Guidance Field Rectification

• Guidance Field Rectification is the process of reshaping dynamic guidance signals, such as denoising score fields and RF magnetic fields, to maintain stability and on-manifold trajectories.
• It employs techniques like Jacobian-informed corrections, energy-aware scheduling, and predictor–corrector schemes to adjust the underlying guidance signals.
• The approach has been applied across conditional diffusion models, rectified flows, and spin–orbitronic FMR, leading to measurable improvements in image quality and signal accuracy.

Guidance field rectification denotes the calibration, reshaping, or theoretically grounded modification of a field that steers system trajectories. In recent work, the term appears most naturally in conditional generative modeling, where the relevant field is a denoising score, drift, or velocity field, and in spin–orbitronic ferromagnetic resonance, where the relevant field is the RF magnetic drive $h_{\mathrm{RF}}$. Across these settings, rectification addresses a common failure mode: the nominal guidance signal improves target adherence or excitation strength, but unrectified use can induce instability, off-manifold drift, train–inference mismatch, or spurious observables that mimic the intended effect (Sanjyal, 13 Jul 2025, Gao et al., 31 Jan 2025, Saini et al., 9 Oct 2025, Fu et al., 29 Jun 2026, Yactayo et al., 16 Feb 2026).

1. Cross-domain meaning and formal scope

In diffusion and flow models, a guidance field is the effective vector field used during sampling. For classifier-free guidance (CFG), this field is obtained by combining conditional and unconditional predictions; for rectified flows, it is the learned velocity field $v_\theta(x_t,t,y)$ or a guided modification of it; for reinforcement-learning-based alignment of flows, the policy itself is represented by the velocity field (Sanjyal, 13 Jul 2025, Saini et al., 9 Oct 2025, Fu et al., 29 Jun 2026). In spin–orbitronics, the RF field $h_{\mathrm{RF}}$ plays an analogous role because it guides magnetization precession and thereby determines both desired spin-pumping or spin-torque signals and unwanted rectification voltages (Yactayo et al., 16 Feb 2026).

Domain	Guidance field	Rectification target
Conditional diffusion	Guided noise/score or denoising drift	Stable energy trajectories, better approximation to optimal guidance (Sanjyal, 13 Jul 2025, Gao et al., 31 Jan 2025)
Rectified flow / flow matching	Velocity field $v_\theta$	Manifold adherence, bounded deviation, RL-optimal alignment (Saini et al., 9 Oct 2025, Fu et al., 29 Jun 2026)
Spin–orbitronic FMR	RF magnetic field $h_{\mathrm{RF}}$	Correct metrology and separation of genuine conversion from AMR/AHE rectification (Yactayo et al., 16 Feb 2026)

A plausible unifying interpretation is that guidance field rectification is a control problem imposed on an already learned or already driven dynamical system. The rectified field is not merely stronger or weaker; it is modified so that its induced trajectory remains compatible with the geometry, energetics, or measurement model of the system.

2. Diffusion-theoretic rectification of guidance gradients

A central theoretical development is the replacement of the usual “scaled marginal” account of guidance by a scaled joint distribution over the entire diffusion chain. In "REG: Rectified Gradient Guidance for Conditional Diffusion Models" (Gao et al., 31 Jan 2025), the practical guided field is written through the modified noise prediction $\bar{\boldsymbol{\epsilon}}_{\theta,t}$, yielding the effective score field

$$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$

The paper argues that the conventional story—rescaling only the endpoint or each marginal independently—is mathematically inconsistent for a DDPM reverse chain. The corrected objective is the scaled joint distribution

$$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$

with expected future reward

$$E_t(\mathbf{x}_t,\mathbf{y}) := \int p_\theta(\mathbf{x}_0\mid \mathbf{x}_t,\mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y})\,d\mathbf{x}_0.$$

Under this formulation, the optimal guided noise network is

$$\bar{\boldsymbol{\epsilon}}^\star_{\theta,t} = \boldsymbol{\epsilon}_{\theta,t} - \sqrt{1-\bar\alpha_t}\,\nabla_{\mathbf{x}_t}\log E_t(\mathbf{x}_t,\mathbf{y}),$$

so the correct rectification target is the gradient of the log expected future reward rather than a local reward surrogate (Gao et al., 31 Jan 2025).

