Stable WarpFusion v0.15 - alpha masked diffusion - Download (Patreon)

Sort of a disclaimer: only nvidia gpu with 8gb+ or hosted env

don't dive headfirst into a nightly build if you're planning to use it for your currect project, which is already past its deadline - you'll have a bad day. This is not a production-ready user-friendly software :D

Make sure you've tried awesome free solutions like Deforum or TemporalKit before subscribing and submitting yourself to the warp.

Hope you like the current development shpeed. You can't stop me now :D

New:

• add alpha masked diffusion
• add inverse alpha mask diffusion
• save settings to exif
• backup existing settings on resume run
• load settings from png exif
• add beep
• move consistency mask dilation to render settings
• hide edge width/dilation from generate flow cell

upd 8.6.23 v0.15.1:

• add pattern replacement (filtering) for prompt
• fix constant bitrate error during video export causing noise in high-res videos

v0.15.2:

• fix typo in cell 1.5

v0.15.3:

• fix full-screen consistency mask error (also bumped missed consistency dilation to 2)

12.06.2023: bumped to nightly.

v0.15.4:

• fix keep_audio not working for relative video path thanks to louis.jeck#2502

26.6.2023

• fix consistency error between identical frames

More about that bug here - https://discord.com/channels/973802253204996116/1122765734532087889

5.7.2023

• fix torchmetrics version thx to tomatoslasher

this one - ImportError: cannot import name '_compare_version' from 'torchmetrics.utilities.imports' (/usr/local/lib/python3.10/dist-packages/torchmetrics/utilities/imports.py)

13.08.2023:
Moved to M tier

Alpha masked diffusion

Same as masked diffusion, which was using consistency mask before, but this time it uses alpha mask. masked_diffusion has been renamed to cc_masked_diffusion.

It works this way:

if current diffusion step is before the masked_diffusion values, both masks are used, and their masked areas (black) are being diffused, unmasked (white) are being fixed.

if current diffusion step is between the masked_diffusion values, the rightmost mask (the one with the higher masked_diffusion value) is used.

if current diffusion step is above the masked_diffusion values, the whole frame is diffused.

Beep

Check Beep to beep. Useful to signal when the render is over.

Save & load settings update

Settings are now saved to frame exif data. You can load them from the frame by specifying a path to it instead of a txt file.

If you are changing your settings during resume run, it will backup existing settings.txt, which were lost before.

Pattern replacement (prompt filter)

You can now automatically replace certain word or phrases.

For example, you can have a lot of animated scene descriptions across numerous frames but want to test different styles. You can set a keyword {style} smewhere in your prompt and replace it with the whole new style without touching your prompts at all to iterate quickly.

It works like this:
go to GUI - Replace patterns and use the following notation:
{0: {"keyword or phrase": "replacement"}}

it can also be scheduled:
{0: {"keyword or phrase": "replacement"}, 10:{"another keyword or phrase": "another replacement"}}

This will replace "keyword or phrase" with "replacement" at frame 0 and "another keyword or phrase" with "another replacement" at frame 10.

You can use multiple filters per frame:

{0: {"keyword or phrase": "replacement", "another keyword or phrase": "another replacement"}


This is applied after captions, so you can use that to filter unnecessary words, for example "cat" if you already have that in your prompt, or want to replace "cat" with "dog" in your captions.