.5 loras. I think I just stumbled into accidentally using 1.5 loras and never set the refiner to anything. it's hard to say if the results are bad or not as it's really my first time messing with loras but I would say I am getting great results. this was actually the whole reason I came to the video so besides saying generally assumed info that using stuff for unintended versions is going to give weird results or not be as good as xl is simply saying xl is better than 1.5 which I would agree with but I wouldn't warn against using 1.5 lora they work just fine and without an unexplained refiner. people also still cling to 1.5 and i'm not really sure why they would beyond not meeting the hardware requirement to use xl but thats a whole other question
1.5 loras or anything trained on SD1.5 will not work (either not at all or correctly) with an SDXL checkpoint. This isn't saying 1 is better than the other, they just are legitimately 2 separate things and not intended to work together.
Fooocus is only programmed to use SDXL all around(other than using a 1.5 checkpoint as a refiner) so that is why 1.5 isnt discussed much. The refiner trick with loras is probably unintended and only called a "trick" because its easier to load up a1111, use a 1.5 checkpoint, 1.5 lora and get the results you want.
If you are using an SDXL checkpoint and a 1.5 lora and everything working as intended that would be news to me (and most people) you might have mixed sdxl loras in with your 1.5s as that is something i did in the beginning because most programs wont show the difference and only way to know is labeling them or separating them.
SDXL and 1.5 main difference is SDXL is trained on larger images 1024x1024. Where 1.5 is trained on 512x512. SO out of the box you are getting higher res images from SDXL but the VRam requirements are higher (that is not the only difference of course but what's important to most)
Reasons for people not switching to SDXL and ditching 1.5 entirely is VRam requirements and the loads of tools that people created for it but have yet to do for SDXL as extensively.
And now my coffee buzz has ended.😉