Full Finetune should work with this adapter removed the residual_x

finetrainers/conditioning/LTXVideoConditionedTransformer3DModel.py

@@ -93,8 +93,7 @@ class LTXVideoConditionedTransformer3DModel(LTXVideoTransformer3DModel):
             return_dict: bool = True,
             residual_x: torch.Tensor = None
         ) -> torch.Tensor:
-            if residual_x == None:
-                print("Residual Not included in the calculation.")
+           
             image_rotary_emb = self.rope(hidden_states, num_frames, height, width, rope_interpolation_scale)
 
             # convert encoder_attention_mask to a bias the same way we do for attention_mask

finetrainers/ltx_video/full_finetune_condition.py

@@ -119,6 +119,10 @@ def conditioned_forward_pass(
         # encoder_hidden_states=prompt_embeds,
         # timestep=timesteps,
         # encoder_attention_mask=prompt_attention_mask,
+
+    if noisy_latents_residual == None:
+        print("Residual Not included in the calculation.")
+
     denoised_latents = transformer(
         hidden_states=noisy_latents,
         encoder_hidden_states=prompt_embeds,

finetrainers/trainer.py

@@ -840,7 +840,7 @@ class Trainer:
                             # That dict says latents but actually tokens.
                             latent_conditions.update({"noisy_latents": condition_tokens["latents"]})
                             # input video noise at level residual information to adapter
-                            latent_conditions.update({"noisy_latents_residual":noisy_residual_tokens["latents"]})
+                            # latent_conditions.update({"noisy_latents_residual":noisy_residual_tokens["latents"]})
                         else:
                             # Default to flow-matching noise addition
                             noisy_latents = (1.0 - sigmas) * latent_conditions["latents"] + sigmas * noise