From Text to Hologram: Creation of High-Quality Holographic Stereograms Using Artificial Intelligence

Abstract

This study simplified the creation of holographic stereograms using AI-generated prompts, overcoming the conventional need for complex equipment and professional software. AI enabled the generation of detailed perspective images suitable for various content styles. The generated images were interpolated, upscaled, and printed using a CHIMERA holoprinter to obtain high-quality holograms. This method significantly reduces the required time and expertise, thereby making holographic content creation accessible. This approach demonstrated that AI can effectively streamline the production of high-fidelity holograms, suggesting exciting future advancements in holographic technology.

1. Introduction

Holographic stereograms, introduced in 1995 by Yamaguchi et al. [1], have since been used successfully in various fields, including art, advertising, the military, architecture, the automotive industry, medicine, and entertainment [2].

Holographic stereograms, which involve capturing a vast series of perspective images of real or virtual 3D models using a camera [3], present several challenges. In the case of a manipulable physical 3D model of a reasonable size (below 80 cm), images can be directly captured in a laboratory but require specific equipment such as cameras, curved rails, turntables, or specialized scanners [4]. Large physical outdoor models require scanning technologies, such as photogrammetry [5], Gaussian splatting [6], and light detection and ranging (LiDAR) [7], which in turn require time, professional tools (cameras, drones), and technical expertise. The creation of virtual models requires not only technical skills but also artistic talent and often expensive professional software (Autodesk 3ds Max 2025, Autodesk Maya 2025) for 3D rendering with a powerful computer. The time required to create a complex 3D model could take 48 h or more, possibly extending to several weeks of work based on the individual artist’s workflow and efficiency. The rendering time for perspective images depends on the computing power of the computers available and the complexity of the textures and lighting in the 3D scene, and can take from several hours to several days. This makes it inaccessible for non-specialists and individuals with limited budgets. The high barrier to entry prevents many interested individuals and small organizations from exploring holographic production.

The advent of artificial intelligence (AI) [8] has revolutionized content creation, enabling the generation of text [9,10], images [11,12], and videos [13,14] from simple text prompts [15,16] without requiring prior artistic or technical skills. Recent research shows that holographic technologies are also beginning to benefit from advances in AI [17,18]. Notably, Gentet et al. [19] reduced the number of initial perspective images to seven for a 120° view and interpolated the missing images using neural networks [20].

This study aimed to further simplify the holographic creation process by generating all the necessary images from a single AI-managed prompt without the need for a 3D model. This new approach facilitates the rapid and straightforward production of high-quality holograms, making this technology accessible to a wider audience. The resulting full-color holographic stereograms can be printed using a CHIMERA holoprinter [21], which employs low-power, continuous RGB lasers, and a silver-halide material Ultimate U04 [22] to achieve a hogel resolution of 250 μm and a parallax of 120°.

2. Materials and Methods

2.1. Holographic Plates for CHIMERA Recordings

CHIMERA holograms were recorded on a silver halide holographic Ultimate U04 glass plate (Ultimate Holography, Bordeaux, France). This material was specially designed to record full-color holograms without any diffusion and was set to be isopanchromatic for all visible wavelengths. The plates were developed in two safe and easy-to-use chemical baths.

2.2. CHIMERA Recordings

Holographic stereogram recordings require the prior acquisition of a series of perspective images of a 3D scene. Half-parallax CHIMERA requires a series of 768 horizontal perspective images on a 120° circular arc, as shown in Figure 1. The axis of rotation is the vertical axis of the final CHIMERA, and everything in front of this axis appears to float in front of the hologram plane.

2.3. Initial Image Generation Using Diffusion AI

The initial image is generated using a stable diffusion model [23]. Advanced AI utilizes deep-learning techniques to create detailed images based on text prompts, thereby facilitating highly customized visual content without the need for manual graphic design. This process involves inputting a descriptive text prompt into an AI system that interprets and translates these descriptions into complex visual representations. This model leverages a trained neural network that learns from a vast dataset of images and their associated text descriptions. It begins by encoding the text and transforming descriptive prompts into a numerical format that guides the initial generation of noisy images. This image is then refined through a decoding process, in which the neural network iteratively improves the clarity and details of the image, effectively reducing noise and enhancing features until a final coherent image is produced. Figure 2 shows the process flow.

2.4. Workflow for Perspective Image Generation

The custom in-house workflow employs a diffusion-based model to transform the initial image input into an expanded set of 768 high-resolution images, as shown in Figure 3.

An advanced, custom-made AI-driven system was employed to generate seven distinct perspective views from a single initial image, each shifted by 20° to cover the 120° field. The resulting images have a resolution of 512 × 512 pixels. The choice of 120° was made to ensure compatibility and optimal display quality with the 120° CHIMERA printer. The key-point component of this process is a consistency mechanism within AI that ensures that no critical visual information is lost during the generation of these perspectives. To ensure consistency, frozen and trainable modules were used. The frozen module contained parameters that were not updated during the process, and the trainable module contained parameters that can be updated. Initially, the input undergoes a diffusion process that incorporates text and image cross-attention mechanisms within a stable, pretrained model to ensure fidelity and consistency in the image characteristics.

