LTX-2 Guidance Parameters: CFG, STG, and Modality Guidance Explained

LTX-2 supports three distinct guidance signals—Classifier-Free Guidance (CFG), Spatio-Temporal Guidance (STG), and modality-specific guidance—that you can independently tune for video and audio generation via the MultiModalGuiderParams dataclass in guiders.py.

The Lightricks LTX-2 diffusion model exposes fine-grained control over generation quality through its multimodal guider system. Each modality (video and audio) receives its own guidance configuration, allowing precise adjustment of prompt adherence, temporal consistency, and cross-modal synchronization. This article breaks down every LTX-2 guidance parameter based on the source code in the official repository.

Classifier-Free Guidance (cfg_scale)

CFG is the standard technique for amplifying text prompt influence. In LTX-2, cfg_scale controls the strength of the (conditional - unconditional) direction computed during each denoising step.

Aspect	Details
Typical range	2.0 – 5.0
Default	~3.0
Disable	Set to 1.0

Higher values force stricter prompt adherence but can reduce motion naturalness. The video and audio guiders accept independent cfg_scale values, so you might prioritize sharper audio generation while keeping video more fluid.

from ltx_core.components.guiders import MultiModalGuiderParams

# Strong text guidance for audio, moderate for video

audio_guider = MultiModalGuiderParams(cfg_scale=7.0, ...)
video_guider = MultiModalGuiderParams(cfg_scale=3.0, ...)

Spatio-Temporal Guidance (stg_scale and stg_blocks)

STG improves temporal coherence by perturbing specific transformer blocks and steering away from the perturbed prediction. This mechanism is unique to LTX-2's video generation pipeline.

stg_scale

Controls the perturbation guidance strength:

Aspect	Details
Typical range	0.5 – 1.5
Default	~1.0
Disable	Set to 0.0

stg_blocks

Specifies which transformer blocks receive the perturbation. The default [29] targets the final block in the architecture.

# Enable STG on the last transformer block

video_guider = MultiModalGuiderParams(
    stg_scale=1.0,
    stg_blocks=[29],  # Perturb block 29 only

)

To completely disable STG, provide an empty list: stg_blocks=[].

Modality Guidance (modality_scale)

The modality_scale parameter implements modality-specific CFG for joint audio-video generation. It pushes the two streams toward synchronized results, preventing audio-visual drift.

Aspect	Details
Typical range	1.0 – 5.0
Default	~3.0 (joint generation)
Disable	Set to 1.0

This parameter is particularly critical when generating matched video and audio simultaneously. Higher values enforce tighter coupling between modalities.

Rescaling and Performance Controls

rescale_scale

Prevents over-saturation by scaling guided predictions to match the variance of conditional predictions:

Aspect	Details
Typical range	0.5 – 0.7
Default	~0.7
Disable	Set to 0.0

skip_step

Accelerator for inference. Skips guidance computation every N steps:

Aspect	Details
Default	0 (no skipping)
Use case	Speed-critical applications

Complete Configuration Example

from ltx_core.components.guiders import MultiModalGuiderParams
from ltx_pipelines.ti2vid_one_stage import run_pipeline

# Video: balanced guidance with temporal coherence

video_guider = MultiModalGuiderParams(
    cfg_scale=3.0,
    stg_scale=1.0,
    rescale_scale=0.7,
    modality_scale=3.0,
    stg_blocks=[29],
    skip_step=0,
)

# Audio: stronger prompt adherence

audio_guider = MultiModalGuiderParams(
    cfg_scale=7.0,
    stg_scale=1.0,
    rescale_scale=0.7,
    modality_scale=3.0,
    stg_blocks=[29],
    skip_step=0,
)

run_pipeline(
    video_guider_params=video_guider,
    audio_guider_params=audio_guider,
    # ... additional arguments

)

Key Implementation Files

The guidance system spans several source files in the LTX-2 repository:

• packages/ltx-core/src/ltx_core/components/guiders.py — Core MultiModalGuiderParams dataclass and guidance merging logic
• packages/ltx-pipelines/docs/multimodal-guidance.md — Official parameter documentation
• packages/ltx-trainer/src/ltx_trainer/config.py — CLI flags for training validation
• packages/ltx-pipelines/src/ltx_pipelines/utils/args.py — Command-line argument parsing
• packages/ltx-pipelines/src/ltx_pipelines/ti2vid_one_stage.py — Pipeline integration example

Summary

• CFG (cfg_scale) controls text-prompt strength per modality—set to 1.0 to disable
• STG (stg_scale, stg_blocks) improves temporal coherence via block perturbation—disable with stg_scale=0.0 or stg_blocks=[]
• Modality guidance (modality_scale) synchronizes audio-video generation—critical for joint generation tasks
• Rescaling (rescale_scale) prevents saturation artifacts
• Skip steps (skip_step) trades quality for inference speed

All parameters live in MultiModalGuiderParams as defined in guiders.py, with independent instances for video and audio.

Frequently Asked Questions

What is the difference between CFG and modality guidance in LTX-2?

CFG (cfg_scale) amplifies text prompt influence by contrasting conditional and unconditional predictions. Modality guidance (modality_scale) specifically targets multi-modal synchronization, pushing video and audio streams toward coherent joint results. You need both: CFG for prompt fidelity, modality guidance for audio-visual alignment.

How do I completely disable STG in LTX-2?

Set stg_scale=0.0 or provide an empty list for stg_blocks=[]. The latter is more explicit and prevents any STG-related computation overhead. Both approaches are valid; the source code in guiders.py checks these conditions to skip STG application.

Can I use different guidance scales for video and audio generation?

Yes—this is the default pattern. Instantiate separate MultiModalGuiderParams objects and pass them as video_guider_params and audio_guider_params to your pipeline. The LTX-2 architecture treats each modality independently, allowing you to prioritize audio clarity (cfg_scale=7.0) while keeping video motion natural (cfg_scale=3.0).

Where are the default guidance values defined in the LTX-2 repository?

Default values reside in packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py. These serve as the reference configuration that the pipelines use when you don't override parameters explicitly.

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