This is a simple but large workflow, i use all the time, using Wan2.2 and SVI Pro for long videos with consistency over multiple sections. It's a extended version of the default SVI Pro workflow. Much simpler than the (great) infinite loop workflow and other fancy svi pro based workflows. But it is more comprehensible and it is much easier, adjusting almost everything for every step individually, including extending the video step by step without re-doing everything every time. . Including the prompt. (disconnect the get_length and so on node from a chunk to set it manually).
I made a little video explaining all this in a ... video. Alternative to the text below. the first 6 or so minutes explain how to slowly develop a video.
Be aware, it is - poorly - ai voiced. But i think it shows it off quite well AND makes clear what i talk about here.
The TL;DW summary is here: Even at 4/2 steps using Wan2.2 Remix 2.1, the results re good, though around halfway through the image gets seriously less detailed. Having run it also with 8/4 steps at 1MPx for 24parts (took a long time, and step by step extending, see below.), the image quality stays remarkably good even after 24 parts. That video stays private though. It is in theory infinitely extendable.
little tips and tricks, applying to other, similar workflows too:
One easy to see and use advantage is that once a chunk is rendered (you visible see it when it's done), ComfyUI stores it's finished frames in node cache, and it will just return these frames without rendering the next time. So you can connect the next chunk (prompt) and run the workflow and all previous chunks are skipped because their frames are already buffered.
All settings including Seed and prompt of the already finished chunks need to remain the same though. Otherwise it will start over.
So you can slowly develop a video bit by bit, even stop and adjust without losing progress. For example if you see in the running preview ) that it does something weird, you can just stop it. then adjust the prompt to prevent that, and run it again and it will do that chunk again with the new prompt, but all previous ones just return their frames again from cache.
btw: (live running preview needs videohelpersuite installed in comfy, than in it's settings you can turn it on. It saves a LOT of time to notice if something goes wrong. )
It's a complete game changer for long multi bit video workflows for me (when i discovered the magic of FIXED seeds) . You're not helpless when something goes wrong halfway through.
I DID run into a problem: it uses ram. When running all 18 parts and at length 101 per chunk, i hit the ceiling of my 96GB ram on the final overlapping step (before the video is saved) and everything got QUITE slow.
On lower ram you may want to keep that in mind when using many chunks/higher length, higher res.
Description
i hadn't connected step 14. so that's V2.
FAQ
Comments (3)
Hello, thanks for the explainations about fixed seed, same prompt X times and 81's frame limits. I now understand SVI subtility.
it's hit or miss. if the scene is relatively simple you often get away with 101 frames too. But if certain things should happen and it just wont do it, it's likely that.
And tbc you don't need the same prompt every time, even in my simple lazy example it changes later. each sampler/chunk (lets say example #8), stores it's finished frames and from then always checks if it's inputs (#8) changed. If not, it will just return it's frames. So what you do in #9 will not affect it at all.
Just don't change #8 or earlier ones OR a general workflow settings, unless they come after (FPS).
You just have to be careful because what enters a scene that is NOT in the start image won't be consistent. like person that wan added later, leaving and then coming back or even turning around for a moment and is not visible in the last 5 frames, next chunk will have no idea what they looked like. While a person visible in the start/anchor image (and named or identified in the initial prompt) will.
the main thing is you can experiment in a new chunk without having to redo everything before every time.
@andyx1976 Hi, Andyx1976, thanks for this explanations. this is helpfull. Have a good Week End