How Self-Targeted Guidance (STG) Functions in LTX-2 and Choosing Optimal `stg_scale` Values

Self-Targeted Guidance (STG) steers the diffusion denoising process in LTX-2 by perturbing self-attention layers in specified transformer blocks, with stg_scale controlling the blend between conditioned and perturbed predictions.

STG is a modality-specific guidance mechanism implemented in the Lightricks/LTX-2 repository that improves temporal consistency and cross-modal alignment. Unlike traditional classifier-free guidance, STG creates a "perturbed" unconditioned prediction by selectively masking self-attention, then blends this with the standard conditioned output. This article explains the internal mechanics based on the actual source code and provides practical guidance for tuning stg_scale values.

Where STG Lives in the LTX-2 Architecture

STG parameters are defined in MultiModalGuiderParams within packages/ltx-core/src/ltx_core/components/guiders.py:

• stg_scale — the guidance strength multiplier
• stg_blocks — list of transformer block indices to perturb

The core blending formula appears at lines 62-66 of guiders.py:

# Simplified view of the calculation

guided_output = cond + cfg_scale * (cond - uncond_text) + stg_scale * (cond - uncond_perturbed)

The uncond_perturbed term is what distinguishes STG from standard CFG. It comes from a forward pass where self-attention is masked in the blocks specified by stg_blocks.

How the Self-Attention Perturbation Works

The Skip-Self-Attention Mechanism

The perturbation type is defined in packages/ltx-core/src/ltx_core/guidance/perturbations.py (lines 13-16):

class PerturbationType(Enum):
    SKIP_VIDEO_SELF_ATTN = "skip_video_self_attn"
    SKIP_AUDIO_SELF_ATTN = "skip_audio_self_attn"
    ZERO_VIDEO_SELF_ATTN = "zero_video_self_attn"  # Alternative perturbation

    ZERO_AUDIO_SELF_ATTN = "zero_audio_self_attn"

During validation inference, the ValidationRunner constructs a BatchedPerturbationConfig that instructs the transformer to skip self-attention for the unconditioned pass. From packages/ltx-trainer/src/ltx_trainer/validation_runner.py (lines 1288-1291):

if video_enabled:
    perturbations.append(
        Perturbation(type=PerturbationType.SKIP_VIDEO_SELF_ATTN, blocks=stg_blocks)
    )
if audio_enabled:
    perturbations.append(
        Perturbation(type=PerturbationType.SKIP_AUDIO_SELF_ATTN, blocks=stg_blocks)
    )

This creates a batch where:

• Conditioned sample: Full self-attention in all blocks
• Unconditioned sample: Self-attention masked in stg_blocks
• Perturbed unconditioned sample: Same latent as unconditioned, but with degraded self-attention

The difference between conditioned and perturbed predictions reveals how much the model relies on self-attention for coherence—STG amplifies this signal.

Default STG Configuration Values

Parameter	Default	Source Location
video_stg_scale	1.0	packages/ltx-trainer/src/ltx_trainer/config.py lines 521-525
audio_stg_scale	1.0	packages/ltx-trainer/src/ltx_trainer/config.py lines 527-531
stg_blocks (LTX-2.3)	[28]	packages/ltx-pipelines/CLAUDE.md line 34
stg_blocks (original LTX-2)	[29]	packages/ltx-pipelines/CLAUDE.md line 34

CLI flags exposing these settings are defined in packages/ltx-pipelines/src/ltx_pipelines/utils/args.py (lines 59-70):

• --video-stg-guidance-scale
• --audio-stg-guidance-scale
• --video-stg-blocks
• --audio-stg-blocks

Optimal stg_scale Values: Practical Guidance

Default Behavior (stg_scale = 1.0)

The default value of 1.0 provides balanced guidance. According to the CLAUDE.md documentation and config.py defaults, this setting:

• Enhances temporal consistency in video generation
• Improves audio-visual alignment in multimodal outputs
• Avoids the artifacts common with excessive guidance

Lower Values (0.0 to 0.5)

Reduce stg_scale when:

• Flickering or temporal inconsistency appears — lower values reduce the "forced" coherence that can cause oscillation
• Running low-VRAM or fast inference — the perturbed pass requires additional computation; lowering the scale or setting to 0.0 skips this overhead
• Seeking more "natural" diffusion trajectories — less intervention preserves the model's native noise-to-data path

Setting stg_scale = 0.0 completely disables STG without affecting standard classifier-free guidance.

