You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the method encode_token_weights in sd3_clip.py, the empty conditioning is generated as:
out = torch.zeros((1, 77, 4096), device=comfy.model_management.intermediate_device())
pooled = torch.zeros((1, 768 + 1280), device=comfy.model_management.intermediate_device())
This seems correct for pooled, but 'out' would presumably need to be 77 * 2, as per the SD3 model diagram. Similarly, not including padded zeroes for CLIPs/T5 when missing seems wrong.
It would depend on how Stability originally trained the model. Either the unconditional generated this way (presuming blank prompts) wouldn't match the tensor size of a conditional with all 3 text encoders (154 long), or the unconditional would need to be created as zeroes based on the length of the conditional.
The text was updated successfully, but these errors were encountered:
Seems perhaps you don't really need to pad the CLIP or T5 sections if not using them. They use different positional embeddings within their sub-range, so the model wouldn't be looking for them at specific indices.
In the method encode_token_weights in sd3_clip.py, the empty conditioning is generated as:
out = torch.zeros((1, 77, 4096), device=comfy.model_management.intermediate_device())
pooled = torch.zeros((1, 768 + 1280), device=comfy.model_management.intermediate_device())
This seems correct for pooled, but 'out' would presumably need to be 77 * 2, as per the SD3 model diagram. Similarly, not including padded zeroes for CLIPs/T5 when missing seems wrong.
It would depend on how Stability originally trained the model. Either the unconditional generated this way (presuming blank prompts) wouldn't match the tensor size of a conditional with all 3 text encoders (154 long), or the unconditional would need to be created as zeroes based on the length of the conditional.
The text was updated successfully, but these errors were encountered: