FunPack Distilled Flow Sampler

Distilled video models (think LTX2.3 with a distilled LoRA) are trained to hit their target in very few steps, and they're picky about how you sample them. Run them with a generic sampler and you get exactly the kind of artifacts that make few-step models look cheap. FunPack Distilled Flow Sampler is the ODE sampler built for that job: Adams-Bashforth 2-step for better trajectory accuracy across the big sigma jumps a distilled schedule throws at you, a Heun predictor-corrector on the final steps for detail, and a pile of experimental extras bolted on.

The inputs a beginner actually sets

Out of a very long list, three controls do the work:

• order (default 2) - 1 is plain Euler; 2 is Adams-Bashforth 2-step, which extrapolates the denoised direction from two consecutive steps for better accuracy at no extra model-call cost. Leave it at 2.
• final_correction_steps (default 1) - the last few steps get a Heun predictor-corrector pass. Each costs one extra model call but noticeably improves final detail. 1 is usually enough on an 8-step run.
• s_noise (default 0) - optional stochastic noise for diversity. The tooltip's advice is worth quoting: 0 is fully deterministic and recommended; small values (0.05–0.15) add variation without wrecking the distilled trajectory. If you want reproducibility, leave it at 0.

Outputs are sampler and sigmas, wired into a SamplerCustomAdvanced-style node: feed your base scheduler's sigmas into this node's optional sigmas input, then take both outputs into the sampler.

The rest, honestly labeled

Everything else is either helpful or experimental, and the tooltips say which is which. ab2_ramp is a genuinely useful freebie: instead of full 2nd-order on every step it ramps the AB2 contribution 0→1 across the schedule, keeping the noisy early steps close to 1st-order to avoid overshoot - no extra model calls. normalize_strength is anti-overbake (latent normalization against oversaturation and color drift), video-only, near-zero overhead.

Then the experimental tier: velocity_bias_* captures/applys an averaged early model velocity as a creative action-injection tool (strength 2–3 approaches full action replacement - the tooltip literally says "creative tool, not for consistency"), rescue_mode steers in-flight toward trajectories you rated good and away from ones you rated Awful (a no-op until you've rated a few gens for the key), alg_enabled is Adaptive Low-Pass Guidance (blur the i2v anchor during early steps so the model doesn't shortcut to a near-static video), and mg_enabled is Momentum Guidance for the fine-motion window. These are the "self-improving engine" features - skip them on your first runs and come back once the baseline looks right.

Install

It's part of ComfyUI-FunPack:

cd ComfyUI/custom_nodes
git clone https://github.com/olivv-cs/ComfyUI-FunPack
pip install -r requirements.txt

or ComfyUI Manager → search "ComfyUI-FunPack". No extra dependencies.

The honest take

For distilled few-step models this is genuinely the right tool - a distilled schedule's giant jumps are exactly where a 2nd-order method earns its keep. The experimental knobs are where the pack's personality lives, and they're marked experimental for a reason: mg_decay above ~0.5 is documented to produce garbage on the default 8-step schedule, so read the tooltips before turning anything on. Start with order=2, one final correction step, s_noise=0, and nothing else.