Hi everyone, we’ve seen some great discussions about Headshot 3 and want to introduce our key technical breakthrough.
To significantly improve the accuracy of 3D head creation from images, Reallusion has invested extensive effort in training a proprietary AI model designed specifically for high-precision facial reconstruction.
This model is trained on a large dataset of facial scans, enabling it to better interpret facial landmarks, depth cues, and subtle anatomical features from 2D images. As a result, it produces a more precise head mesh with improved topology, enhanced identity preservation, and more accurate proportions.
The result? True Geometry. As shown in the images below, the Headshot 3 generated head mesh well aligned with the original source photos.
A comparison of head shape accuracy between Headshot2, Headshot3, and other head generation model, using the same source image for automatic head generation.
As you can see, our self-trained model captures correct head contour, subtle facial structure, and accurate feature proportions, that generic models miss. Here’s another case:
And what about the side view or the back ? Why only from a single image ? There are so many 3d Model creater websites that create from multiple pictures from multiple angeles. Websites that create the whole body and not just the face. And how about multiple obj files comming from 3d scanners and not just Iphones ? If this would need more Vram why not make an OSX Version ? No Vram is cheaper than the Shared memory of Mac. And from M5 it´s even as fast as the 5080.
If you have been around for awhile at RL forums, you will know that that is not going to happen. It would no be simply porting; it would mean redeveloping for a completely different architecture.
EDIT: As this is probably something that is also asked of (small) game developers, there are tools available like: Game Porting Toolkit - Games - Apple Developer. I’m happy with Windows, so I don’t know if that is also a suitable tool for 3D programs like iClone or CC.
Hi! @HansRL are there any plans to improve any features for the mesh to head workflow with Headshot 3? i.e placing symmetrical markers or any quality of life improvements?
It´s about 1/3 slower than Nvidea 5090 about equal to RTX Pro line and faster than AMD. From the money a Mac Studio Ultra with 500 gb shared Memory is about the same as a single 96gb RTX Pro. And you need 5 of them inj order to get the same amount of Vram. And while it comes down to singe applications. I don´t know exactly what AI Realision uses. But Blender for example works on phyton. And python is on OSX same speed than windows.
Yes I´m sure. Benschmarks on all plattforms show the M5 Max already on the same speed as 5080. M5 Ultra isn´t out yet. While the data of the memory shows 25-33% slower memory speed than 5090. And the RTX Pro are also 25-33% slower while having more gb. And if you add the typical numbers for the Ultras in the past to the numbers of the M5 Max you have a speed slightly over the AMD Ryzen Threadripper PRO 9995WX (64 core). It´s a matter of programming and drivers. Nothing else. If you want to build something on a PC base compared to an Ultra with 512 GB(about $10.000). The Mac Studio Ultra M3 is already able to run the 500gb DeepSekk model local. For a PC (Xeon/Threadripper PRO + 5 Nvidea Pro 6000 Blackwell you spend more than $60.000. And 400w power vs. 2500w power. If you spend the $60.000 you have the faster machine. While the Mac is able to do the same just a bit slower for 1/6 of the price and 1/5 of the electrical power.
I think we should go back to the topic at hand as Mac vs PC debates are endless and rather pointless. Just out of curiousity: can you even upgarde a Mac, or is it like an iPhone. You go from M1 to M2 to M3 to M4 to M5 and then perhaps M6. That will add up too.
Unless the M5 is is significantly faster than the M4, it is indeed unlikely. I looked at a 3DMark benchmark for the same game using the fastest Mac for which there is data:
