This is the conversion to GGUF of the Wan2.2-Remix I2V v3.0 - https://civarchive.com/models/2003153
For v0.8, go to https://civarchive.com/models/2097002
For v2.0, go to https://civarchive.com/models/2110981
For v2.1, go to https://civarchive.com/models/2347977
T2V available in https://civarchive.com/models/2094656
Files available also in https://huggingface.co/BigDannyPt/Wan-2-2-Remix
All were using umt5-xxl-encoder-Q8_0.gguf and Wan2_1_VAE_fp32.safetensors
Description
FAQ
Comments (24)
Cab anyone tell me how to use this? TY
You use it as you use any other GGUF, in comfyui with the ComfyUI-GGUF node.
Do I need to connect the lighting LoRA when using it?
No, and check the description from the original model.
any idea why I keep getting a blurry mess when I try to run this? Never had anything like this happen in years of running just about every kind of model. Not really seeing anything on the main page that looks like it'd help me
most likely you don't have the correct steps in both ksamplers.
First should be high with:
- steps 6 or more
- start_step 0
- finish step as half of the steps
Second should be low
- steps same as high
- start_step half of the steps plus 1
- 9999
As an example:
High
- steps 6
- start_steps 0
- finish_steps 3
Low
- steps 6
- start_steps 4
- finish_steps 999
ah right, @Santodan , damn it's been so long since I bothered to change steps on anything lmao. I think I've been running the same steps on everything for a few months now. In the beginning I experimented with it more. Anyway thanks, will see if this works
Q8_0 LOW is broken.
Need more info than that. I used it and got no issues, I see people posting here without issues, so I can only assume that you are doing something wrong with the information that you provided
Some people have mentioned that the output is blurry. I had the same problem when using Q8_0 LOW version. However, when I tried using Q6 version, the blurriness disappeared completely. I also tried mixing Q8LOW versions 2.0 and 2.1, and there were no issues at all.
@a0978750321122 And did you checked the steps like I told in the other comment?
because that is normally from the misalignment of steps, like having the high from 0 to 3 and the low from 2 to 6.
Wan 2.2 should always have half the steps in both sides without repeating the steps, so 0-3 in high and 4-6 in low for a 6 steps workflow.
Can you provide the workflow and what was that you used for clip and vae?
@Santodan My steps are correct, and I've tried other workflows, but the image is still blurry. Versions 3.0Q6, 2.1Q8, and 2.0Q8 are all normal and not blurry; only version 3.0Q8 has the problem. This seems to be the result of using two HIGH versions for output.
@a0978750321122 let me do some tests in my side to see if I get the same problem
@a0978750321122 ok, I'm able to replicate with both files from civitai and hugginface, not sure what happened since I was able to use them, when I created them.
I'm now trying to create the Low Q8 again and will update both places
@a0978750321122 Can i ask you to do one test before I upload it here, becasue I found the reason I was having issues in mine.
I had the high from 0 steps to 3 steps and low from 4 steps to 6 steps, but it shouldn't be like that, it should be high from 0 to 3 steps and low from 3 to 6 steps. after doing this change it worked as intended
@Santodan Whether it's 3-6 or 4-6, the problem is the same. Having tested other versions of the LOW sensor, my conclusion is that your Q8's LOW sensor is faulty.
@a0978750321122 try with the one from huggingface before I upload here. That one I created today and it worked. Tomorrow I'll try with the one from here with the correct steps
yeah it's blurry i can confirm
@raidou88 share the workflow, because it ain't fir me with the one in hugginface
@raidou88 I've jsut test it again and it is working the one from HF, going to replace the file here in CivitAI
@raidou88 @a0978750321122 I've updated the Low, can you retry it? my last videos were created using it, and it works as intended
@Santodan The newly uploaded file works fine after testing, thank you.
@a0978750321122 I really don't know what happened.
I generated the videos thare are here with it, and odd that it was the only one.
But well, now is solved, gladly.
@Santodan yeah i tried the one in hugging face and its working correctly. thanks