From FlowMomentum: a momentum vector, an exponential moving average of past velocities.
From AdaptiveFlow: tolerance control - no fixed step grid; each step must earn its size.

Seren uses the momentum as a predictor. It proposes the next point for free (x_pred = x + m * dSigma); a single model evaluation there plus a trapezoid correction gives the candidate step, and the gap between prediction and correction is the error estimate that grows or shrinks the next step. A typical accepted step costs one model evaluation - the error estimate does not need a second one.

This is a deterministic rectified-flow ODE walk: no stochastic noise term, no look-ahead evaluation blended back into the step, no adaptive noise. The momentum never touches the model input; it only proposes points and measures surprise.

Node

Seren Sampler (momentum x adaptive) - sampling/seren

Inputs:

seed, cfg, flow_shift - as usual for this family of packs.
tolerance (default 0.01) - allowed per-step prediction error, relative RMS. Smaller = slower and more faithful.
beta (default 0.5) - momentum memory: 0 follows only the last velocity, higher values remember longer. Long memory predicts smoothly but can go stale where the trajectory curves hard.
max_steps (default 120) - budget cap, not a step count. If the trial budget (2 x max_steps) ever runs out, the remainder is finished with plain fixed-step Euler.

Outputs: latent, steps_used, evals_used - so you can see what a run actually cost.

Seren SDE Sampler (error-shaken) - sampling/seren

The stochastic sibling. After each accepted step, part of the state's implied noise is swapped for fresh noise along the rectified-flow interpolation line. The swap amount is gated by how hard the step just was (err / tolerance, capped by sde_strength, default 0.3): calm stretches stay nearly deterministic, uncertain stretches get shaken. The noise follows the local error topography of this walk - not a schedule tied to overall progress - and the tolerance controller still governs step sizes. Extra input: sde_strength (0-1; 0 = base Seren).

Seren Color Sampler (chroma steering) - sampling/seren

Colour handled inside the walk instead of as an after-the-fact boost: in the low-sigma stretch (sigma <= color_start, default 0.35) where colour settles, each accepted step's denoised estimate is pushed away from its per-channel spatial mean and the state is shifted to match, bending the trajectory toward richer colour while the model keeps re-denoising it. The per-step gain is scaled by the fraction of the steered band the step covers, so the total bend is set by color_strength alone and does not grow with step count; the per-step shift is also capped at 10% of the state RMS as a runaway guard. Extra inputs: color_strength (default 0.15), color_start.

KSampler registration

The pack also registers seren, seren_sde and seren_color in the standard KSampler sampler_name dropdown, with fixed defaults (tolerance 0.01, beta 0.5; sde_strength 0.3 / color_strength 0.15 for the siblings; shift estimated from the incoming sigma grid). Use any scheduler with them; the walk adapts its own steps between the grid's endpoints. The standalone nodes remain the way to tune the knobs.

Usage

Two ways to run every sampler: the standalone nodes (all knobs exposed) or the KSampler dropdown registrations (fixed defaults, any scheduler - the walk builds its own steps between the grid's endpoints). With a fixed seed everything in this pack is fully reproducible, including the SDE sibling: its noise is seed-driven, so the same seed returns the same image.

Seren (base) - start from the defaults. tolerance is the main quality/speed dial: 0.01 is a good middle, smaller is slower and more faithful, larger is faster and looser. beta 0.5 is the balanced memory; longer memory (0.7+) walks smoother but can drag stale direction through hard curves. max_steps is a budget cap, not a step count - the steps_used / evals_used outputs show what a run cost.

Seren SDE - sde_strength 0 is exactly the base walk. 0.2-0.3 gives visible variation while staying close to the deterministic trajectory; expect the endpoint to move more as strength rises, because every hard step gets shaken. If a result looks unstable, lower strength before touching tolerance.

Seren Color - color_strength sets the total bend across the steered band: each step applies the fraction of it that the step actually covers, so the effect is the same at any step count - a standalone run and a dense refine walk read the number the same way. Start at 0.1-0.15 (the registered default is 0.15, calibrated after 0.35 proved garish in practice). For a stronger effect, raise color_start (steer more of the trajectory) rather than cranking strength; the per-step shift is capped at 10% of the state RMS as a guard.

Two-stage refine recipe - a good pairing is: stage 1 any sampler at denoise 1.0, stage 2 seren or seren_color at denoise 0.3-0.4 for a deterministic finishing pass. One warning: splitting denoise across two KSamplers does not split this walk - a registered Seren run always walks to completion from the sigma it is handed (img2img semantics), so for a true mid-trajectory hand-off put a grid sampler (euler, dpmpp_2m_sde, ...) in stage 1.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/seren-alegro/ComfyUI-SerenSampler

Then restart ComfyUI.

Example

examples/seren_sampler_basic.json - a minimal workflow using the standalone Seren Sampler node (core nodes only; swap in your own checkpoint, TE and VAE loaders as needed).

Tests

tests/test_seren.py - numpy mirror of the core walk (no torch needed): tolerance sweep, evaluation economy (~1 eval per step), beta variations, and the trial-budget fallback. tests/test_seren_sde.py - shake mirror: strength 0 is exactly deterministic, strength > 0 diversifies endpoints and completes in budget (error bounds are loose by design - the test field is expansive and punishes any stochasticity far more than a trained denoiser does). tests/test_seren_color.py - steering mirror: strength 0 is exactly the base walk, strength > 0 stays finite and moves the endpoint.

License

MIT

한국어 요약

FlowMomentum의 모멘텀과 AdaptiveFlow의 허용 오차 제어를 하나로 합친 실험 샘플러예요. 모멘텀이 다음 지점을 공짜로 예측하고, 거기서 모델을 한 번만 평가해 보정하면, 예측과 보정의 차이가 곧 오차 추정이 돼요. 그래서 채택 스텝 하나에 모델 호출이 보통 한 번이면 되고, 스텝 크기는 tolerance가 알아서 조절해요. 결정적 rectified flow 전용이며, KSampler 드롭다운에도 seren이라는 이름으로 등록돼요.

자매 샘플러가 둘 있어요. seren_sde는 확률형으로, 채택된 스텝이 얼마나 힘들었는지(오차/허용치)에 비례해 그 스텝의 노이즈 일부를 새 노이즈로 갈아 끼워요 — 잔잔한 구간은 거의 결정적이고, 불안한 구간만 흔들려요. seren_color는 색을 결과물에 사후 보정하지 않고 보행 도중에 다뤄요: 색이 자리 잡는 낮은 시그마 구간에서 매 스텝의 예측 완성본을 채널 평균에서 밀어내 궤적 자체를 풍부한 색 쪽으로 굽혀요.

사용법 요점: color의 총량은 strength 하나로 정해지고 스텝 수에 따라 불어나지 않아요. 0.1~~0.15에서 시작하세요 (0.35는 실측에서 과했어요). 2단 리파인으로는 1단 아무 샘플러 denoise 1.0 + 2단 seren/seren_color 0.3~~0.4 조합이 잘 맞아요. 고정 시드에서는 확률형 포함 전부 재현돼요.