The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Lower ltxv text encoder vram use. (#11713)
comfyanonymous34751fe9f9ad
diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.pyindex 130ebaeae..dc0694e0e 100644--- a/comfy/text_encoders/lt.py+++ b/comfy/text_encoders/lt.py@@ -98,10 +98,13 @@ class LTXAVTEModel(torch.nn.Module):out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)out_device = out.device+ if comfy.model_management.should_use_bf16(self.execution_device):+ out = out.to(device=self.execution_device, dtype=torch.bfloat16)out = out.movedim(1, -1).to(self.execution_device)out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)out = out.reshape((out.shape[0], out.shape[1], -1))out = self.text_embedding_projection(out)+ out = out.float()out_vid = self.video_embeddings_connector(out)[0]out_audio = self.audio_embeddings_connector(out)[0]out = torch.concat((out_vid, out_audio), dim=-1)