How to Use CFG and STG Guidance Scales for Better Generation Quality in LTX‑2

Use CFG (Classifier‑Free Guidance) scales to control prompt fidelity and STG (Spatio‑Temporal Guidance) scales to improve temporal consistency—tune them together with a guidance rescale factor to balance quality and artifacts in LTX‑2 video and audio generation.

LTX‑2, the open‑source video‑and‑audio generation model by Lightricks, exposes two independent guidance mechanisms that operate during diffusion sampling. Understanding how to configure CFG and STG guidance scales lets you trade off prompt adherence against temporal smoothness and artifact suppression. This guide walks through the configuration files, code paths, and practical tuning strategies based on the official LTX‑2 repository.

What Are CFG and STG Guidance Scales?

LTX‑2 separates guidance into two orthogonal controls defined in packages/ltx‑trainer/src/ltx_trainer/config.py:

CFG (Classifier‑Free Guidance) amplifies the difference between conditional and unconditional model predictions. Higher values force stricter prompt adherence but can introduce oversaturation or structural artifacts.

STG (Spatio‑Temporal Guidance) perturbs specific transformer blocks to improve temporal consistency in video and temporal‑frequency coherence in audio. Unlike CFG, STG operates locally on selected blocks rather than globally.

Parameter	Purpose	Default (Validation)	Config Location
video_cfg_scale	Video prompt fidelity	3.0	config.py lines 509‑511
audio_cfg_scale	Audio prompt fidelity	7.0	config.py lines 515‑517
video_stg_scale	Video temporal smoothing	1.0	config.py lines 521‑527
audio_stg_scale	Audio coherence	1.0	config.py lines 527‑531
stg_blocks	Which blocks receive STG	[28]	config.py lines 533‑535
guidance_rescale	Variance interpolation post‑guidance	0.7	config.py lines 38‑44

The guidance_rescale parameter deserves special attention: it interpolates the variance of the guided prediction with the original variance (1.0). Lower values soften the combined effect of CFG and STG to prevent over‑sharpening.

How Guidance Scales Are Applied During Sampling

During validation and inference, the trainer passes guidance parameters to the diffusion denoiser through the MultiModalGuiderParams structure. In packages/ltx‑trainer/src/ltx_trainer/validation_runner.py:

• Video path receives cfg.video_cfg_scale and cfg.video_stg_scale (lines 1232‑1234)
• Audio path receives cfg.audio_cfg_scale and cfg.audio_stg_scale (lines 1254‑1256)

The underlying denoiser euler_cfg_pp_denoising_loop in packages/ltx‑pipelines/src/ltx_pipelines/utils/samplers.py validates that cfg_scale != 1 and stg_scale != 0 before applying each guidance type. This independence means you can enable CFG without STG, STG without CFG, or both together.

Tuning Strategy for CFG and STG Guidance Scales

The interaction between scales follows a predictable pattern:

1. CFG provides global directional push toward conditioned latents
2. STG adds local block‑level perturbation correcting temporal drift
3. Guidance rescale smooths the mixture to prevent runaway variance

Follow this tuning workflow:

• Start with defaults: video_cfg_scale=3.0, audio_cfg_scale=7.0, video_stg_scale=1.0, audio_stg_scale=1.0

• If motion is jittery despite good prompt match: increase video_stg_scale to 1.5–2.0 while keeping CFG modest

• If audio is noisy or misaligned: raise audio_stg_scale and consider lowering audio_cfg_scale to reduce over‑conditioning

• If colors oversaturate or ring artifacts appear: reduce guidance_rescale to 0.5–0.6

Configuration Examples

YAML Configuration

For training runs or custom inference scripts:

validation:
  video_cfg_scale: 3.0
  audio_cfg_scale: 7.0
  video_stg_scale: 1.5
  audio_stg_scale: 1.0
  stg_blocks: [28]
  guidance_rescale: 0.6

Command‑Line Usage

Via the LTX‑Pipelines entry point with flags defined in packages/ltx‑pipelines/src/ltx_pipelines/utils/args.py:

python -m ltx_pipelines.ti2vid_two_stages \
    --video-cfg-guidance-scale 4.0 \
    --audio-cfg-guidance-scale 8.0 \
    --video-stg-guidance-scale 1.8 \
    --audio-stg-guidance-scale 1.0 \
    --stg-blocks 28 \
    --guidance-rescale 0.5 \
    ...

The argument definitions reside at lines 948‑962 for CFG defaults and lines 961‑966 for STG defaults in args.py.

Programmatic Construction

Building a MultiModalGuiderParams object directly:

from ltx_pipelines.utils.deniosers import MultiModalGuiderParams

guider = MultiModalGuiderParams(
    cfg_scale=4.0,
    stg_scale=1.8,
    modality_scale=1.0,
)

# Pass guider to the sampler's denoise method

Core Implementation Files

File	Role
packages/ltx‑trainer/src/ltx_trainer/config.py	Hyper‑parameter definitions
packages/ltx‑trainer/src/ltx_trainer/validation_runner.py	Scale application during inference
packages/ltx‑pipelines/src/ltx_pipelines/utils/args.py	CLI argument exposure
packages/ltx‑core/src/ltx_core/components/guiders.py	Mathematical combination logic
packages/ltx‑pipelines/src/ltx_pipelines/utils/samplers.py	Denoising loop with scale checks

Summary

• CFG scales (video_cfg_scale, audio_cfg_scale) control global prompt fidelity—higher values increase adherence but risk artifacts
• STG scales (video_stg_scale, audio_stg_scale) enable block‑level temporal correction via transformer perturbation
• STG blocks (stg_blocks) select which layers receive guidance—the default [28] works broadly
• Guidance rescale (guidance_rescale) softens combined guidance to prevent oversaturation
• All parameters can be set via YAML, CLI flags in args.py, or MultiModalGuiderParams objects
• The denoiser in samplers.py applies guidance independently, allowing flexible combinations

Frequently Asked Questions

What is the difference between CFG and STG in LTX‑2?

CFG operates globally on the entire model prediction, amplifying the gap between conditional and unconditional outputs. STG operates locally on specified transformer blocks, perturbing their activations to improve temporal structure. They address different failure modes: CFG improves prompt alignment, while STG reduces temporal jitter and frequency drift.

When should I increase STG scale instead of CFG scale?

Increase STG when your generation follows the prompt correctly but shows motion inconsistency, frame flicker, or audio desynchronization. STG specifically targets temporal coherence without the artifact risks of pushing CFG higher. The default stg_blocks = [28] targets late‑stage transformer layers where temporal structure crystallizes.

How does guidance rescale prevent artifacts?

After applying CFG and STG, the predicted noise variance can become artificially inflated, causing oversaturated colors and ringing. The guidance_rescale parameter interpolates between the guided variance and the original variance (1.0). Values below 0.7 progressively soften this effect, with 0.5–0.6 typically effective for high‑guidance configurations.

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