config.fast.yaml

# LTX-2.3 Image+Audio to Video — Fast Mode
#
# Optimized for speed (~10 min). Lower quality than default config.
# Use with: python generate.py --config config.fast.yaml ...

models:
  checkpoint: ./models/ltx-2.3-22b-dev.safetensors
  distilled_lora: ./models/ltx-2.3-22b-distilled-lora-384.safetensors
  distilled_lora_strength: 0.8
  spatial_upscaler: ./models/ltx-2.3-spatial-upscaler-x2-1.1.safetensors
  gemma_root: ~/.cache/huggingface/ltx2_gemma3

generation:
  height: null
  width: null
  num_frames: null
  frame_rate: 25.0
  num_inference_steps: 10     # Fewer steps = faster
  seed: 42
  negative_prompt: "blurry, low quality, distorted, deformed face, bad lip sync, static, frozen, dull expression"
  image_conditioning_strength: 0.7

guidance:
  video_cfg_scale: 3.5
  video_stg_scale: 0.0        # STG OFF = ~33% faster per step
  video_rescale_scale: 0.7    # Upstream default
  video_stg_blocks: []
  video_skip_step: 0
  a2v_guidance_scale: 3.0
  # Audio guider uses defaults internally (audio is frozen during diffusion)

audio:
  start_time: 0.0
  max_duration: null

performance:
  enhance_prompt: false
  streaming_prefetch_count: null
  max_batch_size: 1
  max_pixels: 409600