FillipandClaude Opus 4.6
b17d44cd6c
Single-knob motion intensity control for LTX-Video. Drives: - Conditioning frame_rate (RoPE temporal scale) — primary motion lever - VAE decoder noise (decode_noise_scale + decode_timestep) — per-frame stochasticity Drop-in replacement for built-in LTXVConditioning, with motion boost on top. Self-contained — no dependencies on other LTX custom node packs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
39 lines
1.5 KiB
Python
39 lines
1.5 KiB
Python
"""VAE patching helpers for LTX-Video."""
|
|
|
|
import copy
|
|
import logging
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def clone_vae_with_decoder_noise(vae, decode_noise_scale: float, decode_timestep: float):
|
|
"""Return a clone of the VAE with decode_noise_scale and decode_timestep
|
|
overridden on the inner causal_video_autoencoder.
|
|
|
|
Falls back to returning the original VAE (with a warning) if it doesn't
|
|
look like an LTX VAE.
|
|
"""
|
|
inner = getattr(vae, "first_stage_model", None)
|
|
if inner is None:
|
|
logger.warning("FL_LTX_MotionBoost: VAE has no first_stage_model — passthrough.")
|
|
return vae
|
|
|
|
has_decode_noise = hasattr(inner, "decode_noise_scale")
|
|
has_decode_timestep = hasattr(inner, "decode_timestep")
|
|
if not (has_decode_noise and has_decode_timestep):
|
|
logger.warning(
|
|
"FL_LTX_MotionBoost: VAE first_stage_model is not an LTX CausalVideoAutoencoder "
|
|
"(missing decode_noise_scale / decode_timestep attributes) — passthrough."
|
|
)
|
|
return vae
|
|
|
|
# Shallow clone the outer wrapper, then deepcopy just the inner model so we
|
|
# don't mutate the user's existing VAE reference. The outer wrapper's other
|
|
# attributes (memory_used_decode, dtype, etc.) are kept by reference.
|
|
vae_clone = copy.copy(vae)
|
|
inner_clone = copy.copy(inner)
|
|
inner_clone.decode_noise_scale = float(decode_noise_scale)
|
|
inner_clone.decode_timestep = float(decode_timestep)
|
|
vae_clone.first_stage_model = inner_clone
|
|
return vae_clone
|