All AI, the init image was made by Midjourney
@fofrai
-
Real-ESRGAN Video Upscaler Now Available on Replicate
By
–
For anyone looking for an open-source video upscaler, @lucataco93 has added the Real-ESRGAN Upscaler to Replicate: https://
replicate.com/lucataco/real-
esrgan-video
… Upscale to Full HD, 2k or 4k -
Video AI Tools: Topaz, FILM, and ST-MFNet Comparison
By
–
I used Topaz AI, but there's also FILM and ST-MFNet: https://
replicate.com/zsxkib/st-mfnet https://
replicate.com/zsxkib/film-fr
ame-interpolation-for-large-motion
… Both do a good job. -
High Resolution Video Generation: Coherency and Noise Parameter Optimization
By
–
On higher resolution videos I've found:
— fofr (@fofrAI) 26 novembre 2023
– coherency takes a hit and there's more warping
– you need lower noise values, like cond_aug 0.01 rather than 0.02
This one was generated at 1792 × 1024 and you can see the warping in the foreground. But I love the rotating planet.
2/2 pic.twitter.com/5gqTc8J8tkOn higher resolution videos I've found: – coherency takes a hit and there's more warping
– you need lower noise values, like cond_aug 0.01 rather than 0.02 This one was generated at 1792 × 1024 and you can see the warping in the foreground. But I love the rotating planet. 2/2 -
Creating HD Videos with SVD: Memory Optimization Techniques
By
–
If you've got the memory, you can make HD videos with SVD, using larger images.
— fofr (@fofrAI) 26 novembre 2023
This example is 1408 × 768 (bigger than 720p) straight out of SVD – no upscaling (but I did add interpolation).
Set the number of frames to decode at a time to 1 (decoding_t) to avoid OOM
1/2 pic.twitter.com/aeRXmlok5QIf you've got the memory, you can make HD videos with SVD, using larger images. This example is 1408 × 768 (bigger than 720p) straight out of SVD – no upscaling (but I did add interpolation). Set the number of frames to decode at a time to 1 (decoding_t) to avoid OOM 1/2
-
Stability AI Video Model Dimensions Technical Specification
By
–
But note that those numbers are `height x width` rather than the usual width x height. https://
github.com/Stability-AI/g
enerative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L117-L120
… -
SVD Excels at Vertical Video Generation
By
–
SVD does really well with vertical videos too. pic.twitter.com/M1iKrUCLx6
— fofr (@fofrAI) 26 novembre 2023SVD does really well with vertical videos too.
-
SVD_XT Base Model with Motion Bucket Configuration
By
–
No, this is just the svd_xt base model with motion_bucket_id set to 28. Using a Dalle3 init image. There's no loras or fine tunes of this model yet AFAIK.
-
Stable Video Diffusion by StabilityAI released
By
–
Stable Diffusion by StabilityAI, it's https://
github.com/Stability-AI/g
enerative-models
… -
Motion Bucket ID in Stable Video Diffusion Controls Camera Motion
By
–
`motion_bucket_id` with Stable Video Diffusion can be loosely approximated to camera motion, not subject motion.
— fofr (@fofrAI) 25 novembre 2023
Lower values can lead to better static shots with moving subjects. https://t.co/1PwUfo03tc`motion_bucket_id` with Stable Diffusion can be loosely approximated to camera motion, not subject motion. Lower values can lead to better static shots with moving subjects.