Drowning
Generative-AI (Video, Speech & Music generation)
Solo Project | Feb 2024
Tools: Stable Diffusion, Comfy UI, Flowframes, Premiere Pro, ElevenLabs, Mubert.
Inspired by T.S. Eliot’s The Love Song of J. Alfred Prufrock, this generative video explores alienation, desire, and the dehumanization of a fragmented world. Using human bio-structures, water, and metal, it portrays a haunting transformation from vibrant, organic life to a cold, metallic object drowning in indifference. Lifelike forms—pulsing and twisting with vitality—gradually merge with reflective metal, symbolizing the erosion of humanity and emotion.
Workflows
Image generation
Individual video generation
Scene transition
Artistic Approach
Using "The Love Song of J. Alfred Prufrock," by T.S. Eliot as the main inspiration of this generative video, there are several emotions and reflections I seek to convey:
Alienation and loneliness
Desire and doubt
A fragmented and chaotic world
Dehumanized emotions and indifferences
Therefore, I decided to use 3 elements: human bio-structures, water, and metal. By setting the storyboard as a process of the intergeation of human body and metal in a watery environment, I hope to decipt how a human gradually grows from a lively body with dynamic desires to a metal object drowning indifferently.
At the beginning, I generated a lot of organ-shaped objects with human skin color style. I also added movements such as pumping and twisting to show the livelihood. Still, the overall unsettling artistic style implies the latter transformation.
Then, I iterated these scenes with more metal elements. I do like how the reflection of metal texture accord with the original light reflection on bio-structures. This also helped me better transit the two themes more smoothly.
At last, I used solely metal image prompts to generate metal-made structure. This took longer time than I thought since there were nothing related to human in the image prompt in my first few generations. It was when I added a picture of eyeballs did the workflow produced what I wanted.
However, 2 things didn’t work out well:
Images and videos are often over-saturated even with negative text prompt and adjusted variables. Later, I learnt from Professor Stavros that we can add some nodes to de-saturize or change the original models/checkpoints to avoid outdated errors.
Using 2 workflows to generated individual scenes and video transitions led to the inconsistency between videos. The scene transitions seemed more blueish compared to other videos and I had to use Capcut to do some color-correction.
Overall, I am quite satisfied with the individual scenes and their video result. The workflow responded with reasonable motions according to different pictures and themes.