WAN 2.2 FLF2V + I2V + Continuation for Infinite Loopage, Long Compositions, Not Sucking and Whatnot

I was going to post this as an update to my previous FLF workflow, but it's now quite different. It will now also run regular I2V as well, with the option of continuation. I have tried so many ways of doing FLF, all the ways, with all the models. I keep coming back to plain old 2.2. With lightx2v. You can bridge almost anything. *I don't know why, but my CRT toggle buttons are not showing up in the workflow images. I will upload a screenshot, won't have an embed of course, but you can see what goes in the holes in the images* Colors below correspond roughly to the colors I set to make the buttons unambiguous.

I whittled this down from a much larger setup that also includes 2.1, VACE and FiP. That's way, way too buggy to share so I made it all 2.2 so the model doesn't need to change for any of the operations. I think I snipped all of the loose ends, but there might be vestigial inputs here are there. I don't think they will cause any problems. I ran through every possible permutation of inputs and it's all good on my end. I tried to make everything that has widgets which might need attention brightly colored. I'm not sure how colors work between setups though, probably depends a lot on your custom nodes. Well, at least they can be seen in the WF images. I included JSONs with the different modes enabled. Not that it's ambiguous. Just two clicks. Turn one button on, turn one off. That's it.

You can set any combination of possible inputs. Image to image, last frame of a video to the first for an infinite loop, last to first of a video for a bridge between them, continue from the last frame of your most recent run, etc. You could pull a frame from wherever you like in the video as well, though I didn't make a specific module for it. The filing system is set up such that your source names will be cleaned up automatically and appended where appropriate for keeping track of what goes where. There is one caveat, see below about the VHS upload choice.

To continue an I2V after the first generation, keep the I2V button on and enable the CONTINUE button to the left. It will pull the last saved image out of your FLF directory and run it. After the first load, you can queue up as many generations as you like, you don't have to wait for anything to end. Before the last click, switch to FLF with the END source set as your original starting image. Bear in mind that quality drops at time increases, and this rate is increased as your latent size decreases. Default directory is yourcomfydefaultfolder/FFLF. You can change this in the ORANGE box in the Filename by Function group.

For I2V you obviously only need to load a start image. However, the first time you run the WF, missing inputs in the other spots may cause errors that require a dummy input. For the VHS load nodes (drag and drop), just give it anything. You should be able to give it a tiny png. Anything will suffice, I think. Same with the path loaders (orange).

The sources in the WF are symmetrical, all the switching is the same, the START sources are on the left, the END sources on the right. There are two choices is you switch from Image to Video source, upload or path.

Note that you must load videos TWICE for each group if using VHS upload video node. This is because you can't reference the node to itself to extract the final frame, this would be a circular dependency. The first loader just provides the video info for the second one. Don't confuse the two loaders in one group for start/end loaders. Start and End have their own groups, left and right respectively. PATH doesn't need duplication, as it uses a string that goes to two nodes.

I haven't figured out how to extract the file name from VHS video loader. Very annoying. Therefore you will have to supply names for each end, that is to say the automatic naming system can't grab the source by itself. If you don't care about that, it will just use whatever is in the RED boxes next to the loaders.

All of this is still built around individual generations. There is no way to automate the editing process such that the result is better than manually editing. There's always something you need to trim or ramp to make a presentable composition. I know everyone loves the 3 second porn, but I hate it so much. And it's so easy to make your stuff suck less. Anyway if you just want to merge your outputs, use Shutter Encoder. It's free and it's awesome. And instantaneous, whereas stitching in comfy takes much longer.

Pingpong is turned on in the embedded images because I hate cuts, and I want the videos for this workflow to show up (a ton of my edited uploads just don't get rated so they're invisible to everyone - I don't know why the limit is 4 minutes if everything that I upload that is longer than 1:30 seems to never get rated unless I start a ticket). Whatever, if you see my workflows but no/little media in the gallery, this is why. Also, what does get rated always goes to XXX. Maybe the site sucks less if you pay for it. I'm not going to find out. I'm just killing time.

The last frame saver spits out a tiny dummy image in the FLF directory when running on FLF. I had some issues when trying to bypass it for this mode, so this is a workaround. You can delete all the nodes except for image list and image batch if you can get it to bypass yourself.

There are plenty of notes in the WF as well. Anyone who's even half-familiar with comfy will have no problem using it. Nobody seems to be interested in this stuff, which blows my mind, as FLF is the only thing that makes video generation actually useful, but such is life. Use it or don't. I don't want buzz, I don't want thumbs. I just want people to make something watchable, and upload it to interrupt the stream of meh once in a while. Mine won't post, maybe yours will.