To enhance this consistency, the workflow integrates the Ip-adapter [24], which ensures uniformity across all the generated images, maintaining consistent style and characteristics. Moreover, ControlNet [25] was used to manage the positioning, specifically controlling the generation of images across the required 120° field. This targeted control is crucial for accurately maintaining the spatial orientation and distribution of images and ensuring that each perspective is correctly aligned.

This is followed by refinement using a U-NET module, which enhances visual details and clarity. Detailed enhancement is achieved using a variational autoencoder (VAE) that compresses and subsequently reconstructs the image and optimizes it for detailed richness.

The interpolation of the initial set of seven perspective images generated was expanded to 768 to cover the required 120° field. This process uses advanced neural network techniques to refine the interpolation, ensuring smooth transitions between each perspective and maintaining image continuity and quality.

The resolution required for a 30 cm × 30 cm CHIMERA hologram was 1320 × 1320 pixels. To achieve this, 768 images were upscaled. A custom ESRGAN x4 [26] advanced algorithm was employed to carefully enlarge these images to the required size while ensuring that the integrity and quality of the original visual content were maintained. This algorithm is specifically designed to prevent any loss of detail or the introduction of unwanted artifacts, such as blurring or pixelation, during the upscaling process.

2.5. CHIMERA Printing

The in-house software generated all the hogels from the perspective images. Each hogel was recorded sequentially using an RGB display system composed of three spatial light modulators (SLMs) and a 120° full-color optical printing head. After interference with the reference beam, the information corresponding to each RGB hogel was recorded on a U04 plate. The hogel size was 250 µm, and the printing rate was 60 Hz. The CHIMERA holoprinter uses three RGB DPSS 20 mW lasers with wavelengths of 640, 532, and 457 nm. CHIMERA can print holograms ranging from 10 cm × 13 cm to 60 cm × 80 cm.

2.6. Computer Specifications

A high-performance PC is essential for the efficient generation and processing of multiple high-resolution images. The computer used for this research had the following specifications: a CPU, AMD Ryzen 9, to efficiently handle complex computations; a GPU, NVIDIA RTX 3090, with 24 GB VRAM, which is crucial for graphics-intensive tasks; a 32 GB RAM to ensure a smooth handling of large datasets; and a 2TB SSD to facilitate rapid data access and storage.

2.7. Evaluation of the Accuracy of the Reconstruction

To assess the accuracy of the images reconstructed by the AI, the structural similarity index measure (SSIM) [27] is employed. The SSIM is a complex metric designed to quantify the visual similarity between two images by considering attributes that significantly impact the perceived quality, such as texture, contrast, and luminance. Unlike basic metrics that only measure pixel-to-pixel differences, the SSIM provides a more comprehensive analysis by assessing how changes in these attributes affect the viewer’s perception of image quality. For a precise evaluation, real perspective images were generated using a 3D model in Blender, which served as a benchmark for comparison. These Blender-created images were then compared with the AI-generated images using the SSIM, allowing for a direct assessment of the accuracy of the AI technology in reproducing detailed and realistic perspectives. In terms of the image quality, the SSIM values can be interpreted, as shown in

Table 1. Interpretation of SSIM values.

SSIM Value	Quality	Description
SSIM > 90%	Very good quality	The differences are minimal and often imperceptible.
80% < SSIM < 90%	Good quality	The differences are slight and may be perceptible but are not disturbing.
70% < SSIM < 80%	Average quality	The differences are noticeable and can affect the visual experience.
60% < SSIM < 70%	Fair quality	The differences are clearly visible and can be disturbing.
SSIM < 60%	Poor quality	The differences are significant, and the image is often unacceptable.

.

3. Results

3.1. Testing the Viability of AI-Generated Perspective Views

The effectiveness of the model in generating perspective views was tested by creating seven views using a 3D Blender model of a robot spanning 120° at intervals of 20°. This setup mirrored the conditions for AI-generated images. The central image from this set was then used by the AI system to autonomously generate similar perspective views. These AI-generated images were compared with the Blender images using the SSIM, as shown in Figure 4.

The SSIM comparison revealed that the AI could replicate the Blender views with good-quality fidelity (the average SSIM was 85%). This is an essential validation step, demonstrating that AI can effectively understand and recreate the spatial and textural nuances required for realistic holographic imaging. However, it is important to note that, while AI demonstrates a significant capability in replicating views, it inherently interprets side views by simulating what might exist beyond the observable scene, much like a human might imagine unseen aspects. For example, AI has interpreted that the robot’s antennae are located at the top of the head rather than at the back and has therefore generated images accordingly.

