🏔 Wide-angle lenses (10mm to 35mm)
🕺 Standard lenses (35mm to 85mm)
👀 Telephoto lenses (85mm to 300mm)
They often correspond to these shots:
Wide angle (wide-angle)
Medium to medium close-up (standard)
Close-up (tele)
Prompting for "close-up", getting medium close-ups ... 😊
We wanted tele, but are probably a bit off toward standard. However, given the depth of field and compression, closer to 80mm than to 50mm. Fair enough.
We just go one step further and ask for an "extreme close-up", ending up with a close-up shot.
If we wanted to go further, we could use different variants of repetition to get to actual extreme close-ups:
Medium shot. I'd say right on target.
Wide angle. Again, a bit too narrow, probably still in the standard lens range?
Enforcing with repetition and we get a more distorted wide-angle shot.
It's probably also possible to reproduce these effects in studio settings with negative prompting and weights. I will look into this here: medium.com/@tristwolff
For more shot sizes, camera types, prompting for props and other cinematic prompt gimmicks, check my three-part tutorial on "Cinematography with #Midjourney" medium.com/design-bootcam…
If you want to follow these explorations into AI & creativity, follow me: @tristwolff
💰💻💡 The industry is eager to save money & resources /w AI efficiency + #GPT4 is leading the charge. I expect production processes to be changed within a year.
The infamous #OpenAI study has shown: Creative tasks are the most exposed to AI automation
The consequences of AI are real, and they're coming to town. Now, how can AI improve the creative process? By helping with structural groundwork, for example: medium.com/design-bootcam… #AI#storytelling
The model is simply called “text-to-video synthesis”. A brief summary:
- 1.7 billion parameters 🔥
- training data includes public datasets like LAION5B (5,85 billion image-text pairs), ImageNet (14 million images) & WebVid (10 million video-caption pairs) 🌎
- open source 💪
Text-to-video synthesis consists of three sub-networks that work together to produce short MP4 video clips:
- a text feature extraction,
- a text feature-to-video diffusion model,
- and a video-to-video diffusion model.
However, 3.6:1 (and higher ratios) seems to work better if you drop the cinematic prefixes (cinematic shot, film still, etc.) 🤷♂️🤔
Here it's just scene & style description. 3.6:1, no letterboxing.
despite the letterboxing, exploring 4:1 is fun ... 🍄🎉 #MidjourneyV5
I generated the images for the game with #midjourney
Prompted for basic 1:1 images, using the tried and tested combination of "white background" and "--no background" to prep transparent PNGs
Then I used #GPT4 via #ChatGPT & gave it the basic story for the game: Squid-Spaceship having to collect little baby squid astronauts, etc...
Asked the AI to come up with HTML/JS code & CSS, then helped it to debug & adjust mechanics (smooth flight paths, etc..)