My (subjective) take-aways, judging by your examples: PuLID cannot compare to the other two. InstantID wins on integrating the face into the environment, and making it look realistic. FaceID wins on capturing the essence of face, if you don't worry about photorealistic shading too much. Great work! Super useful to see all three of them in action with the same parameters, and be able to judge the result. thx!!!!
Thanks for sharing your insights! It's great to hear your perspective on PuLID, InstantID, and FaceID. I'm glad you found the comparison useful. Thx for watching!
wish you had included execution time in your analysis. I've also found that InstantiD is best but also slowest. I wonder if it's possible to beat it with a combination of faceid and upscaling/face detailer approach
InstantID looks the best but FaceID contains sharper details but seems more cartoonish for some reason. InstantID seems to at least try to keep some of the skin texture. I'm not chomping at the bit to get my hands on PuLID after seeing this. Thanks :)
Yeah that’s because faceid is intended to copy the STYLE of a person’s image while defaulting to the base model trained data. Basically if you use faceid with a model that has cartoonish trained data … it’ll produce cartoon that looks “kinda” like the person. Mattheo went into this quite a bit and explained it. He even said it’s best to use a combo of both instant id and Face ID.
I would love to see a video that compares models and loras, and a chart of which ones are best at which things.. some models suck at accurate face swap
I wrote it to other videos as well - the license is the problem here but it is good that you point it out that user needs a license for commercial use. So far I know this is very expensive one. That means those three ones are quite useless at the moment. The InstantID team wanted to retrain without using Insightface. But that was 4 months ago and so far no progress. The only alternatives I know are: 1) IP Adapter using Face model (not Face-ID) or StableIdentity (It does not have ComfyUI nodes and is using weird SD 2.1 model). Any other alternatives out there?
I find InstantID far, very far better than the others, there is a weird thing with it: it excels to extract face features even from a "bad" picture. faceID is just good and pullID project should be closed, the most recent one yet laughably useless
My (subjective) take-aways, judging by your examples:
PuLID cannot compare to the other two.
InstantID wins on integrating the face into the environment, and making it look realistic.
FaceID wins on capturing the essence of face, if you don't worry about photorealistic shading too much.
Great work! Super useful to see all three of them in action with the same parameters, and be able to judge the result.
thx!!!!
Thanks for sharing your insights! It's great to hear your perspective on PuLID, InstantID, and FaceID. I'm glad you found the comparison useful. Thx for watching!
LOVE your eplanation of the workflow logic of the comfyui workflow, makes it super easy to understand what's going on
Awesome, thank you!
mate, this type of videos are soo good ! comparing similar models! u should make more and more !!! XD
Wow, I did not expect PulID to come last.
wish you had included execution time in your analysis. I've also found that InstantiD is best but also slowest. I wonder if it's possible to beat it with a combination of faceid and upscaling/face detailer approach
I would honestly probably just stick with LoRAs for now because all three of them were not consistent and showed some weaknesses.
Please make a tutorial on how to install PuLID on your computer.
InstantID looks the best but FaceID contains sharper details but seems more cartoonish for some reason. InstantID seems to at least try to keep some of the skin texture. I'm not chomping at the bit to get my hands on PuLID after seeing this. Thanks :)
Yeah that’s because faceid is intended to copy the STYLE of a person’s image while defaulting to the base model trained data. Basically if you use faceid with a model that has cartoonish trained data … it’ll produce cartoon that looks “kinda” like the person. Mattheo went into this quite a bit and explained it. He even said it’s best to use a combo of both instant id and Face ID.
Good point!
I agree with you. InstantID is the best at similarity and detail.
Which one will apply a straight forward face to a side facing face the best?
I would love to see a video that compares models and loras, and a chart of which ones are best at which things.. some models suck at accurate face swap
Great video! Can you make a video about how to install PulID? I've downloaded all the files and always fail. Can't figure out what step is missed.
I wrote it to other videos as well - the license is the problem here but it is good that you point it out that user needs a license for commercial use. So far I know this is very expensive one. That means those three ones are quite useless at the moment. The InstantID team wanted to retrain without using Insightface. But that was 4 months ago and so far no progress. The only alternatives I know are: 1) IP Adapter using Face model (not Face-ID) or StableIdentity (It does not have ComfyUI nodes and is using weird SD 2.1 model). Any other alternatives out there?
Can you incorporate a move2Move workflow?
Scarlet really likes others using her likeness for ai so i herd ;) 😂 thanks for the comparsion video, really interesting.
Glad you enjoyed!
I'm getting so many python conflicts after getting pulid working. It's been an absolute nightmare fixing it
Great test! Thank you!
My pleasure!
Amazing videos, thank you so much
My pleasure😊
lol i like with all the controversy lately you use scarletts image lol nice
Better than an over use of musk
this is wonderful ever
Thank you for the kind words
good job, just saw pulid repo has add v1.1 examples
Thanks for the heads-up! I'll check out the v1.1 examples in the pulid repo. Glad you enjoyed the video!
I find InstantID far, very far better than the others, there is a weird thing with it: it excels to extract face features even from a "bad" picture. faceID is just good and pullID project should be closed, the most recent one yet laughably useless
thx, sir
Most welcome