The same work interprets existing guidance methods as no-future-foresight approximations. Standard CG, CFG, and AutoG replace $v_\theta(x_t,t,y)$0 with tractable local terms such as $v_\theta(x_t,t,y)$1, thereby producing a biased time-local guidance field. Rectified Gradient Guidance (REG) modifies this field with a Jacobian-informed multiplicative correction: $v_\theta(x_t,t,y)$2 This is presented as a tractable approximation to the optimal joint-based guidance field (Gao et al., 31 Jan 2025).

Empirically, the paper reports that REG yields a better approximation to the optimal solution in 1D and 2D settings and improves large-scale image generation metrics. On DiT-XL/2 at $v_\theta(x_t,t,y)$3, vanilla CFG improves from FID $v_\theta(x_t,t,y)$4 to $v_\theta(x_t,t,y)$5 and Inception Score from $v_\theta(x_t,t,y)$6 to $v_\theta(x_t,t,y)$7 when REG is added. On Stable Diffusion v1.4, vanilla CFG improves from FID $v_\theta(x_t,t,y)$8 to $v_\theta(x_t,t,y)$9 and CLIP $h_{\mathrm{RF}}$0 to $h_{\mathrm{RF}}$1; on SD-XL with cosine CFG, FID improves from $h_{\mathrm{RF}}$2 to $h_{\mathrm{RF}}$3 and CLIP from $h_{\mathrm{RF}}$4 to $h_{\mathrm{RF}}$5 (Gao et al., 31 Jan 2025). A common misconception addressed here is that guidance is adequately described as a marginal reweighting of the endpoint distribution; the paper explicitly rejects that interpretation.

3. Energy-aware rectification in high-resolution latent diffusion

In "RectifiedHR: High-Resolution Diffusion via Energy Profiling and Adaptive Guidance Scheduling" (Sanjyal, 13 Jul 2025), guidance field rectification is formulated operationally through the latent energy trajectory. The setting is Stable Diffusion 1.5 at $h_{\mathrm{RF}}$6 resolution. The paper defines the latent energy

$h_{\mathrm{RF}}$7

and interprets the collection $h_{\mathrm{RF}}$8 as an energy trajectory during denoising. CFG defines the guided noise prediction

$h_{\mathrm{RF}}$9

which in turn induces the sampling drift. The paper’s central claim is that fixed CFG scales distort this field at high resolution, especially at large $v_\theta$0, and that the distortion is directly visible as spikes, oscillations, or premature flattening in $v_\theta$1 (Sanjyal, 13 Jul 2025).

The experiments profile energy over CFG scales $v_\theta$2, samplers DDIM, Euler A, and DPM++ 2M, and $v_\theta$3 denoising steps. Under high fixed CFG, the paper reports energy spikes near the end of sampling, oscillations in the last third of steps for some samplers and guidance scales, and higher variance across runs and prompts. These deviations correlate with oversaturation, loss of fine details, unrealistic or “crunchy” textures, and distortion or halos around edges (Sanjyal, 13 Jul 2025).

Rectification is introduced through time-varying schedules $v_\theta$4, including linear-decreasing, cosine ramp, step function, exponential, and sigmoid schedules. The linear-decreasing schedule is emphasized: $v_\theta$5 The stated rationale is that strong guidance is useful early, when latents are mostly noise and global structure is being set, whereas large guided drifts are dangerous late, when the sample is close to the data manifold and only small refinements are needed (Sanjyal, 13 Jul 2025).

The best empirical combination is DPM++ 2M with linear-decreasing CFG scheduling, which reaches a stability score up to $v_\theta$6 and a consistency metric up to $v_\theta$7. The paper further defines energy-based metrics for stability, consistency, efficiency, and convergence, and notes that energy clipping and adaptive energy thresholds can further reduce occasional spikes. The method is described as plug-and-play, with negligible overhead beyond a norm computation per step and no retraining or sampler rewrite (Sanjyal, 13 Jul 2025). This work also addresses a frequent practical misconception: a single fixed CFG scale is not a neutral default in high-resolution sampling; it is a globally mis-scaled guidance field whose late-step behavior can dominate artifact formation.

