As per NVIDIA’s white paper of 2024 technologies, image to video ai can generate as much as 12 types of video outputs from a single image such as 3D rotational display with rotation rate being equal to 30° per second, dynamic scene expansion with elevated resolution being four times of original resolution, and macro motion simulation with accuracy levels ±0.2mm. For instance, Shopify merchants leverage Runway ML’s ai video creator to create 360-degree presentation videos out of the key product images. Production costs have been reduced from $500 a unit for standard 3D modeling to $0.8 a unit, and the conversion rate has improved by 23% (statistics in the 2024 Meta E-commerce Report). In television and movie production, Disney used the Pika Labs tool to inflate the concept art to 10-second dynamic storyboard, reducing the production time of a single shot from 72 hours to 19 minutes. A 15% perspective distortion error, however, needs to be adjusted manually (the case is referenced from the SIGGRAPH 2024 demo).
Physical simulations with high complexity have been made possible by advancements in technology. MIT Media Lab experiments show that in capturing a 12-megapixel still frame image of a flame, the image to video ai can generate a 5-second smooth simulation video (8 million smoke particles per frame), and the visual similarity to the actual shoot material is 92%. But the maximum power consumption by the GPU becomes 580W (the specification of NVIDIA RTX 6000 is 600W). In car commercials, BMW applies Stable Video Diffusion to mimic dynamic videos of vehicles driving through rainforests. The cost of a single video has been reduced from $470,000 in on-location shootings to $900, but the simulation error in tire water splash height is up to ±18cm (the industry standard of ±5cm). In the realm of scientific visualization, Harvard Medical School created 3D cell division videos (24 frames per second) by feeding pathological sections in a mere 8 minutes, 140 times faster than conventional microscopic photography (case in point, see the September 2024 issue of Nature Medicine).
Artistic boundaries keep on widening. The Sora V2 of OpenAI is able to convert Van Gogh’s “The Starry Night” into a 30-second dynamic canvas. The brush stroke rate is precise to 0.5mm/second, and the error in the change of colors ΔE is ≤2.3 (ΔE is ≤1.5 in case of animations done by expert painters manually). TikTok influencers’ actual tests show that after entering selfies, the ai video creator can create 20 different dance style videos (e.g., inter-frame Angle change of Hip-hop movements ±7°). The completion rate is 41% higher than static content (Social Media Today Q3 2024 analysis data). In restoration in history, Luma AI technology was used at the British Museum to restore broken sculptures’ photographs into complete spinning videos. The accuracy of geometric restoration was up to 89%, but simulation error on reflectivity rate remained up to 18% (the case is referred from ICOMOS 2024 Conference on Cultural Heritage).
Industrial applications at the production level redefine production processes. Tesla’s factory applies image to video ai to convert part inspection charts into fault simulation videos (thermal distribution variations at 60 frames per second), increasing the quality inspection efficiency 7 times, and reducing the misjudgment rate from 3.1% to 0.7% (statistical data might be included in the 2024 International Manufacturing Summit report). In aerospace, Boeing supplies engine blade images to produce a video of metal fatigue formation in a period of time (time compression ratio of 1:1000) with an early warning rate of 92% and 83% of cost savings compared to traditional X-ray detection (case source: Aviation Week, October 2024). Despite the dynamic blur accuracy limitations (currently up to 120fps) and ultra-wide light simulation (with only 10,000 nits of brightness support), Gartner predicts by 2025 that 65% of B2B organizations will use ai video generator to replace more than 30% of static visual content.