Out of curiousity, I decided to do a direct comparison of Stable Diffusion vs. Disco Diffusion on a music video animation I originally did in Disco Diffusion.
The original Disco Diffusion version took about 6 days to render on an nVidia RTX a6000 courtesy of GPU Cloud service RunPod.io. It was rendered at 768x768, cropped to a 16:9 aspect ratio and then uprezzed in Topaz Labs Gigapixel. I added a few post FX in After Effects like lens flares and glows.
The Stable Diffusion version uses pretty much the exact same prompts without editing. It was rendered in less than 24 hours on an nVidia Tesla T4 through Google CoLab at a resolution of 512x512 using the Euler sampler. I used Deforum's v.0.5 Stable Diffusion Notebook. (Link below)
Under "Coherence," it had 'diffusion_cadence' set to 5, so it essentially rendered every 5th frame and used MIDAS to interpolate the inbetween frames. It was uprezzed to 1024x1024 in Topaz Labs Gigapixel. The frame rate was increased from 24 fps to 120 fps using Topaz Labs Video Enhance AI.
Again, the prompts used in the SD version were virtually identical to the ones used in the original DD version. No editing of the SD version prompts were done to adapt to the SD "way of doing prompts."
------------------------------------------------
Need some extra GPU power? Check out RunPod.io's GPU Cloud service. Really great folks running it. Very helpful! I wouldn't have been able to do the original DD version of this video without them!! They have a wide range of GPUs available at really competitive prices. Check them out here:
[ Ссылка ]
The Stable Diffusion version of this animation was done using Deforum Stable Diffusion v.0.5. This Google CoLab Notebook can be found at this link:
[ Ссылка ]