3.2. Creation of the Initial Image

An initial image with 1024 × 1024 pixels was generated within a few seconds using a stable diffusion AI model, as shown in Figure 5. To achieve this, the AI is provided with the following prompt: “A whimsical, full-bodied cartoon monster that exudes charm and vibrant colors. This delightful creature stands in a friendly posture, showcasing its unique and humorous features: large, expressive eyes; a wide smile with quirky teeth; and a mix of textures across its body, including fluffy fur, smooth scales, and soft spots. Its multicolored body ranges through a rainbow palette, with each limb displaying a different hue, enhancing its playful appearance”. This prompted the stable diffusion model to generate a high-resolution image that captured the details. This image closely resembles a computer-generated image (CGI) generated by a 3D artist using a modeling software.

After the generation of the CGI-like monster, four images were generated with other styles of images, colors, and schemes using other prompts: a Japanese woodblock print (Figure 6a), a comic-book red mage (Figure 6b), a pile of realistic fruits (Figure 6c), and a large dream-like elven scene (Figure 6d).

3.3. Perspective Image Generation and Detailed Upscaling Process

Seven perspective images of 512 × 512 pixels, as shown in Figure 7, were generated for the monster using the AI workflow within 2 min. Subsequently, the interpolation process expanded these images to a complete set of 768 images, covering the full 120° arc, which was completed within 3 min. The final set of perspective images shows a smooth reconstruction of the monster and its shadow, as shown in Video S1, but presents some small, unnatural deformations in the hands. Perspective image generation was repeated for the four other models with similar results.

However, the subsequent upscaling process, which is essential for preparing the images for high-resolution holographic displays, proved to be more time-consuming, requiring approximately 1 h for the 768 images. The upscaled images were adjusted to meet the specific resolution of 1320 × 1320 pixels needed to print a 30 cm × 30 cm CHIMERA hologram with a hogel resolution of 250 μm.

3.4. Final Holograms

Hogels were generated from these perspective images and recorded in 6 h with a resolution of 250 µm one after another into a 30 cm × 30 cm U04 holographic plate. The monster hologram was then developed using two chemical baths and sealed with an optical glue to prevent variations in the emulsion thickness. When CHIMERA was illuminated at an angle of 45° with an RGB LED lamp, a fine full-colored, half-parallax, pop-out, and ultra-realistic reconstruction of the monster was generated, as shown in Figure 8a and Video S2. The four other images were printed together on another 30 cm × 30 cm U04 plate in the same manner and presented the same ultra-realistic reconstruction, as shown in Figure 8b and Video S3. In all the cases, the plane of the hologram was level with the center of the recorded object. The observations showed that the unnatural deformations present in some interpolated images were not visible in the final holograms.

4. Discussion

This study demonstrated that it is possible to generate high-quality half-parallax holographic stereograms from a simple idea using a method based on AI-generated prompts. This new approach successfully simplifies the creation process, allowing for the generation of original holographic content in 5 min, as opposed to the conventional lengthy and skill-intensive methods required for 3D rendering using specialized software or an actual object with precise scanning equipment.

The results showed that the AI reconstruction of the perspective images works for all styles of content (CGI, comic, painting, realistic, etc.), all object sizes, scenes with single or multiple objects, and even very complex scenes with a detailed background (such as the “large dream-like elven scene”). It is important to note that while the AI-generated perspective views may present certain defects, these imperfections are concealed when the images are combined during the hologram printing process, rendering the defects invisible.

The images were upscaled to create 30 cm × 30 cm holograms. However, this scale can be increased to allow the printing of holograms of any size up to 60 cm × 80 cm, which is the current limit of the CHIMERA printer. The reconstruction was set to 120° to match that of the CHIMERA printer; it is technically possible to extend the total reconstruction to 360° using the same process.

The method has certain limitations. It is unpredictable. The operator has no control over the perspective image generation, which can vary randomly for the same model. However, because the process is quick and requires only a few minutes to complete, if it fails, it can be repeated until a satisfactory result is achieved. Moreover, the axis of rotation is always located at the center of the object and cannot be modified, which means that the final hologram will always have a “pop-out” effect.

This study was conducted using AI-generated images; however, the proposed method can also work with real images such as photographs, paintings, and drawings. The SSIM results for the robot showed a good quality reconstruction. However, it was not perfectly accurate in reality.

Therefore, holograms created from real images may not be acceptable for certain projects, as they can be interpreted differently by AI. This is particularly true for portraits, in which resemblance is a crucial factor. The micro-details of faces, such as subtle expressions and fine lines, may not be accurately reconstructed, leading to a lack of authenticity in the generated images. The specific issue of accurate portrait reconstruction will be the subject of a future study.

The generated holograms only have horizontal parallax, which is sufficiently large in most cases because novice observers generally do not notice the absence of a vertical parallax. This study can be extended to generate missing vertical views, which will be the subject of future research.

In figurative use, the term “chimera” refers to a dream, an unreal or unrealizable object. Thus, it can be said that CHIMERA holograms have been aptly named.

5. Conclusions

This groundbreaking method of creating holographic images using AI technology can significantly improve the quality of CHIMERA production. By leveraging AI algorithms to convert text into intricate holograms, the need for time-consuming 3D modeling is eliminated, making content creation more streamlined and user-friendly. This innovative approach aligns with the progress in image generation through neural networks, paving the way for exciting advancements in holographic technologies.