0:00 Intro 0:23 How To Use Flux Pulid Paid (Replicate) And Free (Huggingface) 1:28 Best And Worst Use Cases - Strengths And Weaknesses Of Pulid 5:06 Lets Use Flux Pulid - Uploading Our Image 5:28 Prompting Tips 6:28 Additional Settings 8:35 The Generated Images And Additional Tips With More Complex Prompts 10:23 BONUS TIP - Use Seeds To Create Variations Of An Image 11:36 BONUS - Pulid Is Perfect For Generating Datasets To Train A Lora 13:00 Outro
Flux Pulid allows you to create character consistency from a single image. You can generate images of a character in completely different clothing and lighting scenarios. There is no need to use any Loras. In this video I'll highlight both it's strengths and weaknesses. It's great to use if you want to generate an entire dataset which you can then later use to train your own Lora to generate images using a free open source solution like Comfy UI or Forge.
Currently testing this, however this works better with training a Lora. For instance, I have a character Lora trained on the Mona Lisa so I can generate infinite images of her in Forge for free. Then I trained another Lora on an item of a shoe so the shoe is also consistent when it’s generated . So now in forge I can utilize both loras at the same. I can make the Mona Lisa wear the real shoe item and it remains consistent. Using 2 loras at once can get a bit tricky though. I think we will see a dedicated platform for fashion customisation with consistency, it’s already happening.
@@lukaszlr6163 A lora is is a way to teach an AI model new tricks without changing its whole brain. It's like adding a small, specialized skill on top of what the AI already knows. So to simplify that, if you use FLUX, it's a big brain pretrained model that generates images. When I train a LORA, I still utilize the big brain but add something new to it to help me extend the big brains capabilities. So instead of FLUX just giving me random images of cats, I can now train a LORA on my own cat and I can teach the big brain how to generate these images of my cat while still utilizing it's full power and potential. Loras allow for customization when generating images, they are extremely useful.
Referring to the information at the end of the video , I would love to see your workflow for achieving consistent character design in advance live portrait and FOFE, in combination with Pulid, to get different poses. It would be nice if Pulid had some ControlNet features added for pose reference inside replicate it self
Thank you Puja! I can certainly put something like this together, basically a deep dive into lora training and how I actually trained the Mona Lisa Lora with the different facial expressions and poses.
Planning to release a video on this, won't be focused on humans though as there is many videos on youtube for training loras on your own likeness, i'm going to do something different.
True BUT i've got a great solution for facial expressions, i already recorded a tutorial on how you can generate different expressions, check it out ua-cam.com/video/DC3sYgj_byc/v-deo.html
@@thehuskyrocky7051 on replicate you need to search for FOFR, their tool is called consistent-character. When you upload an image it will generate different images in different poses while retaining the likeness.
Yup this is unfortunately true however I did mention this in the video when I was discussing best and worst use cases at 4:44 when I tested pictures on myself. For getting good likeness of yourself it’s best to train a Lora for Flux on Replicate or Fal.
@@TheGoodContent37 I think it holds up very well in comparison to Midjourney but to be honest, Midjourney cref is better as it can do some slightly varied expressions as well. The biggest plus here is that this offers a service similar to Midjourney for a fraction of the cost or even completely for free on hugging face even though you can only generate a couple images. Once a pulid for flux workflow like this exists for comfy ui, then we can do this with open source tools and generate infinite consistent characters from a single image for free, forever, using our own hardware. For me personally I think that open source is the future of Ai tools.
0:00 Intro
0:23 How To Use Flux Pulid Paid (Replicate) And Free (Huggingface)
1:28 Best And Worst Use Cases - Strengths And Weaknesses Of Pulid
5:06 Lets Use Flux Pulid - Uploading Our Image
5:28 Prompting Tips
6:28 Additional Settings
8:35 The Generated Images And Additional Tips With More Complex Prompts
10:23 BONUS TIP - Use Seeds To Create Variations Of An Image
11:36 BONUS - Pulid Is Perfect For Generating Datasets To Train A Lora
13:00 Outro
Flux Pulid allows you to create character consistency from a single image. You can generate images of a character in completely different clothing and lighting scenarios.
There is no need to use any Loras. In this video I'll highlight both it's strengths and weaknesses.
It's great to use if you want to generate an entire dataset which you can then later use to train your own Lora to generate images using a free open source solution like Comfy UI or Forge.
Generating historical figures like Einstein driving a Tesla? That’s the kind of crossover I never knew I needed! 🚗💡
Sounds like a great idea hahahaha, have fun.
Modern Mona Lisa looks cute :D
I think so too hahaha
Thank you. Feels like a breakthrough for me 🙂
It's actually super awesome! I'm using it to create datasets for training loras.