4. Velocity-field rectification in rectified flows and RL alignment

Rectified flow models replace stochastic denoising with deterministic ODE transport. In this setting, guidance field rectification has acquired two distinct forms: geometry-aware inference-time correction and reinforcement-learning-derived velocity rectification.

In "Rectified-CFG++ for Flow Based Models" (Saini et al., 9 Oct 2025), the unrectified field is standard flow-based CFG,

$v_\theta$8

which the paper characterizes as an extrapolation that provokes severe off-manifold drift in rectified flows. The proposed rectification is a predictor–corrector rule. First, a half-step predictor advances along the conditional field,

$v_\theta$9

Then a corrector evaluates conditional and unconditional velocities at the predicted point and constructs

$h_{\mathrm{RF}}$0

Under Lipschitz continuity, bounded guidance difference, bounded integrable schedule, and bounded conditional velocity, the paper proves bounded single-step deviation, bounded tubular-neighborhood proximity to the manifold family $h_{\mathrm{RF}}$1, and bounded KL divergence from the target conditional marginals (Saini et al., 9 Oct 2025).

The empirical results are reported on Flux, Stable Diffusion 3 and 3.5, and Lumina. On MS-COCO 10K for Flux-dev, FID improves from $h_{\mathrm{RF}}$2 to $h_{\mathrm{RF}}$3, CLIP from $h_{\mathrm{RF}}$4 to $h_{\mathrm{RF}}$5, Aesthetic from $h_{\mathrm{RF}}$6 to $h_{\mathrm{RF}}$7, and PickScore from $h_{\mathrm{RF}}$8 to $h_{\mathrm{RF}}$9. For Flux on MS-COCO 1K, the $\bar{\boldsymbol{\epsilon}}_{\theta,t}$0-step comparison is particularly large: CFG FID $\bar{\boldsymbol{\epsilon}}_{\theta,t}$1 versus Rectified-CFG++ FID $\bar{\boldsymbol{\epsilon}}_{\theta,t}$2 (Saini et al., 9 Oct 2025). The paper also emphasizes that the method is training-free and flow-specific, rather than a direct transfer of diffusion-oriented guidance heuristics.

A different rectification principle appears in "FlowAWR: Online Adaptive Flow Reinforcement via Advantage-Weighted Rectification" (Fu et al., 29 Jun 2026). Here the field to be rectified is the generative policy’s velocity field itself. Starting from a KL-constrained reward maximization objective, the paper derives the optimal terminal policy

$\bar{\boldsymbol{\epsilon}}_{\theta,t}$3

propagates it to intermediate times, and obtains an optimal velocity field of the form

$\bar{\boldsymbol{\epsilon}}_{\theta,t}$4

where

$\bar{\boldsymbol{\epsilon}}_{\theta,t}$5

This is called advantage-weighted rectification: better-than-average samples pull the field toward their conditional target direction, worse-than-average samples push it away, and the magnitude of rectification is advantage-dependent rather than fixed (Fu et al., 29 Jun 2026).

FlowAWR trains by supervised regression to a stochastic version of this optimal field and proves optimization equivalence between the AWR loss and the ideal loss. It is explicitly described as SDE-free optimization and CFG-free generation. On SD3.5-Medium, it reaches PickScore $\bar{\boldsymbol{\epsilon}}_{\theta,t}$6 in $\bar{\boldsymbol{\epsilon}}_{\theta,t}$7k steps, compared with $\bar{\boldsymbol{\epsilon}}_{\theta,t}$8 in $\bar{\boldsymbol{\epsilon}}_{\theta,t}$9k steps for DiffusionNFT and $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$0 in $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$1k steps for FlowGRPO, corresponding to a reported $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$2 to $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$3 convergence acceleration (Fu et al., 29 Jun 2026). The paper also characterizes DiffusionNFT as a binary quantized special case of the same rectification template.

Taken together, these two works show that in rectified flows the central issue is not only the strength of conditioning but the geometry of the induced velocity field. One rectification route operates at inference through predictor–corrector sampling; the other internalizes rectification into the trained field through advantage-weighted regression.

