WAN 2.2 14B workflow for t2v, i2v, long video, video extension AIO with switch combination.
Description
1st version uploaded
FAQ
Comments (39)
For those that still don't get it- wan2.2 is light years ahead of wan2.1, and those trying to force 2.1 LoRAs on 2.2 are mostly missing the true wonders of the new model.
Curiously, the 2.1 lightX speedup LoRA works insanely well on 2.2, keeping both motion and full prompt understanding. So the extended videos shown in this workflow are highly feasible in decent time on reasonable Nvidia GPUs.
We are getting closer to being able to create long form narrative videos rather than well animated 5-second clips.
That's why I put those switch to bypass the lightx2v lora. For those who want to wait 20 steps for the wonder of 2.2
Best of both: set lightx to only low noise, so the basic prompt understanding and initial denoise is already done, and forward lightx to low noise only.
yorgash how do you change your steps for each model with this method? Since high noise is not running lightx.
123sirako123521 I use 10 / 8, but lightx does speed up individual steps 2x too, since I can use cfg1 instead of 3.5 for low denoise, so even going 10/10 is worth it.
receiving this error KSamplerAdvanced
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 21, 60, 104] to have 36 channels, but got 32 channels instead
What mode do you on? t2v, i2v?
i also - i2v
kukalikuk i2v
bball23 did you try with the default setting first? Extend always using i2v. I made this only with 12gb VRAM and I tested it in more than 500x generations but never got that error you mentioned.
delete flow2-wan-video node
You need to update your comfyui to the latest version and update all nodes
dread_mannequin
just remove this node, this workflow does not need this node, and this node has a lot of bugs.
SKKpapa which node? I don't remember using flow2 wan node. I made this from scratch and I checked flow2 wan is not installed
kukalikuk Go to the custom_nodes folder and have a look. If the node is there, delete it.
I'm fairly sure people experiencing this problem used Wan2.2 Vae instead of 2.1 like the template has. The only time I ever encounter these errors is with the 2.2Vae (in the "Load VAE" node).
I'm curious, how you made FFLF with WAN 2.2?
Great, Thank you for your time and efforts.!
If it is possible, Can you make WAN2.2. [ 5 B AIO GGUF ] , because Kijai Already published Wan2.2_5B (FastWan_FullAttn_lora_rank_128)
I haven't try the 5B because some examples did not do better than 14B, only faster. I'll look into it later.
kukalikuk Thank you for the reply. Have a great time.
This has to be by far my favorite WAN 2.2 Workflow!! Amazing generation speeds and accuracy, while also being fairly beginner friendly! I am able to generate 7 second 720p videos in under 200 seconds (4070ti super). I will be experimenting with adding loras (mainly movement based) to this workflow but I'm not sure how to adjust their strengths, any advice? (High vs Low)
I got that in my plan previously, I'll put the nodes later because it add more complexity to the workflow for loras that still mostly for wan2.1
kukalikuk How do I use the extend options? Do I have to write my original prompt then write 3 more prompts in Xtend 1, 2 and 3? My idea is to start with an image then slowly develop a story from it.
123sirako123521 you don't have to. If you write it, turn on extend prompt switch. If you don't, all use main prompt.
For some reason the color is changing constantly to bluish every 5 seconds or more (depending on your length settings) also the contrast changes progressively. Please fix it.
thats the effect of lightx2v, you can bypass the lora (go full 20 steps), or make your prompt longer and detailed, or add color match node just before video combine, but meh, i'm to tired to do it
Any chance of rolling WAN2.2 Loras into this? Preferably via Lora Manager :)
I tried but it looks like a mess and not sure about clip model links for the high and low.
Great workflow, but why no frame overlap?
Frame overlap in with WAN i2v node for wan22 doesn't work as good as VACE, it sometimes create strange moves, that's before lightx2v wan22. Try my VACE workflow for comparison. I'll try again later for the updates
I just tried again adding frame overlap. The problem, the node did support multi frames as input but if we go more than 1 frame (as overlap) the next first frame is always bad. It creates a "blink" of quality in the video. Go try my VACE workflow with a working overlap frames.
No idea why, but the Get_scheduler nodes don't work for me. Everything else is fine, but I had to disconnect those from all the KSamplers or it only generated a single image and nothing else.
That's a workaround. I add it since it is bothersome to change it one by one in 6 ksamplers. But if you rarely change it then its fina.
Post some of your results here please.
Do you have the comfy-image-saver-main nodes? Get the nodes, install the requirements.txt in your env? If you did it should work fine.
@kukalikuk I think I found the issue, but by installing a new environment. However now I have other issues, changed hardware and sageattention no longer seems to want to work. Incidentally to run it with the xtend on the new hardware I have to bypass the vram purge nodes. I don't think that's a problem with your workflow, I don't think the RTX 6000 series is properly supported by most things yet.