Another Wan 2.2 SVI I2V workflow.
I tried to keep this relatively basic with no subgraphs and only a few set/get nodes, hopefully it is digestible for beginners.
There are three separate workflows used to create extended videos:
i2L start - create first clip, save to latent file
Once you're happy with the clip, copy the latent file to the input folder
i2L extend - create subsequent clips loading previous latent file and saving to new latent file
Select previous clip latent file, generate subsequent clip, once happy copy it to the input folder
L2V combine - combine all latent clips and convert to video
Once you have all the clip latent files, load them into the combine workflow and generate the final video
The 3-workflow system is nice in that you can easily re-roll clips, managing the latent files externally as needed, allowing you to curate to your liking. You can keep extending as long as you like then output to video when you're happy with it. I've found many advantages to working with clips as latent files. However it also means you have to babysit each clip, manually moving the latent file over after each gen - this system isn't designed for "set-and-forget" type of work.
Custom nodes
KJ
WanMOEScheduler (technically optional)
Impact (for wildcard encoder, technically optional)
VHS (for combine stage, optional, just replace with save-to-video technique of choice)
VFI (for combine stage, optional)
Adding loras
Keep the additional lora weight on low noise to 1.0 or less. I've found that too much lora on low noise causes very noticeable degradation.
Discussion
This is basically a "classic" triple-sampler setup inspired by one of KJ's workflows, but using WanMOEScheduler instead of shift. It is set up for 4/2/3 steps, 9 total. The goal was to ensure the sigma values aligned for the handoff between each sampler, and to have the high/low transition at 0.9 sigma as I've seen officially recommended.
I do 480p for nearly everything and still sometimes get artifacts, particularly with realistic images where the character blinks the eyes can look weird for a frame which is off-putting.
Feedback/tips/pointers always appreciated
Description
FAQ
Comments (10)
been looking for a while for a workflow using latent for svi. im preety used to mess with workflow but i admit im lost with yours. can you make a video on how to use it?
I appreciate the suggestion but honestly I'm probably not going to make a how-to video, that's just not really my thing, sorry. Can you elaborate which part is confusing? I can try to explain better and am happy to answer questions.
1. Generate first clip using the "start" workflow. The output will be a .latent file in the output/latents folder. (To load the starting image with the workflow as-is, put it in the output/wan_todo folder, or replace the load images node with whatever node you use to load an image.)
2. Copy that .latent file to the comfy input directory.
3. In the "extend" workflow, load that .latent file in the load latent node in the prompt section.
4. Generate second clip using "extend" workflow. Similarly, the output will be a .latent file in the output/latents folder. It will also show a preview of each clip and them combined so you can see if it's what you want.
5. Repeat steps 2-4 as many times as you like.
6. In the "combine" workflow, load all the .latent files in order and run it to generate the final video.
thanks for the quick reply!
Step 1 and 2 are fine!
Step 3 in the extend workflow, the loadlatent node has partial_00001_.latent selectected in it and give an red error without message. i can't change the value therefore can't put my saved latent to use it.
i don't think step 4 5 6 will cause problem but i'll follow up if i can pass step 3 haha
@BottleForkBerry77 quick follow-up, i bypassed the fact that i can't edit file name in the load latent node by renaming my latent with the default "partial_00001_.latent". he took it, so i can manage to go further by renaming everytime each latent. by i'm confronted to the same problemetic in the combine workflows. i can combine 2 latent since the first and second are named svi_00001 and svi_00002 but can't go further since all extension are named svi_00002 and i can't change it. it must be something really stupid that i didn't get bacause i've never messed with latent before i guess.... so the only problem could be resumed by how do i edit the loadlatent file name???
@logicalpop The LoadLatent node is a very basic "beta" built-in node that will only read files that are in the comfyui/input directory, similar to all the vanilla built-in loader nodes. That means you have to copy your generated .latent file from the output/latents folder to the comfyui/input folder manually outside of comfy. Then in the workflow press "r" to refresh, then it should show up in the dropdown. If you're sure the latent file is in the input folder and you press "r" and it still doesn't show up in the LoadLatent dropdown there may be something funky with your comfy setup. There are likely custom nodes for loading latents you could look into as well.
@BottleForkBerry77 oh man i feel stupid now! i even changed the node for other latent node haha! but i never thought about refreshing comfy! haha of course that was it and it work now! i<ll test it out more extensively now! Thaks mate! :)
follow-up!
Been experimenting and works great! however i still have the usual color shift and degradation aat the usual 20-30 sec mark. It thought working with latent was fixing this issue but doesn't with me. You think it's normal or i'm probably doing something wrong?
@logicalpop It's possible that I'm getting similar degradation it's just not as noticeable due to the style of my starting image and scene. I definitely see "softening" of details after a few clips, which seems more noticeable on realistic images. And if the scene gets too far from the starting image things can fall apart. But the only thing I've seen cause a very noticeable color shift at the scene transition is too much lora on the low noise steps. The only other thing I can think of is are you using the exact same svi and lightx versions as in the workflow, and are you using a finetune or a lower gguf model?
@BottleForkBerry77 i did experiment with the full model instead of gguf for better consistency of my original frame made with t2v full model. i do have fews lora plugged in used in low strength to avoid that problem but i didn't try lowering low noise only. i'll try that!
@logicalpop I get color shift with more than 1.0 of extra lora on low noise (on top of SVI at 1.0 and lightx at 1.0). But it also seems to depend on the lora and what's going on in the scene relative to the starting image. Try a simple scene where not much changes and you don't need any lora (like someone just turning around in front of a blank wall) and see if you still get the color shift.