EMAG (Exponential Moving Average Guidance) Guider Node is a drop-in replacement for CFGGuider that applies EMA-based attention perturbation to create hard negatives for improved guidance.

Based on: "EMAG: Exponential Moving Average Guidance for Diffusion Models" Paper equations implemented: Eq. 12 (EMA), Eq. 15 (EMAG update), Eq. 16 (CFG with EMAG)

EMAG + SyncCFG Hybrid Guider Node combines Exponential Moving Average Guidance with Synchronization-Enhanced CFG for
improved audio-video generation alignment.

Modes:

• EMAG_ONLY: Standard EMAG guidance (original behavior)
• SYNCCFG_ONLY: Pure SyncCFG without EMA perturbation
• HYBRID: EMAG perturbation + SyncCFG guidance structure

Based on:

• "EMAG: Exponential Moving Average Guidance for Diffusion Models"
• "Harmony: Harmonizing Audio and Video Generation through Cross-Task Synergy"

Based on the node by Kijai, this expands this node by providing the ability to schedule the Enhance-A-Video tau value. EAV can disrupt and suppress high frequency (fine) details that are desired in later steps. This allows you to mitigate that occurrence while gaining substantial coherence improvements with lower frequencies earlier in the process.

Frequency-Decomposed Temporal Guidance for video diffusion models applies wavelet decomposition to the model's velocity prediction, then applies differentiated guidance to low-vs-high frequency bands with optional temporal consistency enforcement on high-frequency content.

RF-Solver: Pure deterministic ODE sampler for rectified flow models.

Use this when you need deterministic output (reproducibility, inversion). For best generation quality on LTX-2, use SA-RF-Solver with eta=1.0.

SA-RF-Solver v2: Proper SDE sampler for Rectified Flow models.

The key parameter is ETA: 0.0 = Deterministic ODE (like Euler/RF-Solver) 1.0 = Full ancestral SDE (fresh noise each step — best for LTX-2)

The predictor controls how x̂₀ is estimated: euler = 1 NFE (matches SA-Solver predictor_order=1) rf_solver_2 = 2 NFE (better x̂₀ via second-order correction) ab2 = 1 NFE after warmup (reuses velocity history)

Empirically, eta=1.0 + euler matches SA-Solver's best LTX-2 settings.

Drop-in replacement that resizes the input image to the target latent dimensions WITHOUT center-cropping. The only change from the stock node is "center" -> "disabled" in the common_upscale call, which forces a direct resize (stretch-to-fit) instead of cover-then-crop. This allows for multi-staged multi-segmented generation of i2v workflows without your images becoming misaligned from the original node.