Explore your endless real-time interactive dreams
Imagine sitting back in your easy chair dictating your dreams your voice and using commands like zoom in/out, pan left/right/up/down and seeing the results instantly. Imagine being surprised when a famous person appears in your dream and you realize; Oh, I have the TV on in the background and they were just talking about that person. Imagine drawing into the video with noise in real-time and getting iteresting effects. Imagine merging another video stream, family picture, or an openpose animation into your dream as it unfolds and seeing the result immediately.
You might say: Oh, but the video is jittery. Of course, with other excellent video generators you can generate a very smooth video that takes 5(?) minutes to get perhaps 10 seconds of someone turning their head or taking a few steps. And, yes, the quality is high. However, EndlessDreams generates continuous variety for hours, days, weeks, ... as you interact with it in real-time. It is a different use case and those that have seen EndlessDreams think it has its niche in this world.
Jan 2022: Retired as a database performance architect at Microsoft.
Aug 2022: Learned of Stable Diffusion and was hooked.
Dec 2022: Bought hi-end system with Ubuntu, i9-13900K and a 4090.
Spent little time on "art?". Most of my time was spent pushing the performance of SD to the max. For me the question was: How many milliseconds do I need to generate an image? I was learning python, stable diffusion internals, and optimization techniques making continuous progress.
Oct 2023: Simian Luo announces [LCM].(https://github.com/luosiallen/latent-consistency-model)
Prior to Oct the dreams of real-time AI video generations were just a fantasy. With LCM I saw that the time was NOW. Within hours of LCM dropping I had what might have been the first jittery real-time video generations running on my 4090 and I posted examples to reddit and elsewhere. By real-time I don't mean the Youtube demoes starting to appear in Nov/Dec 2023 claiming real-time SD videos at 3 to 5 fps. I started out of the gate at 10fps and quickly got to 15 fps with SD1.5. Then 33 fps for 800x800 SDXL, 22 fps 1024x1024, 17 fps 1280x1024, and I continued to this day to make things faster.
Oct 30, 2023: Deepfakes had been around for awhile. But not real-time generated DeepFakes. Here is a poorly lit crude early demo of mine: RT DF Demo
Nov 5, 2023: Coined the term RTSD and demoed the ability to, in real-time, generate images on the fly as one changed the prompt or other paramaters using sliders. Instant feedback vs making a change, clicking generate and then waiting.
Soon afterwards all that I've learn about performance and other ideas coalesced in the multimodal experience of EndlessDreams. It has been a bumpy ride.
My Twitter where you can see some older demoes: https://x.com/Dan50412374/
My YouTube channel with my first try at a YouTube demo: https://www.youtube.com/watch?v=irUpybVgdDY
Enjoy EndlessDreams. It is my fault and my fault alone for sitting on this for so long. It was a breakthrough when I first created it and all I ask is for people to give it a look.
There are a number of reasons I'm starting with "no license". I'll only mentions this one: I gave a screen share demo of this to a guy that does Hackathons. He recorded my demo, and having not written a line of code of this, he posted it on Linked-in as one of his projects. I got mentioned as a member of his team. Egads.