storytold/FineTrainers-Conditioning
2
0
Fork 0
mirror of https://github.com/storytold/FineTrainers-Conditioning.git synced 2026-10-18 07:33:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Full Finetune should work with this adapter removed the residual_x

Browse Source
This commit is contained in:
CrossProduct
2025-01-30 00:14:44 +00:00
parent e399684065
commit ea3cd5c753
3 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finetrainers/conditioning/LTXVideoConditionedTransformer3DModel.py
+1 -2
View File
@@ -93,8 +93,7 @@ class LTXVideoConditionedTransformer3DModel(LTXVideoTransformer3DModel):
return_dict: bool = True, return_dict: bool = True,
residual_x: torch.Tensor = None residual_x: torch.Tensor = None
) -> torch.Tensor: ) -> 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) 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 # convert encoder_attention_mask to a bias the same way we do for attention_mask
finetrainers/ltx_video/full_finetune_condition.py
+4
View File
@@ -119,6 +119,10 @@ def conditioned_forward_pass(
# encoder_hidden_states=prompt_embeds, # encoder_hidden_states=prompt_embeds,
# timestep=timesteps, # timestep=timesteps,
# encoder_attention_mask=prompt_attention_mask, # encoder_attention_mask=prompt_attention_mask,
if noisy_latents_residual == None:
print("Residual Not included in the calculation.")
denoised_latents = transformer( denoised_latents = transformer(
hidden_states=noisy_latents, hidden_states=noisy_latents,
encoder_hidden_states=prompt_embeds, encoder_hidden_states=prompt_embeds,
finetrainers/trainer.py
+1 -1
View File
@@ -840,7 +840,7 @@ class Trainer:
# That dict says latents but actually tokens. # That dict says latents but actually tokens.
latent_conditions.update({"noisy_latents": condition_tokens["latents"]}) latent_conditions.update({"noisy_latents": condition_tokens["latents"]})
# input video noise at level residual information to adapter # 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: else:
# Default to flow-matching noise addition # Default to flow-matching noise addition
noisy_latents = (1.0 - sigmas) * latent_conditions["latents"] + sigmas * noise noisy_latents = (1.0 - sigmas) * latent_conditions["latents"] + sigmas * noise