Download
1 variant available
I forgot to fix one thing in color match: if enabled you may get one more frame than if not enabled (depending on selected fps and duration). Furthermore, I didn't mention the specific reason for not truncating frames, which is this: if you're dealing with serial transitions, even with the anchor frames I included, you still want as much opportunity as possible to get a smooth comp in editing if they are left in. I deliberately did not remove them in the example posts to accurately reflect this intended output. The color issues are not from the color match.
LTX is currently at bat in the quest to get FLF right. Everyone advertises it, and and nobody shows it. Liars.
First version is strictly I+I. I have a video version that I will add as soon as I have some finished comps to post for it, and after that I will add the context-sampling, ie SVI-style continuation. I started simple to see how far I can push the model with just two images. Nothing novel here, just the way I do things.
The backbone is essentially the same as the comfy template. I've added presets and automation to suit my needs. All that needs to be set is duration, framerate and resolution.
There are two preset resolutions set with an INT, HD style- 720 or 1080. Now this setup, unlike some other implementations of LTX, will always spit out a 720 input as 704. You CANNOT force it. Keep that in mind- whatever you change the presets to, check to make sure the resize value you set matches the final output, as the whole point here is to make transitions, so the resolutions need to match if you're using frames pulled from video.
Orientation is automatic - if you change the preset sizes, enter them as landscape orientation only! It will detect if the image is tall and flip the values by itself.
There are two options for guidance. The default is a KJ node. It uses each input image twice to create a hard anchor and the start and end while giving the model lots of freedom in between. If you flip the guidance switch over to LTX, it uses the template style guidance, which uses two nodes, one for each image. It also modifies the conditioning, which the KJ node does not. I think the KJ node is much, much better, so it is the default.
24fps, 25fps, 30fps all work fine for me. I'm posting a 24fps portrait example and a 30fps landscape example. Both 704.
The last step is a color match. Default is an MKL match at 0.5 strength gradually blends in throughout the entire duration, so it starts with a match to the first image as normal and ends matched to the last image. I'm still dialing this in, as can be seen in the examples. The compression on this site also does its best to make everything look as awful as possible. You can bypass the entire step with the appropriate toggle.
You MUST use the transition LoRA! This is the secret sauce. Don't tell anyone. https://huggingface.co/valiantcat/LTX-2.3-Transition-LORA/tree/main This LoRA has the training needed to get the effect that is desired here - sometimes the WF will work by chance without it, but don't bother. Trust me. I would not be posting this without it.
I prefer Eros and the Frog, with heretic as the text encoder. The rest is standard LTX fare. Fill in the blanks. If it's not naughty, anything will do. But it's set up to run distilled.
Audio will come out with the video as usual, but I've not paid any attention to it in this version of the WF. When using images only (as opposed to frames) I'm always going to use a musical track instead of what is generated. You can prompt in what you like.
Prompt is set up as a concat - the bottom spot you can leave alone, as it is all good stuff for making the modified model do what you want- smooth morphing bridges, as opposed to crossfades or jumps. The top is for throwing in specifics if it needs help. Stuff like camera movement, character changes (she gets more body positive, she identifies as a cucumber, etc.). If you do get an unwanted jarring effect, here is where you can add in the cues that are needed.
I think that's it. There are copious notes throughout every group. I'm extremely pleased with this, the FLF gauntlet has been a real headache, sometimes feel like I go backwards more often than I progress. This is most definitely progress. I love this model so much. WAN can go suck a futa dong.
The video to video version works just as well as this, and you can stitch the output right in the WF as well, using either the last+first frame of an existing video to spit out an infinite loop, or two different videos to create a seamless transition between them that can then be continued ad infinitum. That version shall be added post-haste. The only think I'd ask is for a thumbs up, or down if you find it useful, or awful. Just so I can tell if any actual human got something out of it either way.