i2v
Thanks for the gguf versions you uploaded! I wanted to ask if you plan to make the i2v version too
test the t2v first; see does it work or not; i2v has another clip to check; and comfy side is working on the code support
Is it different from city96 Q6 version ? To understand if I can avoid downloading it for nothing; I'm using the native comfyui workflow and it works; his q6 model and yours weigh the same too. I had to use the text encoder and vae provided by comfy-org though, as I wrote here. The encoder and vae i discarded were the ones made by city96 though, not yours.
About the i2v gguf version, when/if it will be available I think I'll test first the clip vision provided by comfy-org
thanks; testing
thanks; testing
"thanks for testing" ? You're welcome, luckily I already made some test t2v-ideos with city's model.
the latest one from comfy
The comfy-org models, I suppose. Yeah, I think so too. At least the gguf unets still work...
Also, are you using llamacpp to convert the wan model? (I get the architecture error but if I edit the config file then the code says it doesn't recognize the new architecture)
the encoder and vae from comfy side have different structure; should redo everything right away otherwise it won't match
When you update them I can test them, then, maybe. Anyway, what are you using to convert them? Do you need to build the architecture "from scratch" and pass it to the llamacpp gguf converter?
Can you upload a t5xxl um q8_k too? I use the q8 as not to lobotomize the encoder too much, and it saves some space too