Edit: I've tried both lora's recently, and I will argue that https://civarchive.com/models/2088559/cinematic-hard-cut is better, for fast cutting, too. Especially for maintaining likeness. I will integrate these training clips in the next version of Hard cut and hopefully improve it even more.
"Fast cutting" is apparently the correct term: https://en.wikipedia.org/wiki/Fast_cutting
This lora is trained on "quick cuts", an editing technique that tells the story of a whole scene in a couple of seconds. I figured it would be suitable for the constrained context window local video producers have to work with.
Consider it experimental, as the dataset is quite limited at the moment. It may collapse completely on unseen data.
It's trained on shot concepts like "wide-angle shot", "mid-shot", "close-up shot" and (which is often used for quick cuts, "extreme close-up shot".
The format is:
A series of quick cuts:
[shot one]
[shot two]
[shot three]
...
Each cut has (about) one sentence description. You may specify angle, too.
It's trained on between 3 and 5 shots, over a very short time. Going for the full 81 frames might make it lose strength.
Tested both with T2V and I2V (but trained on I2V).
Only high noise required.
Description
Larger dataset, slightly larger resultion. Less epochs.
FAQ
Comments (19)
Can it only really show variations of the original image or can it add more creatively? Like if the original is a close-up can the next cut show his entire body? Or show him in a different setting? Or with different clothes?
I haven't really tried wide usage, yet. What I can say is that it's trained on data that are recent in temporal space, usually the same scene, but often with different perspectives. It can change perspective, going to wide-angle has proven to give varying results. Sometimes it works, sometimes it fails. If you look at examples of the technique in movies, it's usually "in scene".
With the "hard cut" lora, I've been able to make scene changes, sometimes.
It's pretty powerful. I've tested for science going from a portrait, to a closeup bj, and ending with missionary sex ... it just works if you use something like the DR34ML4Y all-in-one loras that know multiple concepts.
Wan 2.2 knows various cuts natively to some extent and I've used them with decent success
Hard cuts are mainly used to express the current action timeline in the current scene
Jump cuts does the same as hard cuts but with brief jumps in time usually to maintain suspense, it can also be used to change scenes
Match Cut changes scenes while matching something in scene between the 2 transitions
Smash Cut an abrupt scene change that has no context to previous scene
If your having trouble getting these cuts to trigger you might have to increase your high noise step count since these concepts are only triggered by the high noise model.
@AnonBlah Interesting thank-you.
I did try a test and what I got was a result that looked like it was really underbaked/lacking in steps which made sense to me since it basically had to render multiple scenes and not just one scene with gradual changes in each frame. I typically run a high ksampler with 6 steps (out of 24 steps total so 25%) with no lightx lora, and then 6 steps low ksampler (out of 8 steps total so 75%) with a lightx lora. Wonder if I need more steps both high and low.
@AnonBlah So would you say this lora isn't necessary?
@robbiebunny1883 This lora helps with executing fast consecutive hard cuts for high impact story telling so yes download it if you want I have which sent me down the rabbit hole discovering the other cut types, I also want to note its really hard to force fast camera movements/cuts in my recent trials wan likes to have about a ~12 frame delay before executing them natively and i had to do a bunch of stuff to force it to do them as early as frame 2. I only posted my list of cuts since different cuts are useful for different purposes.
@ptangwai700 It sounds like a very interesting workflow but i would try a more traditional workflow since from my understanding i think your leveraging the noise wrong for what the models want. An abrupt scenery change shouldn't really matter for the wan and should actually help the transition
So i've been playing around with your Hard Cuts and this Fast Cuts lora all of today. Very nice! Getting a good variety of results from each. Thanks!
how do you train this sort of thing do you just use 1 video?
This version uses 16 short clips, that demonstrate the effect, and hand written captions.
i'm not getting quick cuts at all. basically just getting the camera panning. strength at 1
Are you using the prompt format "A series of quick cuts"? And on high noise model? Did you try a few different seeds? And/or a different set of prompts? (it may not work well for everything)
@neph1 yes to everything.
@robbiebunny1883 One more thing: Are you using several other loras, and/or high strength lightning loras? I've noticed they can diminish the effect. I usually just use one at strength 1 (many workflows suggest 2 or 3).
@neph1 i had the same thought and disabled it but same issue.
I really love this Lora!! Will there be a T2V version?
This, and other LoRA like it, but specifically this one is so good at generating very fast small clips of interesting and prompt adherent scenes, its amazing for clip and frame farming for setups for new videos or exploring the limits of the model and finding ways to tease out things from the model without spending forever to do so, thank you for this!