storytold/FineTrainers-Conditioning
2
0
Fork 0
mirror of https://github.com/storytold/FineTrainers-Conditioning.git synced 2026-10-09 00:09:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

The encoder hidden states and encode attention mask sizes do not match up in the validation pipeline.

Browse Source
This commit is contained in:
CrossProduct
2025-01-29 05:24:06 +00:00
parent 4ab3b25c13
commit 707e001f7b
3 changed files with 13 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finetrainers/conditioning/LTXVideoConditionedTransformer3DModel.py
+2 -2
View File
@@ -106,7 +106,7 @@ class LTXVideoConditionedTransformer3DModel(LTXVideoTransformer3DModel):
# inject the condition and the residual then project it into the pretrained proj_in
hidden_states = self.adapter(residual_x=residual_x,
conditioned_x=hidden_states)
# whats this value ?
hidden_states = self.proj_in(hidden_states)
temb, embedded_timestep = self.time_embed(
@@ -134,7 +134,7 @@ class LTXVideoConditionedTransformer3DModel(LTXVideoTransformer3DModel):
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
hidden_states = torch.utils.checkpoint.checkpoint( # crashes here
create_custom_forward(block),
hidden_states,
encoder_hidden_states,
finetrainers/conditioning/conditioned_pipeline.py
+5 -3
View File
@@ -180,7 +180,7 @@ class LTXConditionedPipeline(LTXPipeline):
height,
width,
num_frames,
torch.float32,
torch.bfloat16,
device,
generator,
)
@@ -192,7 +192,7 @@ class LTXConditionedPipeline(LTXPipeline):
patch_size=self.transformer.config.patch_size,
patch_size_t=self.transformer.config.patch_size_t,
device=device,
dtype=torch.float32,
dtype=torch.bfloat16,
generator=generator,
)["latents"]
@@ -202,7 +202,7 @@ class LTXConditionedPipeline(LTXPipeline):
patch_size=self.transformer.config.patch_size,
patch_size_t=self.transformer.config.patch_size_t,
device=device,
dtype=torch.float32,
dtype=torch.bfloat16,
generator=generator,
)["latents"]
@@ -280,6 +280,8 @@ class LTXConditionedPipeline(LTXPipeline):
# change this
# latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
# latent_model_input = latent_model_input.to(prompt_embeds.dtype)
# might be the bug ..
latent_model_input = noisy_latent_tokens
latent_model_input = latent_model_input.to(prompt_embeds.dtype)
finetrainers/trainer.py
+6 -6
View File
@@ -813,7 +813,7 @@ class Trainer:
noisy_latents = noisy_latents.to(target_video_latents.dtype)
# add pose information to both channels
pose_noisy_latents = noisy_latents + pose_video_latents["latents"]
pose_noisy_latents = noisy_latents # + pose_video_latents["latents"]
pose_img_ref_latents = img_refs_latents["latents"] + pose_video_latents["latents"]
# expand channel information # B x 2C latent will be projected to adapter to scale it back to 128d using adapter.
@@ -821,14 +821,14 @@ class Trainer:
# project this turn into patches
# concat tensor
pose_img_ref_latents = post_conditioned_latent_patchify(latents=pose_img_ref_latents,
pose_img_ref_tokens = post_conditioned_latent_patchify(latents=pose_img_ref_latents,
num_frames=latent_conditions["num_frames"],
height=latent_conditions["height"],
width=latent_conditions["width"],
patch_size = 1,
patch_size_t = 1)
# target video as a input residual
noisy_latents_residual = post_conditioned_latent_patchify(latents=noisy_latents,
noisy_residual_tokens = post_conditioned_latent_patchify(latents=noisy_latents,
num_frames=latent_conditions["num_frames"],
height=latent_conditions["height"],
width=latent_conditions["width"],
@@ -836,9 +836,9 @@ class Trainer:
patch_size_t = 1)
# Target Latent Patchified concat tensor
# pose template noisey input [cat] img_ref + pose video
latent_conditions.update({"noisy_latents": pose_img_ref_latents["latents"]})
latent_conditions.update({"noisy_latents": pose_img_ref_tokens["latents"]})
# input video noise at level residual information to adapter
latent_conditions.update({"noisy_latents_residual":noisy_latents_residual["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
@@ -1029,7 +1029,7 @@ class Trainer:
image = load_image(image)
if video is not None:
video = load_video(video)
# loading videos for inference ..
if self.pose_condition:
if pose_video is not None:
pose_video = self.preprocess_condition_video(pose_video,max_num_frames=self.dataset.max_num_frames)