@@aitechwitht This is actually what I've been wanting to create consistent characters for a story I want to animate. It is awesome 🙂
Helpful tutorial. Really helped me anyway. 👍👍
Thank you so much! I'm glad it helped.
Thanks for this video.
Could you explore ways to customize the clothing or adding product placement?
Currently testing this, however this works better with training a Lora. For instance, I have a character Lora trained on the Mona Lisa so I can generate infinite images of her in Forge for free. Then I trained another Lora on an item of a shoe so the shoe is also consistent when it’s generated . So now in forge I can utilize both loras at the same. I can make the Mona Lisa wear the real shoe item and it remains consistent. Using 2 loras at once can get a bit tricky though. I think we will see a dedicated platform for fashion customisation with consistency, it’s already happening.
@@aitechwitht would be nice to see you make a video about it when that platform comes up.
@@aitechwitht This may be obvious to many, but what is "Lora"?
@@lukaszlr6163 A lora is is a way to teach an AI model new tricks without changing its whole brain. It's like adding a small, specialized skill on top of what the AI already knows. So to simplify that, if you use FLUX, it's a big brain pretrained model that generates images. When I train a LORA, I still utilize the big brain but add something new to it to help me extend the big brains capabilities. So instead of FLUX just giving me random images of cats, I can now train a LORA on my own cat and I can teach the big brain how to generate these images of my cat while still utilizing it's full power and potential. Loras allow for customization when generating images, they are extremely useful.
very cool, thanks for sharing!
Thank you so much bud! I've actually watched a couple of your tutorials and they've been very useful! Keep it up!
You can also just use the lasso tool to cut it out in ms paint if you want the exact same face over and over...
great tutorial !
Thank you!
perfect x2
Thank you Mosen!
Referring to the information at the end of the video , I would love to see your workflow for achieving consistent character design in advance live portrait and FOFE, in combination with Pulid, to get different poses. It would be nice if Pulid had some ControlNet features added for pose reference inside replicate it self
Thank you Puja! I can certainly put something like this together, basically a deep dive into lora training and how I actually trained the Mona Lisa Lora with the different facial expressions and poses.
How can we change the posture of the face and emotions?
Checkout 12:28, I explain another service that can generate different postures.
not possible for local install?
Not that I'm aware of BUT I feel like we will probably get a workflow of this for comfy ui soon or something similar.
yeah an implementation for comfyUI would be delighting.
can you explain how to use Lora please?
Planning to release a video on this, won't be focused on humans though as there is many videos on youtube for training loras on your own likeness, i'm going to do something different.
It's a great way to create consistent characters, but it still can't handle facial expressions like sad, angry, or crying faces.
True BUT i've got a great solution for facial expressions, i already recorded a tutorial on how you can generate different expressions, check it out ua-cam.com/video/DC3sYgj_byc/v-deo.html
@aitechwitht Thank you so much. I just subscribed to your channel today, and that's why I asked about this. 🙂🙏
@@AI-EleganceStudio truly appreciate the sub and support, thank you!
how do you create the side view of monalisa?
like 12:36
@@thehuskyrocky7051 it's explained at 12:36 which service is used that generates different poses.
@@thehuskyrocky7051 on replicate you need to search for FOFR, their tool is called consistent-character. When you upload an image it will generate different images in different poses while retaining the likeness.
What about testing a unique object/ product.
It unfortunately can't do products or objects, scroll to 4:53. For stuff like this you have to train a Lora.
@@aitechwitht cheers
Does anyone know how to fix this problem?
Failed to restore node: Ultimate SD Upscale
Please remove and re-add it.
Hmmm, wish I could help but i've never seen this error code, did you manage to figure it out?
@@aitechwitht No, I told me it is because of a knot that must be removed. But I don't know which one and how to find it
Does it allow you to generate NSFW images?
Nope.
With your own photos, it dosn't work. Thats why i want use it.
Yup this is unfortunately true however I did mention this in the video when I was discussing best and worst use cases at 4:44 when I tested pictures on myself. For getting good likeness of yourself it’s best to train a Lora for Flux on Replicate or Fal.
But how does it fair against Mid journey's cref which does the same?
@@TheGoodContent37 I think it holds up very well in comparison to Midjourney but to be honest, Midjourney cref is better as it can do some slightly varied expressions as well. The biggest plus here is that this offers a service similar to Midjourney for a fraction of the cost or even completely for free on hugging face even though you can only generate a couple images. Once a pulid for flux workflow like this exists for comfy ui, then we can do this with open source tools and generate infinite consistent characters from a single image for free, forever, using our own hardware. For me personally I think that open source is the future of Ai tools.
why does youtubers say further ado.
Hahaha I've always said this on all of my tutorials, it's become more of a habit and a way to say "let's get started".