3D2I: Create Stunning AI-generated Images and Product Concepts from 3D Models on AWS with Amazon?Bedrock

Learn to quickly create AI-generated product concepts from 3D models using TurboSquid, Blender, Amazon Bedrock, and the Amazon Titan Image Generator foundation model.

In this post, we will learn how to turn 3D models from TurboSquid by Shutterstock into AI-generated images and product concepts using a combination of Blender , Amazon Web Services (AWS) Bedrock, and the Amazon Titan Image Generator foundation model.

Most current image generation foundation models support text-to-image (T2I) and image-to-image (I2I), also called image variations. However, creating just the right image with Generative AI, even with a combination of finely crafted text prompts and existing images, is still tricky. The reference images you start with often don’t have the proper layout, lighting, camera angle, or overall composition. To increase our success rate at generating better images, we can turn to 3D2I?—?using 3D models and high-quality 3D renderings as the basis for our image-to-image generation.

Using leading 3D software, such as Autodesk 3ds Max , Autodesk Maya , SketchUp , Maxon Cinema 4D , or our choice for this post, Blender , we can control most compositional aspects of our source image, such as the position and scale of objects in the scene, textures, lighting, camera position, and depth of field. Once our composition is perfect, we can render high-quality images for image-to-image generation using the Amazon Titan Image Generator foundation model.

TurboSquid

According to their website, TurboSquid , acquired by Shutterstock in 2021, is “a platform that makes it easier than ever to buy or sell stock 3D assets. TurboSquid features more than a million stock assets, which are all available to download and use in your own projects at the click of a button.” TurboSquid’s 3D models are used by game developers, news agencies, architects, visual effects studios, advertisers, and creative professionals worldwide. Their 3D models range in price from free for simpler, lower-quality models to thousands of dollars for larger, more complex models. We will use free 3D models compatible with Blender for this blog post.

Blender

Their website states Blender is “the free and open source 3D creation suite. It supports the entirety of the 3D pipeline?—?modeling, rigging, animation, simulation, rendering, compositing and motion tracking, even video editing and game creation.” Blender is cross-platform and runs equally well on Linux, Windows, and Macintosh computers. The interface uses OpenGL to provide a consistent experience. For this post, I am running Blender on my MacBook Pro and an Amazon WorkSpaces GraphicsPro.g4dn graphics bundle VDI environment, running on an NVIDIA T4 Tensor Core GPU.

Amazon Titan Image Generator Foundation Model

According to AWS , the Amazon Titan Image Generator is “an image generation model. It generates images from text, and allows users to upload and edit an existing image. This model can generate images from natural language text and can also be used to edit or generate variations for an existing or a generated image. Users can edit an image with a text prompt (without a mask) or parts of an image with an image mask. You can extend the boundaries of an image with outpainting, and fill in an image with inpainting. It can also generate variations of an image based on an optional text prompt.” For this post, we will generate variations of our product renderings.

AI-generated Images and Product?Concepts

For this post, let’s assume we’re an upscale furniture company that has designed a trendy solid wood and metal chair. Using the existing CAD drawings of the chair, we will use Generative AI to help our product designers and marketing team develop new concepts for our successful chair design, including more environmentally friendly options and ones that are affordable to a broader customer base.

According to Imaginationeering , “Product design conceptualization is an important step during the early design phase of a new product development, or during an upgrade of an existing product.”

Using Blender, we can precisely position the 3D elements within the scene. We can also duplicate elements, like the chair, to create a more natural room arrangement. With Blender, we can create and apply new materials to the objects, such as changing the chair seat from wood to plastic, leather, or fabric. We can also change the dimensions of objects, such as making the chair’s seat larger, the side table’s legs shorter, or the plant larger. Lastly, with Blender, we can tightly control the camera position to compose shots and lighting to set the mood. These options give us tighter control over the reference image used to generate image variations with Amazon Bedrock.

Once we create a composition we are happy with, we generate a high-quality rendered view and save the rendering as a JPEG image.

Using the JPEG of the rendered view as a reference image, we generate variations in the Amazon Bedrock Image Playground, CLI, or SDK using the Amazon Titan Image Generator foundation model. Along with the rendering, we include a well-engineered positive and negative text prompt to guide the model’s image generation and the reference image. The prompts below result from multiple tests to optimize the resulting image variations based on the prompts.

“smooth light blue plastic chair, single large aloe plant on side table in bright yellow pot, highly-textured medium brown wood flooring, bright natural light, shadows on the floor from chair and table, highly-detailed realistic 3D rendering”?—?positive prompt example 1

“vintage distressed red leather chair with baseball glove style stitching, large barrel cactus in light blue flower pot on side table, deep-grained medium brown wood floor, bright natural light, dramatic shadows”?—?positive prompt example 2

“light orange, course-grained fabric, cloth chair with thick wood legs, cactus on side table in a light blue pot, highly-textured light brown wood floor, realistic, hyper-detailed, bright natural light, shadows on the floor from chair and table”?—?positive prompt example 3

“floating chair seat, deformed table legs, deformed chair legs, asymmetric table legs, asymmetric chair legs, tight cropping, dark shadows, low-res, low-quality, JPEG artifacts, humans, animals, people, wording, text, letters, symbols”?—?common negative prompt

Using a variety of renderings as reference images, positive and negative prompts, and inference parameters, we can quickly generate numerous concepts for new chair designs using materials such as leather, fabric, plastic, metal, and wood.

Conclusion

In this post, we discovered how to quickly develop multiple product concepts using 3D model renderings as reference images with Generative AI services like Amazon Bedrock and the Amazon Titan Image Generator foundation model. Stay tuned for my upcoming accompanying video if you want to learn more about turning 3D models into AI-generated images. In the meantime, I recommend trying some 3D models from TurboSquid by Shutterstock and checking out Olivio Sarikas , who introduced me to this method. His YouTube video, 3D to AI?—?THIS is the REAL Power of AI , is a great place to start.

This blog represents my viewpoints and not those of my employer, Amazon Web Services (AWS). All product names, images, logos, and brands are the property of their respective owners.