FunPack Distilled Flow Sampler

ODE sampler for distilled few-step video models (e.g. LTX2.3 distilled LoRA). Adams-Bashforth 2-step multistep for better trajectory accuracy across large sigma jumps, Heun predictor-corrector on final steps for quality, optional controlled noise for diversity, optional velocity bias + reactive rescue (shared with the Hybrid/RF samplers), experimental ALG anchor de-staticking for i2v chunks, and experimental Momentum Guidance smoothing for the complementary (fine-motion) sigma window.

By digital-garbage·Created about a year ago·Updated 24 days ago· 16

sigmas

sampler
sigmas

◄order2►

◄final_correction_steps1►

◄s_noise0.00►

◄velocity_bias_modeoff►

◄velocity_bias_strength0.00►

◄velocity_bias_sourcemean►

◄velocity_refinement_keydefault►

◄rescue_modefalse►

◄rescue_threshold0.15►

◄rescue_strength0.20►

◄ab2_rampfalse►

◄normalize_strength0.00►

◄normalize_start_sigma0.900►

◄alg_enabledfalse►

◄alg_strength2.0►

◄alg_sigma_threshold0.975►

◄mg_enabledfalse►

◄mg_strength0.50►

◄mg_decay0.50►

◄mg_sigma_threshold0.975►

◄quality_sharpness0.00►

CategoryFunPack/Sampling

Inputs (22)

Name	Type	Default	Description
order	INT	21–2	Multistep order. 1 = standard Euler ODE. 2 = Adams-Bashforth 2-step: extrapolates the denoised direction from two consecutive steps for better accuracy at no extra model-call cost.
final_correction_steps	INT	10–3	Number of final steps that use a Heun predictor-corrector pass. Each costs one extra model call but significantly improves final-step detail. 1 is usually enough for 8-step runs.
s_noise	FLOAT	0.000–0.5	Optional stochastic noise for diversity. 0 = fully deterministic ODE (recommended). Small values (0.05–0.15) add variation without strongly disrupting the distilled trajectory.
velocity_bias_mode	COMBO	off	Experimental: capture/apply averaged early model velocity around normalized sigma 0.9/0.72/0.42. Off preserves the plain distilled ODE. Note: few-step schedules may only land on a target or two, so it fires less often than on an 8-step run.
velocity_bias_strength	FLOAT	0.000–3	Strength of the remembered velocity (action) injected at the structure sigma. 0 disables. ~0.15 = subtle spice; 0.3-1.0 = clear action crossover; 2-3 approaches full action replacement (capped so the current gen isn't wiped). Creative tool, not for consistency.
velocity_bias_source	COMBO	mean	How velocity bias / rescue pick a good direction. 'mean' = prompt-blind global average. 'nearest' = single best-matching prompt cluster — preserves one real good gen's detail instead of a washed-out average. Affects both apply and rescue.
velocity_refinement_key	STRING	default	Memory key used to capture/apply early velocity bias and rescue trajectories.
rescue_mode	BOOLEAN	false	Reactive in-flight rescue, rating-gated. Steers each eligible step toward trajectories you rated good and away from ones you rated Awful (matched to the current prompt). A no-op until you've rated a few gens for this prompt/key.
rescue_threshold	FLOAT	0.150–1	Fires when the step has diverged from the good trajectory by more than this (1 - cosine) OR aligned with a bad trajectory by more than this (cosine). Lower = corrects more eagerly. 0.10-0.20 typical.
rescue_strength	FLOAT	0.200–0.5	How hard to pull toward good / push away from bad when triggered (magnitude preserved, no energy injected). Keep moderate; 0.5 is a strong correction.
sigmasopt	SIGMAS		—
ab2_rampopt	BOOLEAN	false	Graduated 2nd order (free). Instead of full AB2 on every step, ramp the AB2 contribution linearly 0->1 across the schedule: early/noisy steps stay near 1st-order euler (less overshoot), late/detail steps get full AB2. No extra model calls. Helps low-step distilled runs. No effect at order=1.
normalize_strengthopt	FLOAT	0.000–1	Video-only latent normalization (anti-overbake / oversaturation / colour drift) stacked on this ODE. 0 = off. 0.5 = gentle. Audio is never touched. ~zero overhead.
normalize_start_sigmaopt	FLOAT	0.9000–1	Sigma at/below which latent normalization activates and anchors its reference (above it the x0 estimate is meaningless). Only used when normalize_strength > 0.
alg_enabledopt	BOOLEAN	false	EXPERIMENTAL: Adaptive Low-Pass Guidance (arXiv:2506.08456). Blurs the i2v anchor frame while sigma is above alg_sigma_threshold, then swaps back to the sharp anchor — counters the model's tendency to shortcut to a near-static video that just matches the reference image. No-op without an i2v anchor.
alg_strengthopt	FLOAT	2.01–4	Downsample factor for the anchor blur (paper default 2.5, but 2.0 held character/i2v consistency noticeably better in testing here). Higher = blurrier anchor during the affected steps. Only used when alg_enabled; guide-frame blur has its own controls on the Scene Chain Sampler (alg_blur_guides + alg_guide_blur_*).
alg_sigma_thresholdopt	FLOAT	0.9750.5–0.999	Anchor stays blurred while sigma is above this value (the near-pure-noise steps), then swaps to sharp. Higher = narrower blurred window. Only used when alg_enabled; guide-frame blur has its own controls on the Scene Chain Sampler (alg_blur_guides + alg_guide_blur_*).
mg_enabledopt	BOOLEAN	false	EXPERIMENTAL: Momentum Guidance (arXiv:2602.20360). Keeps a running average of the per-step direction and blends the current step toward it once sigma drops below mg_sigma_threshold — the complementary window to ALG's blur. Smooths the fine-motion/refinement steps; may damp motion as a side effect, untested for video.
mg_strengthopt	FLOAT	0.500–1	Blend weight toward the momentum average when active (0 = no effect, 1 = fully replace the step's direction with the average). Only used when mg_enabled.
mg_decayopt	FLOAT	0.500–0.99	EMA decay for the momentum average (higher = longer memory of past steps' directions). On our 8-step schedule, high decay (e.g. 0.9) keeps the EMA anchored to the very first, near-pure-noise step's direction for nearly the whole run — wrong scale, produces garbage regardless of mg_strength. Tested safe (and good) at 0.5 even at mg_strength=1.0. Only used when mg_enabled.
mg_sigma_thresholdopt	FLOAT	0.9750.5–0.999	Momentum guidance applies while sigma is BELOW this value (the opposite window from alg_sigma_threshold) — defaults to the same boundary as ALG for a clean handoff. Only used when mg_enabled.
quality_sharpnessopt	FLOAT	0.000–1	Restores fine detail via temporal-average unsharp on the x0 prediction, applied only during the final Heun-correction steps (final_correction_steps > 0 required to have any effect). 0 disables. 0.2-0.4 typical. Free (no extra model eval), video-only.

Outputs (2)

Name	Type	Description
sampler	SAMPLER	—
sigmas	SIGMAS	—