Realistic AI Humans...

[Rough1]

Hello everyone,
I work for a company that produces medical devices for physically handicapped people. We have the problem that it is difficult from a marketing perspective to photograph actual patients in the devices. I would like to generate scenes with C4D of the device, and have some sort of solution to insert a realistic AI generated person into the picture. The boss is not happy with the results from Renderpeople for example. He wants it absolutely photorealistic, as far as it gets...
I have messed around with ChatGPT, but the results are not acceptable. The generated person is NEVER in exactly the correct position, although I give it an exact example of what I want with a C4D puppet model.
Does anyone have any ideas how to approach this?
Thx in advance
Simon

[Dr. Sassi]

Hi Simon,

The problem is, whatever we used to see our whole life will make us the most critical viewers.

To recreate human appearances is not just about doing some things for the model; the light interaction is surely a major part of it. That goes deep, literally. That you try to prompt it indicates to me that you don’t have a lot of development time.

Being aware of the problems, it is not an easy task, for a still being already a challenge, but for an animation, which is extremely difficult, if you like “photorealism”. I placed this before my suggestions, as I am aware, there is no quick fix to do that, if one sets the bar to “photorealism”.

How about https://www.metahuman.com/en-US
Depending on the licensing that applies, it might be a solution or a good start.
https://www.metahuman.com/en-US/license
This is a tip, and I can’t give support for this.

My best wishes for your project.

[Rough1]

 Thanks for the reply Doctor Sassi. He hasn't talked about animating yet, he's focused on Pictures. Its frustrating, because what I would think should be relatively simple, seems not to be. Here is my imagined workflow... I create a scene with one of our products, and a rigged puppet taking the place of the patient. Fix a camera. Make a picture with the a render tag removing the puppet, but keeping the shadows its created etc... Make another picture of just the puppet, keeping the light environment. Then I want to give AI the puppet picture, and prompt that AI should create a photorealistic human in exactly the same pose, perspective and lighting as the example I supplied, and then I can composit the 2 pics in photoshop.
... In theory...
Thx again
Simon
EDIT: I tried to upload a pic as an example, but even at 500x300 pixels, its telling me the dimensions are too big!!?

[Dr. Sassi]

Hi Simon,

The picture should be fine, 500 pixels vertical is the max, and wide can be larger.

I can't tell how to write a prompt in AI, nor is that the forum to do so. I can provide some tips. But since you use AI, ask what the qualities of Skin appearance are. To create perhaps a communication base with the AI entity. AI will not know the difference in images of retouched toon skin, perhaps, if not defined what qualities you are looking for.

Perhaps find an AI engine that allows you to use reference images, and then mark the area where you would like to have that quality.

Eventually use a subsurface scattering (SSS) material and get the lighting already close to the final, before asking AI to take over.
Test as well the Skin Material of Redshift:
https://help.maxon.net/c4d/2026/en-us/Default.htm#html/COM_REDSHIFT3D_REDSHIFT4C4D_NODES_CORE_SKIN.html

Skin is typically the main challenge as the light goes through several layers. Even small emotions change the blood flow, and tiny changes in color might occur. We are all reading this, more or less conscious.
When it is missing, it often works like with shadows; most people can tell it looks like CG, but what goes wrong is often not clear. Like, often you might notice in CG comps that the light wrap is not even established.

The list of Skin qualities or influences is long and not limited to the ones in the list:

• Subsurface Scattering (SSS)
• Detailed Texturing (Albedo and Displacement):
• Displacement/Normal Maps
• Layered Shading Model
• Dynamic Changes
• Anatomical Accuracy
• Lighting

More about light: Think also of the effect of small hairs catching light. Using only "white" light is typically a mistake; bouncing light is typically colorful. Abstract painters have exaggerated this extremely. Cinematographers sometimes use even a color-full bounce board.
Lots of Reference and Observation (Avoid web-based stuff, shoot your own with a gray card)

If nothing helps, add a slightly smaller model (Normal Shift) and create the "under the Skin" results, then blur them a little bit and comp those to a small percentage into the image. (Fake SSS, but only as extra, not replacement of SSS)

All the best