storytold/FineTrainers-Conditioning
2
0
Fork 0
mirror of https://github.com/storytold/FineTrainers-Conditioning.git synced 2026-10-22 08:05:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Moved the norm back to prepare latents step.

Browse Source 
Now its time to add the adapter layer.
This commit is contained in:
CrossProduct
2025-01-18 04:18:25 +00:00
parent 44df9e3bce
commit 2720d24f62
3 changed files with 4 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finetrainers/trainer.py
+1 -11
View File
@@ -799,14 +799,6 @@ class Trainer:
if self.pose_condition:
# condition with the pose information
# grab noise pre patchification shape
noise = torch.randn(
latent_conditions["latents"].shape,
generator=self.state.generator,
device=accelerator.device,
dtype=weight_dtype,
)
# normalize and patchify
# create noise for latent to feed diffusion transformer to get pred noise
noisy_latents = (1.0 - sigmas) * latent_conditions["latents"] + sigmas * noise
@@ -819,15 +811,13 @@ class Trainer:
# project this turn into patches
latent_final = post_conditioned_latent_patchify(latents=latent_final,
latents_mean=(latent_conditions["mean"] + pose_video_latents["mean"] + img_refs_latents["mean"]) / 3.0,
latents_std=(latent_conditions["std"] + pose_video_latents["std"] + img_refs_latents["std"]) / 3.0,
num_frames=latent_conditions["num_frames"],
height=latent_conditions["height"],
width=latent_conditions["width"],
patch_size = 1,
patch_size_t = 1)
latent_conditions.update({"noisy_latents": latent_final})
latent_conditions.update({"noisy_latents": latent_final["latents"]})
else:
# Default to flow-matching noise addition
noisy_latents = (1.0 - sigmas) * latent_conditions["latents"] + sigmas * noise