jbilcke-hf HF Staff commited on
Commit
5d15b02
·
verified ·
1 Parent(s): 0e02098

Lightricks have updated their pipeline

Browse files

Applying same changes as: https://github.com/Lightricks/LTX-Video/commit/93af6864b942e931908fcf1a3f04feebf97dbb36#diff-e4522277a659ca45d158c90142b86a4518428bda49b31d8027e7f5482039a74f

Files changed (1) hide show
  1. handler.py +3 -10
handler.py CHANGED
@@ -89,10 +89,9 @@ class GenerationConfig:
89
  quality: int = 18
90
 
91
  # STG (Spatiotemporal Guidance) settings
92
- stg_scale: float = 1.0
93
- stg_rescale: float = 0.7
94
  stg_mode: str = "attention_values" # Can be "attention_values", "attention_skip", "residual", or "transformer_block"
95
- stg_skip_layers: str = "19" # Comma-separated list of layers to block for spatiotemporal guidance
96
 
97
  # VAE noise augmentation
98
  decode_timestep: float = 0.05
@@ -147,10 +146,6 @@ class GenerationConfig:
147
  elif self.stg_mode.lower() == "stg_t" or self.stg_mode.lower() == "transformer_block":
148
  self.stg_mode = "transformer_block"
149
 
150
- # Convert STG skip layers from string to list of integers
151
- if isinstance(self.stg_skip_layers, str):
152
- self.stg_skip_layers = [int(x.strip()) for x in self.stg_skip_layers.split(",")]
153
-
154
  # Check if we should enhance the prompt
155
  if self.enhance_prompt and self.prompt:
156
  prompt_word_count = len(self.prompt.split())
@@ -407,7 +402,7 @@ class EndpointHandler:
407
  width=768, # Using smaller resolution for faster warm-up
408
  height=416,
409
  num_frames=33, # Just enough frames for a valid video
410
- guidance_scale=3.0,
411
  num_inference_steps=4, # Fewer steps for faster warm-up
412
  seed=42, # Fixed seed for consistent warm-up
413
  fps=16, # Lower FPS for faster processing
@@ -505,7 +500,6 @@ class EndpointHandler:
505
  stg_scale=params.get("stg_scale", GenerationConfig.stg_scale),
506
  stg_rescale=params.get("stg_rescale", GenerationConfig.stg_rescale),
507
  stg_mode=params.get("stg_mode", GenerationConfig.stg_mode),
508
- stg_skip_layers=params.get("stg_skip_layers", GenerationConfig.stg_skip_layers),
509
 
510
  # VAE noise settings
511
  decode_timestep=params.get("decode_timestep", GenerationConfig.decode_timestep),
@@ -589,7 +583,6 @@ class EndpointHandler:
589
  generator=generator,
590
  output_type="pt", # Return as PyTorch tensor
591
  skip_layer_strategy=skip_layer_strategy,
592
- skip_block_list=config.stg_skip_layers,
593
  stg_scale=config.stg_scale,
594
  do_rescaling=config.stg_rescale != 1.0,
595
  rescaling_scale=config.stg_rescale,
 
89
  quality: int = 18
90
 
91
  # STG (Spatiotemporal Guidance) settings
92
+ stg_scale: float = 0.0
93
+ stg_rescale: float = 1.0
94
  stg_mode: str = "attention_values" # Can be "attention_values", "attention_skip", "residual", or "transformer_block"
 
95
 
96
  # VAE noise augmentation
97
  decode_timestep: float = 0.05
 
146
  elif self.stg_mode.lower() == "stg_t" or self.stg_mode.lower() == "transformer_block":
147
  self.stg_mode = "transformer_block"
148
 
 
 
 
 
149
  # Check if we should enhance the prompt
150
  if self.enhance_prompt and self.prompt:
151
  prompt_word_count = len(self.prompt.split())
 
402
  width=768, # Using smaller resolution for faster warm-up
403
  height=416,
404
  num_frames=33, # Just enough frames for a valid video
405
+ guidance_scale=1.0,
406
  num_inference_steps=4, # Fewer steps for faster warm-up
407
  seed=42, # Fixed seed for consistent warm-up
408
  fps=16, # Lower FPS for faster processing
 
500
  stg_scale=params.get("stg_scale", GenerationConfig.stg_scale),
501
  stg_rescale=params.get("stg_rescale", GenerationConfig.stg_rescale),
502
  stg_mode=params.get("stg_mode", GenerationConfig.stg_mode),
 
503
 
504
  # VAE noise settings
505
  decode_timestep=params.get("decode_timestep", GenerationConfig.decode_timestep),
 
583
  generator=generator,
584
  output_type="pt", # Return as PyTorch tensor
585
  skip_layer_strategy=skip_layer_strategy,
 
586
  stg_scale=config.stg_scale,
587
  do_rescaling=config.stg_rescale != 1.0,
588
  rescaling_scale=config.stg_rescale,