Stable WarpFusion v0.5.12 (Patreon)

A cumulative announcement post for v0.5.12 and v0.5.11
Now getting even closer to some stable Stable Warp version thanks to your support and feedback!
Here's the changelog:

New stuff

Added VRAM usage fix, now supports up to 1536x1536 images on 16gig GPUs
Added VRAM usage fix from this awesome repo - https://github.com/Doggettx/stable-diffusion

It allows running output sizes up to 1536x1536 (or any with the same pixel area) @ 16Gigs of VRAM. Turnin on any loss scales will negate this effect, as computing gradients for loss functions is still memory intensive. 

Added frame range (start-end frame) for video inits
Animation Settings -> Video Input Settings ->
start_frame
end_frame

You can now select a frame range to extract from your videos. 0 end frame means it will use the video's end frame i.e. use the whole video after the start_frame. 

Added pseudo-inpainting by diffusing only inconsistent areas
stable-settings -> mask_result

This setting applies diffusion only to areas deemed inconsistent by the consistency checking script, the other part of the frame is kept intact with respect to flow_blend, and is still warped around

Added Create a video from the init image
Animation Settings ->
use_looped_init_image - enable feature
video_duration_sec -> set video duration (1 sec = ~24 frames)

You can now create a video from your init image. It will be just a looped still image. Useful for testing settings or animating stills and backgrounds.

Added Fixed start code toggle \ blend setting
stable-settings ->
fixed_code - enables the feature
blend_code  - add some random noise back to the fixed noise vector
normalize_code - restores the initial noise range after blending the fixed noise with random (it may drop in range significantly resulting in a blurry image)

Fixed code is analogous to the fixed seed (video seed continuity) setting in DiscoDiffusion, but it only fixes the starting noise added to the init image, so it doesn't have that "behind a dusty window" effect compared to fixing global seed.

Added Latent frame scale
Animation Settings -> Video Input Settings -> Coherency Settings ->
frames_scale

This is a latent init scale loss for frames, similar to the init latent scale

Fixes:
- Fixed changing width-height not working correctly
- Removed LAMA inpainting to reduce load and installation bugs
- Fixed multiple image operations being applied during intermediate previews (even though the previews were not shown)
- Moved Stable model loading to a later stage to allow processings optical flow for larger frame sizes
- Fixed VideoFrames being saved correctly without google drive\locally
- Fixed PIL module error for Colab to work without restarting
- Fixed prompt scheduling
- Returned init scale \ frames scale back to its original menus
- Hidden all unused settings

You can grab v0.5.12 here - https://www.patreon.com/posts/stable-v0-5-12-71655145

P.S. Added settings for that video from twitter here as well