MORE Consistent Characters & Emotions In Fooocus (Stable Diffusion) 

ComfyUI Please Please 🙏🙏

That link isnt working.. neither one (in the description or above)

@DesignerVIsuals hmm still works for me, and it's still getting downloads according to the stats. ❓🤔

Just dropping in to share something exciting with my fellow content creators! As a newcomer to RU-vid, I've been all about storytelling and creative video-making. Recently stumbled upon VideoGPT, and it's been a game-changer, giving my videos that polished, professional touch.

Can’t find the initial tutorial you talked about.. :/

Thanks! I appreciate the support 😉. 👍

Thanks! I'm glad the videos are something to look forward to. Much appreciated. 😊

No problem! Glad to help show people other ways to get results.

No Problem!

You're welcome 😊

Welcome! Glad you liked it!

I have done LoRAs from generations with mixed results. I haven't put enough work into training since it can be frustrating, lol.

When doing it this way, having the prompt, the pyracanny grid, and the same seed. We can get emotion without compromising the original image because we are only adding in one or two words to all of that information. When putting in an image and trying to get emotion without anything to holding the face to its original form, you will get what you want but not without the character details changing too much. What you can try is putting an image into the "Upscale or Variation" and using the "Vary(subtle)" and simply put in the emotion in the text prompt. It's better if the face is majority of the image as then it wont change as much as if the face was far away. You can also try going into the debug menu on the debug tools tab and go down to (Forced Overwrite of Denoising Strength of "Vary" ) and change the denoise level of that. Vary (subtle) starts at 0.5 so try going down to 0.4 or lower if the image is changing too much.

​@@JumpIntoAI Hi, also thanks from my side for your excellent tutorial! It helps a lot. I was first creating multiple models to then later choose the one that I'd like the most. When going back to create the emotion grids, would you approach it in the same way as pointed out above, or does the seed + details from the log help me in any form to start from the selecting your model point? Thank you in advance! Edit: Nevermind - the log file usually helps to get back into the editing

@@JumpIntoAI Hi, I have the same problem, I generated the model I wanted, but when I tried your process, it changed the model even if I lowered the Forced Overwrite of Denoising Strength of ‘Vary’. I get stuck on this step

you can prompt for it in the generation, if that doesnt work well enough you can change it in inpaint after. If wanting to change the grid of 4 you can mask all 4 subjects hair and try to change at once with inpainting.

Used a background removal tool after I split them. If you are on windows even MS paint has one, not as great as others but a pure white background like that is easy to remove with it.

thanks for clarifying @@JumpIntoAI

I used 9 head positions before which worked but since the heads were smaller the quality of each was lower. So i switched to 4. stop at and weight can be difficult to understand at first. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-0vzunoCYiMI.html that explains it some but might still be confusing. As far as the array function and adding actions I haven't tried. All depends on the image really, asking a closeup face to smile is different than asking a character to do an action. But always worth a shot!

@@JumpIntoAI Ah ok, I'll check that video. Do you know situations other than facial expressions, where this type of array function can be used? Or other possible array functions? For example, I was wondering if it's possible to vary other parameters like adding a awway for the function to generate 4 images, each one with values equal to 0.5, 0.6, 0.7 and 0.8, in the same prompt, like the facial expression array, in order to generate each image with a different set of variables.

And another question, if I'm not already being an annoyance... Kkkkk Is there a good way to alter specific characteristics of a character while prompting. For example, if I just want the right arm to be up, or the left leg to be sideways... I was struggling with that a lot yesterday. @@JumpIntoAI

IT seems that this array function works for clothes colors, interesting.... But I couldn't find online a reference for the term "array function" used in this context @@JumpIntoAI

@@BiancaMatsuo The array function was only added in 2.2 a few weeks ago so haven't tested it myself extensively. But right now generating different weights with a lora automatically isn't possible, actually the lora text prompt command like that doesnt work in fooocus (but i think they are looking to add it) so possibly in the future.

