n0p/mlx-video
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

improve dev color and quality

Browse Source
This commit is contained in:
Prince Canuma
2026-03-13 09:51:24 +01:00
parent 835ba33202
commit 387d4fc301
1 changed files with 14 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
mlx_video/generate.py
View File

@@ -545,6 +545,7 @@ def denoise_dev(
transformer: LTXModel, transformer: LTXModel,
sigmas: mx.array, sigmas: mx.array,
cfg_scale: float = 4.0, cfg_scale: float = 4.0,
cfg_rescale: float = 0.0,
verbose: bool = True, verbose: bool = True,
state: Optional[LatentState] = None, state: Optional[LatentState] = None,
use_apg: bool = False, use_apg: bool = False,
@@ -554,6 +555,9 @@ def denoise_dev(
"""Run denoising loop for dev pipeline with CFG or APG guidance. """Run denoising loop for dev pipeline with CFG or APG guidance.
Args: Args:
cfg_rescale: Rescale factor for CFG (0.0-1.0). Normalizes guided prediction
variance relative to conditional prediction to reduce over-saturation.
PyTorch default is 0.7. Set to 0.0 to disable.
use_apg: Use Adaptive Projected Guidance instead of standard CFG. use_apg: Use Adaptive Projected Guidance instead of standard CFG.
APG decomposes guidance into parallel/orthogonal components APG decomposes guidance into parallel/orthogonal components
for more stable I2V generation. for more stable I2V generation.
@@ -667,6 +671,14 @@ def denoise_dev(
else: else:
# Standard CFG # Standard CFG
x0_guided_f32 = x0_pos_f32 + (cfg_scale - 1.0) * (x0_pos_f32 - x0_neg_f32) x0_guided_f32 = x0_pos_f32 + (cfg_scale - 1.0) * (x0_pos_f32 - x0_neg_f32)
# Apply CFG rescale if enabled (std-ratio rescaling to reduce over-saturation)
# factor = rescale * (cond_std / pred_std) + (1 - rescale)
# pred = pred * factor
if cfg_rescale > 0.0:
v_factor = x0_pos_f32.std() / (x0_guided_f32.std() + 1e-8)
v_factor = cfg_rescale * v_factor + (1.0 - cfg_rescale)
x0_guided_f32 = x0_guided_f32 * v_factor
else: else:
x0_guided_f32 = x0_pos_f32 x0_guided_f32 = x0_pos_f32
@@ -1381,6 +1393,7 @@ def generate_video(
latents, video_positions, latents, video_positions,
video_embeddings_pos, video_embeddings_neg, video_embeddings_pos, video_embeddings_neg,
transformer, sigmas, cfg_scale=cfg_scale, transformer, sigmas, cfg_scale=cfg_scale,
cfg_rescale=cfg_rescale,
verbose=verbose, state=video_state, verbose=verbose, state=video_state,
use_apg=use_apg, apg_eta=apg_eta, apg_norm_threshold=apg_norm_threshold use_apg=use_apg, apg_eta=apg_eta, apg_norm_threshold=apg_norm_threshold
) )
@@ -1477,6 +1490,7 @@ def generate_video(
latents, positions, latents, positions,
video_embeddings_pos, video_embeddings_neg, video_embeddings_pos, video_embeddings_neg,
transformer, sigmas, cfg_scale=cfg_scale, transformer, sigmas, cfg_scale=cfg_scale,
cfg_rescale=cfg_rescale,
verbose=verbose, state=state1, verbose=verbose, state=state1,
use_apg=use_apg, apg_eta=apg_eta, apg_norm_threshold=apg_norm_threshold use_apg=use_apg, apg_eta=apg_eta, apg_norm_threshold=apg_norm_threshold
) )