The non-speed lora version of the WAN 2.2 T2V Template (which is bypassed by default) has a mistake:
The low pass KSampler is unused and the first/high-pass KSampler is set to execute all 20 steps, meaning that the low pass model/KSampler are never used.
As can be seen in the screenshot (workflow is the template copy-pasted straight from this repo, 100% unmodified) the high-pass sampler executes all 20 steps instead of ending at 10, and the latents are passed straight from that first sampler to the VAE Decoder.
The speed-lora version up top is fine, which I imagine is why this wasn't noticed sooner.
It appears to have been introduced in this commit:
df5ebb3
It doesn't seem to be present in the I2V version of the workflow template.
Follow up edit:
I just checked the other WAN 2.2 templates modified in the same commit and it looks like this one was the only one with the issue, the rest properly go through the low pass sampling.
The non-speed lora version of the WAN 2.2 T2V Template (which is bypassed by default) has a mistake:
The low pass KSampler is unused and the first/high-pass KSampler is set to execute all 20 steps, meaning that the low pass model/KSampler are never used.
As can be seen in the screenshot (workflow is the template copy-pasted straight from this repo, 100% unmodified) the high-pass sampler executes all 20 steps instead of ending at 10, and the latents are passed straight from that first sampler to the VAE Decoder.
The speed-lora version up top is fine, which I imagine is why this wasn't noticed sooner.
It appears to have been introduced in this commit:
df5ebb3
It doesn't seem to be present in the I2V version of the workflow template.
Follow up edit:
I just checked the other WAN 2.2 templates modified in the same commit and it looks like this one was the only one with the issue, the rest properly go through the low pass sampling.