Yes there are a few ways to achieve this, the best is to describe the subject from top to bottom. Like, "A woman wearing a t-shirt, jeans and cowboy boots." adding in their shoes helps as it will want to include what you are describing so it will be more zoomed out. Also describing their surroundings. especially items that are bigger than them like. "A woman wearing a t-shirt, jeans and cowboy boots. Standing next to a lamppost. " If still having trouble, you might try removing the 3 default fooocus styles in the "Styles" tab. they tend to lean more towards portraits.

@@JumpIntoAI Ok, i'll try that. And this works with any models? Or some models are trained to create portraits only?

google collab has limited gpu credits in the free tier which fooocus quickly maxes out. Only solution to this is to either switch to a paid tier or run it on your own system.

​@@dhruvarora2995 so is there anyway to generate high quality background or blend things perfectly with face swap and then typing prompt manually?

This method really only works with generating an image first, all the factors like the prompt/seed/grid faces/ stop at & weight settings, all hold the image close to the original and changing just a word or 2 for emotion is why it works.

CivitAI, just set the filter on there to SDXL 1.0>Checkpoints. Those are all the base models that can be used in Fooocus.

Yes, I also need it

Fingers and hands are still a pain. Especially if they are closer to the camera, and holding hands or ones with the fingers interlocking are nightmares. I usually end up hiding them if i can. But I have fixed them in the past with some basic photoshop (just separating the fingers) and then running again in Fooocus inpaint. If I can come up with something that isn't too complicated that works I will make a vid.

@@JumpIntoAI thanks a lot bro, nice idea, cut them off and run it again, should be easier for the AI to regenerate whole new thing.

@@AxelWXChannel Yep, I have even found a 3d hand that i spent hours getting the right angle and then low denoising in inpaint over and over to blend in naturally. But it was way too much work than it was worth.

all models from CivitAI civitai.com/models/152525/realism-engine-sdxl

One thing to remember is those people probably have some resources in post processing and perhaps some well trained LoRAs. And I bet they still generate hundreds of images for every one good photo. But to do this yourself isn't impossible if the clothing you want is not isn't very specific, doesn't have lots of minor details, and no words or logos. With a lot patience and skill with inpainting it could work. If I come up with a workflow that makes sense I will make a video about it.

@@JumpIntoAI I've thought that using a body created in 3D (created with daz3D, Blender or similar) , you can render it and after that, use AI to make a faceswap on the render. So, you'll always have the same body (sizes of chest, hips, etc..) Even you can change clothes, make physics on clothes, etc..) I don't know if it'd be possible ,I'm only thinking in loud voice :)

Just to check. you have used fooocus fine before trying this yes? Are you on google colab or using this on your own computer? What model/checkpoint are you using? What is your GPU if using on your computer?

AMD Ryzen 7 3800X 32gb Ram 4070 Ti 12gb Vram Images take around 12 seconds to generate each

First is to uncheck the 3 default "styles". They lean towards that studio look.

is the 4x4 grid set to pyracanny?

@@JumpIntoAI Thanks a lot! I've missed that part in the video..

If you describe the clothing in detail you can get similar clothes in generations. But its hard to get exactly the same. there will always be slight variations in shape and color, especially if the clothes you want are specific looking. You can also go back and try to inpaint clothing to change it, especially if its close to what you want and just needs a color change or a missing button, etc.

AMD gpu aren't officially supported (its a stable diffusion problem not a fooocus problem) but there are ways to get it to work as long as it is a beefy card. Fooocus install page has a few things you can try to get it to work. github.com/lllyasviel/Fooocus?tab=readme-ov-file#linux-amd-gpus Or this person created a little tutorial to get his amd gpu to work github.com/lllyasviel/Fooocus/discussions/2552

@@JumpIntoAI Thank you!)

With this specific method no. It will be a different face. The reason I can create different emotions is because I am using the exact settings/weights/prompts/seed/model that created the original and tweaking just one or two words. So it wont change much.

@@JumpIntoAI ok. Thanks. Makes sense

You need to download additional models. CivitAI is the main site for this. Just use the filter in the model tab and look for SDXL 1.0 Checkpoints.

