This is the conversion to GGUF of the Wan2.2-Remix I2V v0.8 - https://civarchive.com/models/2003153
For v2.0 go to https://civarchive.com/models/2110981
For v2.1, go to https://civarchive.com/models/2347977
For v3.0, go to https://civarchive.com/models/2472759
T2V available in https://civarchive.com/models/2094656
Files available also in https://huggingface.co/BigDannyPt/Wan-2-2-Remix
All were using NSFW-Wan-UMT5-XXL (FP8 version) and Wan 2.1 FP32 VAE
Description
FAQ
Comments (13)
Oddly this runs slightly slower than, https://civitai.com/models/2053259?modelVersionId=2367702 , on my 5080. Same workflow, input image, settings, everything.
yeah, the more quantized a model, the better it fits in memory, but typically it's gonna be slower. in my experience Q8 GGUFs are the fastest, each further quant like Q6, Q4 etc. add about a second or two onto your gen times, ironically making the smallest possible Q2 model also the slowest
@7satsu But with a Q4 or Q2 your GPU doesn't cost 1000$+ xD
I'm still uploading the other version, hopefully I can upload all today, was having some issues yesterday
workflow???
Tried alternative getting error:: The expanded size of the tensor (22) must match the existing size (44) at non-singleton dimension 4. Target sizes: [1, 48, 1, 30, 22]. Tensor sizes: [16, 1, 60, 44]
Ghe workflow is the default one from comfyui for Wan 2.2 I2V with the text encoder that is in the description.
What have you used?
@BigDannyPt ok...Great....Working with correct vae and TE.
Usually that happens when using the wrong Text Encoder. Did you try with the recommended one by @BigDannyPt ?
Was this meant to be a reply to someone?
Ah sorry, correct, it was for @durgeshvp93
Thank you very much for these. Hats off
this is amazing, but not perfect as it keeps changing faces and adding nipples to swimsuit tops.
same problem, doesn't keep the original face. =(