xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2026-09-09 05:17:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

[HiDream] Hard-cap T5 embeddings to 128 length

Browse Source
This commit is contained in:
tdrussell 2025-04-16 13:42:36 -05:00
parent b4dc03ad76
commit bc388f8ffa
1 changed files with 2 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
comfy/text_encoders/hidream.py
View File

@ -108,6 +108,8 @@ class HiDreamTEModel(torch.nn.Module):
if self.t5xxl is not None:
t5_output = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
t5_out, t5_pooled = t5_output[:2]
# Quality rapidly degrades if T5 embeddings are >128 length. Llama3 is not as bad, so we allow that one to be longer.
t5_out = t5_out[:, :128, :]
if self.llama is not None:
ll_output = self.llama.encode_token_weights(token_weight_pairs_llama)