Default CFG Scales for Video and Audio in LTX-2: Complete Technical Reference

LTX-2 uses a default CFG scale of 3.0 for video generation and 7.0 for audio generation, defined in the pipeline constants and exposed via the command-line interface.

LTX-2, Lightricks' open-source video generation model, relies on classifier-free guidance (CFG) to control how closely outputs follow text prompts. The default CFG scales for video and audio in LTX-2 are hardcoded in the pipeline utilities and propagate through the argument parser to provide out-of-the-box generation behavior. These values determine the strength of guidance applied during the diffusion process for each modality.

Default CFG Scale Values in LTX-2

The repository defines distinct defaults for video and audio modalities to account for their different generation characteristics.

• Video: 3.0 (defined at line 40)
• Audio: 7.0 (defined at line 50)

These values represent the baseline guidance strength before any user overrides are applied.

Where Default CFG Scales Are Defined in the Source Code

Pipeline Constants (constants.py)

The canonical values reside in packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py. This file contains the VIDEO_DEFAULT_CFG_SCALE and AUDIO_DEFAULT_CFG_SCALE constants, which initialize the video_guider.cfg_scale and audio_guider.cfg_scale attributes respectively.

Command-Line Argument Parser (args.py)

The packages/ltx-pipelines/src/ltx_pipelines/utils/args.py file constructs the CLI parser by referencing these guider instances. When default_1_stage_arg_parser() builds the argument list, it pulls the current values from video_guider.cfg_scale (3.0) and audio_guider.cfg_scale (7.0) as the defaults for the --video-cfg-guidance-scale and --audio-cfg-guidance-scale parameters.

How to Check Current CFG Scale Defaults

You can inspect the active defaults programmatically using the argument parser:

from ltx_pipelines.utils.args import default_1_stage_arg_parser
from ltx_pipelines.utils.constants import VIDEO_DEFAULT_CFG_SCALE, AUDIO_DEFAULT_CFG_SCALE

parser = default_1_stage_arg_parser()
args = parser.parse_args([])  # No CLI overrides

print(f"Video CFG scale: {args.video_cfg_guidance_scale}")   # Output: 3.0

print(f"Audio CFG scale: {args.audio_cfg_guidance_scale}")   # Output: 7.0

How to Override Default CFG Scales

Override these values at runtime using command-line flags:

python -m ltx_pipelines.ti2vid_two_stages \
    --reference-video input.mp4 \
    --video-cfg-guidance-scale 5.0 \
    --audio-cfg-guidance-scale 9.0

The values 5.0 and 9.0 replace the defaults (3.0 and 7.0) for this specific generation run.

Summary

• Video default CFG scale: 3.0, optimized for temporal consistency and smooth motion
• Audio default CFG scale: 7.0, providing stronger prompt adherence for sound generation
• Source location: packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py (lines 40 and 50)
• CLI exposure: packages/ltx-pipelines/src/ltx_pipelines/utils/args.py via --video-cfg-guidance-scale and --audio-cfg-guidance-scale
• Modification: Pass explicit values via command-line arguments or edit the constants file directly

Frequently Asked Questions

What is the default CFG scale for video in LTX-2?

The default CFG scale for video generation in LTX-2 is 3.0, defined in the VIDEO_DEFAULT_CFG_SCALE constant within packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py. This lower value helps maintain natural motion and temporal coherence in generated video sequences.

What is the default CFG scale for audio in LTX-2?

The default CFG scale for audio generation in LTX-2 is 7.0, specified in the AUDIO_DEFAULT_CFG_SCALE constant in the same constants file. This higher value ensures stronger adherence to text prompts, which is necessary for generating audio that accurately matches descriptive content.

How do I change the CFG scale in LTX-2?

You can change the CFG scale by passing the --video-cfg-guidance-scale or --audio-cfg-guidance-scale arguments when running pipeline scripts such as ti2vid_two_stages.py. For example, use --video-cfg-guidance-scale 5.0 to increase video guidance strength from the default 3.0 to 5.0 for a specific generation task.

Why are the video and audio CFG scales different?

The video and audio CFG scales differ because audio generation typically requires stronger guidance (7.0) to produce coherent, text-aligned sounds, while video generation uses a lower value (3.0) to prioritize smoother motion and temporal consistency over strict literal interpretation of prompts. This separation allows each modality to optimize for its specific perceptual requirements.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}