🚨🚨🚨MISSION CRITICAL: PLEASE READ🚨🚨🚨

Would you mind sharing your workflow?

iokk I am using the default wan2.2a14b_i2v workflow provided by Kijai. You can find it in the Kijai node package. I only modified the generated resolution and the corresponding model address. All other parameters use the default parameters of the Kijai workflow. The large model uses the Kijai quantized version a14b_i2v, and then the v0.06 version is added in front of the high-noise model with an intensity of 1. The generated results are good. I have not yet tried whether to add it in front of the low-noise model. I look forward to your tests. Thank you.

No issues running it on a 3080 card here. I started w/standard wan 2.2 from the "browse templates" workflow. I added one of the accelerator models (wan_lcm_r16_fp32_comfy) @ weight 1 to both high and low noise, then add other loras as wanted. cfg 1.0 on both ksamplers, 10 steps on both samplers, and have high noise end at step 5, low noise start at step 5.

I've also added frame interpolation, and get 4-second clips in about 220 seconds.

niche34filler Okay so it’s a skill issue with me 🤣 I am quite new, gotta learn what all the nodes do

Using 3080 10GB, Kijai nodes not working well for me ,crashing with OOM even on Q3,but with native nodes using any model (best in term of speed and quality for me is Q5_0 gguf, but i could run with fp16 but it was like 20 minutes for 65 frames in 480p) , in workflow with image generation → picture captioning → prompt enhancer → wan with loras (with lightx and this nsfw lora) → interpolation and upscaling average time is about 500 sec for 480p 81 frames. if i disable image gen and upscaling with interpolation,then average generation time is about 220 sec

Ambercleaved im observing the same. if i just take the example i2v workflow form comfy org, i can run the full fp16 models on 16Gb vram, although 480p 81 frames are made in 30mins. but even then i can compromise time and make 720p 145 frame sin 3hours. but if i use wan wrapper nodes, it just ooms at higher settings. would appreciate if someone could help get this lora working in the comfy org example workflow.

You need ram + virtual memory (ssd) to load models https://civitai.com/images/92003430 was generated on 6gb vram + 40gb ram + virtual

ice02 Ah I think I should be covered then, 32GB Ram, SSD & 4070 with 12GB vram, I’ll see if I can mess around to get something going here

get the wan2.2 GGUF Q4 ; works perfect with 8gb vram

CubeyAI the problem is my pc, has 4gb vram, so i use a virtual machine, but now the problem is, there no space for wan 2.2, is to heavy in hd

Leamas_ Leamas_  there not much hope with 4gb vram unfortunatly

heatchv01812 4GB of VRAM barely runs SDXL, but I've accepted that, so I use other sites to run comfyui. The biggest problem with these sites (besides the time limit) is the space limit. Either I remove everything and leave only WAN, or I don't have enough space.

Can you share the workflow?

That's called Wan Vace
https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_wan_vace_14B_v2v.json

2P2 For some reason my computer keeps freezing up at a certain %. Would this be not too draining on my pc? i have like 12g of ram

2P2 I tired the work flow but for some reason my image gets over taken by the reference video. How do i make it so the image is in the reference video?

Aiartsenpai Did you try adjusting workflow? is the reference somehow related to the initial image? and sorry, I can't allow notifications and neither I usually use VACE for that

Q6 is 11.1gb...
Q8 is 14.3gb

I would use Q6 to have decent room for doing large images etc
idk if it would spill for long 5s videos etc
T5 XXL will definitely get unloaded (and reloaded) for each generation, but that will only add a few seconds

Curious too about this, Pls tell me how it goes.

sidenote. respect for hololive gens. keep up the good work baerat lol

Because you're smashing it with too many loras or with too much strength on the loras. Or not using enough steps.
Or you're using a quantized model of which theres not much you can do because compressed models dont have much quality.

I had blurriness when I was doing an accelerator, but didn't adjust steps accordingly. I was doing 10 steps, but high/low noise crossover step was above step 10, so no low-noise steps were active. It seems about 20% high-noise is about right with the accelerator I use.

because you haven't specified otherwise what they should be doing.
put "talking" in negative prompt, or if you're using 1.0cfg put "moaning", "smiling" or something that isn't talking in the positive prompt

tmp88 i will try again thank you

You might have to weigh the phrases to prevent this from happening, i.e. (talking, speaking:1.2) in negative prompt. Just analyze what you don't like about the generated video and adjust your prompt accordingly.

Another thing to watch out for is jittery movement, this is often caused by LoRAs that are weighed too high... which unfortunately is often necessary if you're using Wan 2.1 LoRAs in 2.2, but you can mitigate this by putting something like "jitter, shaking, head jitter, shaking tongue" in your negative prompt

magicballoon ok thank you bro

you need it on both I believe, that's what I did and it turned out great. will be posting my test video soon, just waiting for the upscaling to finish. Will be posting the separate vids too so people can look at the workflow.

Oscarc Yesterday I dl the Wan2.2-Lightning_T2V-A14B-4steps-lora_HIGH(LOW)_fp16 lora,but it can`t work,anyone try it?

sekaiwlc07860 The optimization LoRAs are a mixed bag. I've used lightx2v personally and it works. However I think it works better if you only use it on the low noise model, probably has something to do with it being a Wan2.1 LoRA. What many people don't tell you is that these LoRAs will diminish your video quality.

magicballoon Wan2.2-Lightning_T2V-A14B-4steps-lora_HIGH(LOW)_fp16 lora I have work it out...But i don`t know whether faster or not

Nice ! .. so you did it just like in the video and this is what you got? How long did it take to generate all of this?

iokk yeah pretty much, I also added start and end img-to-vid. if you want the workflow you can download one of the separate vids (any of the ones that are 5 seconds) and drag it into an empty comfyui workflow it will hopefully give it you the workflow used to make it (I'm new to Comfyui so might be wrong >.<). and it roughly took about 3 mins 25 seconds for 5 seconds of video (with 8gb vram). So to find the right seeds for each picture took quite a while 😅 (a few hours)

Does your tests with 1 img only works well too? Mine are looking bad even using Q8.

anquinoum I been trying out this LoRA (the Phantom one https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/tree/main/FusionX_LoRa) it seems to help to increase the quality. So I would recommend adding it to the LoRA loader and hopefully it will help fix any problem you're having. also having a very descriptive prompt gives you a better chance of getting a good out come. hope this all helps!

anquinoum What's your hardware? If you have low VRAM, you should not use Q8, go down to Q5 or Q4 instead

man.. thats really nice stuff, good job and thanks for sharing! question how did you seamlessly combine it? manual editing?

LewdAi hey there, yeah here's what I did.

1. generate the images for the start and end frames (for me I just took ones from my LoRA showcase here https://civitai.com/models/438394/pov-cowgirl-looking-down-updated).

2. then you can start generating! once you've got vids you like, you then load the vids into any video editing software, as the all the vids share a start or end frame with another vid it's and simple as putting them in the right order then export the finished vid.

hope this helps out!

Oscarc Wait, am I reading this right? You got first-last-frame generation working for Wan 2.2? How'd you do that? I thought you needed a separate model for this.

Oscarc thanks a lot, ill give it a try!

magicballoon yeah it's works well in comfyui. If download one of the separate vids (any of the ones that are 5 seconds) and drag it into an empty comfyui workflow it will hopefully give it you the workflow I used to make it. 🙂

Oscarc Oh, duh. I don't know why I keep forgetting this.

Wan2.2 handles both 480p and 720p from the same models

it seems wan is not available anymore since today or maybe yesterday