mirror of
https://github.com/hanzoai/live.git
synced 2026-10-30 18:29:41 +00:00
36 lines
853 B
YAML
36 lines
853 B
YAML
# Self-Forcing DMD Configuration
|
|
# Based on Wan2.1-T2V-1.3B with Self-Forcing training
|
|
|
|
# Model paths will be set by pipeline manager
|
|
model_dir: ""
|
|
checkpoint_path: ""
|
|
text_encoder_path: ""
|
|
|
|
# Video generation settings
|
|
height: 320
|
|
width: 576
|
|
seed: 42
|
|
fps: 24
|
|
|
|
# Self-Forcing specific settings
|
|
chunk_size: 16 # Number of frames to generate per chunk
|
|
use_kv_cache: true # Enable KV caching for faster inference
|
|
autoregressive: true # Enable autoregressive generation
|
|
|
|
# Model architecture
|
|
model_kwargs:
|
|
in_channels: 16
|
|
out_channels: 16
|
|
cross_attention_dim: 4096
|
|
num_layers: 28
|
|
attention_head_dim: 128
|
|
num_attention_heads: 24
|
|
use_linear_projection: true
|
|
upcast_attention: false
|
|
|
|
# DMD (Distribution Matching Distillation) settings
|
|
dmd:
|
|
enabled: true
|
|
num_inference_steps: 4 # Reduced steps for faster inference
|
|
guidance_scale: 7.5
|