Higher Values (1.0 to 2.0)

Increase stg_scale when:

• Strong temporal anchoring is needed — audio-driven video synthesis benefits from aggressive STG to lock lip-sync and motion to audio cues
• Cross-modal consistency is critical — values of 1.5 to 2.0 can reinforce alignment between video and audio streams

Avoid exceeding 2.0. Higher values typically introduce:

• Over-sharpened edges and plastic appearance in video
• Audio distortion or "metallic" artifacts in generated sound
• Reduced diversity in outputs

Special Case: HQ Pipelines

The High-Quality pipelines (*_hq variants) deliberately disable STG. From packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py (lines 101-113), these pipelines set:

video_stg_scale=0.0,
audio_stg_scale=0.0,

The second-order samplers used in HQ pipelines provide fine-grained control through their internal schedules; STG interference would disrupt this precision.

Practical Configuration Examples

YAML Configuration File

# configs/v2a_lora.yaml — video-to-audio with tuned STG

video_stg_scale: 1.0      # Enable video STG for motion coherence

audio_stg_scale: 0.0      # Disable audio STG (audio is the target, not source)

stg_blocks: [28]          # Perturb block 28 (LTX-2.3 architecture)

Command-Line Override

uv run python scripts/train.py \
    configs/t2v_lora.yaml \
    --video-stg-guidance-scale 0.8 \
    --video-stg-blocks 27 28 29

This applies reduced STG strength across three consecutive blocks for smoother transitions.

Programmatic MultiModalGuider Creation

from ltx_core.components.guiders import MultiModalGuiderParams, MultiModalGuider

params = MultiModalGuiderParams(
    cfg_scale=3.0,
    stg_scale=1.2,          # Stronger STG than default

    stg_blocks=[28],        # Single-block perturbation

    modality_scale=1.0,
)
guider = MultiModalGuider(params=params)

guided = guider.calculate(
    cond=cond_pred,
    uncond_text=uncond_pred,
    uncond_perturbed=perturbed_pred,
    uncond_modality=uncond_pred,
)

Debugging Perturbation Masks

from ltx_core.guidance.perturbations import (
    PerturbationType,
    PerturbationConfig,
    BatchedPerturbationConfig,
    Perturbation,
)

# Configure skip-self-attention on block 28 for video

pert = Perturbation(
    type=PerturbationType.SKIP_VIDEO_SELF_ATTN,
    blocks=[28]
)
cfg = PerturbationConfig(perturbations=[pert])
batched_cfg = BatchedPerturbationConfig([cfg], num_blocks=32)

# Inspect the actual mask tensor

mask = batched_cfg.mask(PerturbationType.SKIP_VIDEO_SELF_ATTN, block=28)
print(mask.shape)   # (batch_size, 1, 1)

print(mask)         # 0.0 where attention is masked, 1.0 otherwise

Key Source Files for STG Implementation

File Path	Purpose
packages/ltx-core/src/ltx_core/components/guiders.py	MultiModalGuiderParams, MultiModalGuider, and the STG blending formula
packages/ltx-core/src/ltx_core/guidance/perturbations.py	PerturbationType enum and perturbation configuration classes
packages/ltx-trainer/src/ltx_trainer/config.py	User-facing video_stg_scale, audio_stg_scale, and stg_blocks defaults
packages/ltx-trainer/src/ltx_trainer/validation_runner.py	BatchedPerturbationConfig construction for inference
packages/ltx-pipelines/src/ltx_pipelines/utils/args.py	CLI argument definitions for STG parameters
packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py	HQ pipeline defaults that disable STG

Summary

• STG perturbs self-attention in specified transformer blocks to create a degraded "unconditioned" prediction, then blends this with the conditioned output using stg_scale
• Default stg_scale = 1.0 works well for most generation tasks in both video and audio modalities
• Lower values (0.0–0.5) reduce artifacts and computation; 0.0 fully disables STG
• Higher values (1.0–2.0) strengthen temporal consistency but risk quality degradation above 2.0
• HQ pipelines disable STG entirely to avoid interference with second-order samplers
• Block selection matters: [28] for LTX-2.3, [29] for original LTX-2

Frequently Asked Questions

What happens when stg_scale is set to 0.0?

Setting stg_scale = 0.0 removes the STG term from the guidance calculation, reducing the formula to standard classifier-free guidance. The perturbed pass still runs but contributes zero weight. This is how HQ pipelines operate according to constants.py lines 101-113.

Why do different LTX-2 versions use different stg_blocks?

The LTX-2.3 architecture places critical self-attention mechanisms at block 28, while the original LTX-2 uses block 29. These specific blocks were identified as optimal intervention points through architectural analysis. The CLAUDE.md documentation at line 34 documents these version-specific defaults.

Can STG be applied to only one modality?

Yes. The configuration system separates video_stg_scale and audio_stg_scale, allowing independent control. Set the unused modality to 0.0 while keeping the target modality at your desired strength. This is common in video-to-audio tasks where audio STG is disabled.

How does STG differ from standard classifier-free guidance?

Standard CFG compares conditioned and unconditioned predictions. STG introduces a third "perturbed unconditioned" prediction where self-attention is masked. The difference between unconditioned and perturbed reveals self-attention's contribution to coherence, which STG amplifies through stg_scale.

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