This is a workflow base from kijai wanvideowrapper. very user friendly. I added for loop and also added lora for each video step. I Took idea from some other workflow then did it with kijai node and added some note in the workflow. I cleaned it up also. Im using an rtx 3080 so everything is set for 10 to 12GB. but you can tweak thing as you want.
For those who doesn't know what for loop it is basically longer video and it take the last frame of your first generation and continue your video from this last frame etc... so you can get long video.
EDIT: I saw that some people was still downloading my old version and had some question so i updated it and made it a bit more clear.
Description
FAQ
Comments (27)
Thanks for making such a clean layout. Excellent. Initial test worked great.
I really like how you setup the HIGH/LOW CFG 3/1 and 2/4 step split. That has produced the best motion for me out of various workflows.
thanks for the appreciation. Yes its hard to know whats is best. Cause sometime a cfg of 3.5 in high noise is also viable but sometime it overcook the image and other time it work better on some image. Also in the workflow i found out that putting 2.0 str on the lora high noise per steps was a bit to much and killed the prompt adherence a bit to much so i reduced it to 1 str on the high noise lora and the lower one also at 1str and put another similar lora that do the same thing. but im still checking whats work the best. Also 8 steps total end at 3 on high noise 3.5 cfg then start a 3 on the low noise yield me some good result also.
donzo15 I tried this light lora combo and it works really well: https://www.reddit.com/r/StableDiffusion/comments/1mkc6xf/psa_with_wan_22_combine_the_new_light_22_v2i/
It's faster without having to use the 2-3 CFG on high and the movement/quality seems really good. It might be worth it to try and if you like it add an Acceleration LORA for high noise node to your great workflow to give option to easier switch between HIGH CFG mode and light 2.2/2.1 combo CFG 1 mode..
iodrg244 i tested it had some good result aswell thanks for the hint. Still looking for whats best also.. Thanks,
donzo15 returning to this just to say that after 100's of tests, you really do need a higher CFG than 1 with HIGH noise. I'm currently using CFG 2 along with Lightning 2.1 and 2.2 at str 1 on HIGH with good results that has a little speed boost. but CFG 3 and no lightning loras on HIGH noise will give the best results if time is no issue. MediaSyncer is a very low resource media player that makes it easy to compare AI video clips which makes it easy to pick out the differences.
Recently tried something and it gave me some good result: 3ksampler total. First ksampler HIGH, NO LORA, 2 steps, 2 cfg, second ksampler HIGH, with LORA lightxv2 rank64 at 1.5-2 str, 2 steps, 1 cfg. 3ksampler LOW, with LORA lightxv2 rank64 at 1.5-2 aswell. For 6 total steps but you can try 6 to 12 steps also.
Just link 2ksampler with the high model and one sampler with the low model. Play with the denoise on top of the ksampler, for the second High ksampler.
donzo15 wow triple sampler solution. I was going to say you could use RES4LYF's Constant scheduler node to adjust the CFG using the same sampler but I don't think you can scheduled the loras so I guess you'd need the third KSampler. I'm pretty happy with 2 CFG HIGH with 1 str 2.2 and 2.1 lightning and 1 CFG 1 STR 2.2/.25 STR 2.1 LOW lightning, 4 steps low, 3 steps high. Hehe. That's my secret sauce for now. It still feels fast to me (4090) with great lora/prompt adherence and motion. I think we'll eventually get another version of 2.2 lightning.
donzo15 I'm back to no speed loras on HIGH. I'm seeing if I can get by with 2 CFG and 4 steps on HIGH and 3 steps LOW with CFG 1 and 1/.25 2.2/2.1 lightning. It's about 1min45secs for 101 frames (704x480 which seems to be a resolution WAN2.2 likes) without the speed loras on LOW with those settings. Just need to do some some side-by-side comparisons between various CFG settings now on HIGH.
donzo15 check out this new Ksampler for WAN2.2
https://github.com/stduhpf/ComfyUI-WanMoeKSampler
it combines high/low in one node and allows you to set CFG separately. it also has a "boundary" setting .875 i2v and .9 for i2v which auto selects the LOW noise when timestep appropriate so you one need to set the total steps and it figures out when to switch. I've only done one test so far but it worked great. Very promising.
iodrg244 Yes i saw that ksampler. Make thing easier there also the advanced if you want to split it also.
Very nice workflow; I tested it with default prompts so I put a dance lora on the second column high and low but it affected the first video as well, not an expert but it looked like all the loras loaded instead of the individual column's. I didn't find the quality good so copied the settings from the original for the samplers along with cfg schedular, resulted in much better quality output. Yet no matter what at some point soon character consistency takes a dive. I'll prolly only use 2 loops instead of 4 but that's just me. Thanks for the WF.
I just did a test. there no influence from the other lora in the first video. But sometime the first image if you put a person they can dance by their own even if you doesnt prompt it.
donzo15 I see, well then it should be easy to prevent with additional prompts
Very good WF and nice use of loop and great using Kijaj nodes as one can test new functions not available with native nodes.
I'm now testing with WAN 2.2 native Lightning 2+2 steps loras + Sigmas and seems very good. I'm shocked how good it follows prompt in just 4 steps.
Thanks i mainly took idea from https://civitai.com/models/1829052/wan-22-for-loop-video-with-scenario?modelVersionId=2070152 then cleaned it to be user friendly. did some adjust and did it for kj. Thanks. And i will test your suggestion also.
There is a high noise link under low noise download locations in the workflow
If you have a 20xx card set the attention_mode to sdpa in both WanVideo Model Loaders in the part of the workflow labelled: 14B GGUF Model loader.
Not sure if there's another attention_mode suitable for 20xx cards. sdpa works for me.
the missing node "wanvideoxxx" ,how to download
[Internal Download link wrong]
Thanks in advance for the kickstart with this workflow. Can't tell yet if it works, but I'm about to find out:
Found a lil paste/syntax problem; In your workflow's listed download-links there's the very same link pasted in for the high noise and low noise model. It's both leading to the high noise one.
Can you confirm the low noise model you were referring to was this one: https://huggingface.co/bullerwins/Wan2.2-I2V-A14B-GGUF/blob/main/wan2.2_i2v_low_noise_14B_Q4_K_M.gguf ?
yes it is. sorry im not uploading a new one but if your interested someone took my workflow and made it a bit better. my workflow was just a basic one when wan 2.2 came out. so its a bit outdated: here the workflow the person made: https://civitai.com/models/1945794/i2v-lynxpusamoelightning-x2-speed-face-consistency-with-loras-looping-workflow-wan2214b
@donzo15 Sure thing. It's only a minimal typo and you're surely busy with a million other thingies right now, so can comprehend. I just wanted to make sure if I picked the right and report the typo at the same time. Thanks also for the link, I'll have a look in comparison!
(Just came back from a sobering WanGP run as an introduction to AI-Video creation. Though it was easier, but turns out it's cheaper/worse/less adjustable than ComfyUI, so around here I am now.)
VERY NICE WORK THX ! but i dont understand how i can make longer Videos? ich have a RTX 3090
i might just do a small update of the workflow and add the new lightx and some minor improvement. But THE TOTAL STEP COUNT node is the number of time the video extend the last frame of your first generation. So if its set to 1 it will only do one video set to 2 it take the last frame of the first video and extend it to a second video and etc...
this workflow is abit outdated i will update it soon just to tweak some thing to make the video more fluid and responsive. Just tell me what you dont understand ill try to help ya. And sorry im not checking my workflow alot since it date a bit already.