Have you tried using a Lightning Model?

@@JumpIntoAI I get this from Civit Ai that’s name Realities edge XL lightning+ turbo that one brother ? I’m installing fooocus on hdd if I change to install on SSD that’s have some change for me or not brother ? And if this 1660S 6GB too slow for this Fooocus ?

@@zeezee4760 Yes an SSD will help. and Yes any of the lightning and or turbo models will help since they use less steps (3-8). They do need specific settings and changes made in the advanced/Debug mode to get them to work correctly. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-DRcxsqnhjws.html That will show you what settings need to be changed. that gpu is on the low end but I would try out the lightning models and see how much it improves.

@@JumpIntoAI I’m followed ur step brother but still too slow to generate a image with only 4-6 steps but quality is quite good love it maybe I need to upgrade my GPU

you can download it from CivitAI and place it into your fooocus\models\checkpoints folders civitai.com/models/152525/realism-engine-sdxl

This method isn't tied to Fooocus. I can't say for sure the results without testing it but I have used something similar in A1111 in the past with no problem. So I don't see why not.

in the 4x4 grid select PyraCanny

@@ljimenez29oh thx

AMD Ryzen 7 3800X 32gb ram 4070 Ti 12gb (I do speed up the video during generations)

github.com/lllyasviel/Fooocus?tab=readme-ov-file#download Easy install if you have Windows.

thanks, i watched some other videos while i was waiting and found comfydo you know if its also possible to generate multiple angles of a face from a single reference photo like you did in the original video?@@JumpIntoAI specifically "Create Consistent Character Face/Body/Clothes From Multiple Angles" at 0:33

@@infectioussneeze9099 Yes you can face swap with a single photo and get different angles. It can be hard to do and keep the consistency since you will need to turn down the steps and weights, otherwise the face will always be the same angle or close to the original photo. One way to help is using a photo of a similar face in the position you want set to pyracanny slot, and then your reference photo set to face swap onto that.

Is the grid set to pyracanny?

Yes! A well created LoRA is much more efficient in the long run for a single subject. But it can be too confusing for most people to do themselves and not free to do online. Other alternatives are always nice.

I agree that it is overall a better tool out of the box, but it still suffers its own issues with blurry or too smooth skin. And really, this is for people stretching the limits of Fooocus.

Using the 3d face grid in the input image section? Pyracanny selected? 1024x1024 resolution? if its blending into a single face by the end, the stop at needs to be higher. If its generating a single face from the beginning than there is a step your missing in this process.

@@JumpIntoAI thanks I fixed , was not on PyraCanny , good videos

Make sure the 4 face grid is set to pyracanny

Thanks, also had this problem.

@@JumpIntoAI yes but still not working

Indeed! I Kind of wish I got this done a few days earlier, lol. But thats ok. Like you said, paid vs free!

I agree! I prefer Reactor in auto1111/forge. But lots of people don't want to move beyond the easy interface of fooocus so i like finding ways to stretch the limits. I might be beating a dead horse by now, but, oh well.

@@JumpIntoAI Your def not overdoing everything. This is great and cool to learn. Gives people who are willing a bit more insight to how all this is working.

@@JumpIntoAI roop just takes less time, its exactly for something like this, imoi you should mention it at the beginning that theres easier faster way but needs auto11

Well, you can't please everybody. This method works fine in A1111 also.

Omg, just be thankful

It’s not a type of SD! It’s just an interface. You can make it even without any interface using script.

You do get some images that are 4 faces yes? If it only morphs to one image some times then you might need to turn the "stop at" up some on the image grid. The grid is set to pyracanny correct?

@@JumpIntoAI So it is! It's because I didn't set the grid to pyracanny!!! a million thanks to you!!!!❤‍🔥❤‍🔥❤‍🔥❤‍🔥❤‍🔥

@@QianruLiu No problem. I have done a few videos on the basics of this so I went a bit fast. I realize I didn't mention that setting this time. My fault.