5. RF-field rectification in spin pumping and spin-torque FMR

Outside generative modeling, "RF field characterization and rectification effects in spin pumping and spin-torque FMR for spin-orbitronics" (Yactayo et al., 16 Feb 2026) uses a closely related notion in experimental metrology. The guiding field is the RF magnetic field $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$4, whose amplitude and orientation determine both the magnitude of spin-pumping signals and the presence of spurious rectification voltages. The paper states that quantifying $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$5 correctly is essential because the pumped spin current scales as $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$6.

The work derives a calibration protocol based on AMR photoresistance under an applied DC current and obtains a compact formula for $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$7 that is valid for any antenna geometry. It also analyzes the voltage line shape by decomposing symmetric and antisymmetric Lorentzian components. A key result is that odd-symmetric Lorentzian voltages, which perfectly mimic spin-pumping or spin-torque FMR signals, can arise purely from rectification effects due to anisotropic magnetoresistance when the orientation of $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$8 is parallel to the ferromagnetic surface (Yactayo et al., 16 Feb 2026).

The geometry dependence is central. In the GAP geometry, the RF field is mainly out-of-plane and the signal is dominated by a symmetric even Lorentzian, consistent with AHE mixing. In the SHORT geometry, the field is mainly in-plane along the strip and the signal is dominated by a symmetric Lorentzian odd in field from transverse AMR or planar Hall effect; this has exactly the symmetry of an ISHE spin-pumping signal. In the FULL geometry, both $$\mathbf{g}_t(\mathbf{x}_t,\mathbf{y}) := -\frac{\bar{\boldsymbol{\epsilon}}_{\theta,t}}{\sqrt{1-\bar\alpha_t}}.$$9 and $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$0 are present and mixed symmetries appear (Yactayo et al., 16 Feb 2026). In ST-FMR, the paper further shows that even a single ferromagnetic layer can exhibit a sizable symmetric odd voltage, with magnitudes up to approximately $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$1 for NiFe, purely from self-induced AMR rectification.

The material dependence is equally sharp. For $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$2-nm ferromagnetic layers, Fe and CoFeB exhibit minimal rectification, whereas Ni and NiFe generate strong rectified signals that must be corrected. The thickness dependence is summarized by the result that these rectification effects become negligible for ferromagnetic thicknesses $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$3 nm, validated in NiFe/Pt bilayers (Yactayo et al., 16 Feb 2026). The paper also uses geometry-independent $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$4 in NiFe(6 nm)/Pt(5 nm) as a validation criterion, whereas NiFe(10 nm)/Pt(5 nm) retains geometry dependence, indicating rectification contamination.

This literature addresses a misconception analogous to those seen in generative models: a measured signal with the expected symmetry need not originate from the intended mechanism. In FMR metrology, line-shape symmetry alone is insufficient unless the RF guidance field has been calibrated and its rectification channels isolated.

6. Comparative themes, misconceptions, and open problems

Several recurrent themes emerge. First, guidance rectification is not equivalent to increasing or decreasing a scalar knob. In diffusion, REG replaces a local reward gradient by a Jacobian-corrected field (Gao et al., 31 Jan 2025), and RectifiedHR replaces a fixed CFG scale by energy-aware schedules and optional clipping (Sanjyal, 13 Jul 2025). In rectified flows, Rectified-CFG++ replaces unconditional extrapolation by conditional anchoring plus a scheduled corrector (Saini et al., 9 Oct 2025), while FlowAWR derives a magnitude-aware correction from a KL-constrained RL objective rather than using a fixed heuristic magnitude (Fu et al., 29 Jun 2026).

Second, several papers reject endpoint-only or inference-only intuitions. REG argues that a scaled endpoint marginal without modifying the rest of the reverse chain is theoretically invalid (Gao et al., 31 Jan 2025). FlowAWR eliminates the training–inference mismatch associated with SDE rollouts at training and deterministic ODE plus CFG at inference by learning a rectified deterministic velocity field directly (Fu et al., 29 Jun 2026). Rectified-CFG++ similarly treats instability as a trajectory-level geometric problem rather than a per-step semantics problem (Saini et al., 9 Oct 2025).

Third, stronger guidance is repeatedly shown to be non-monotonic in quality. High fixed CFG produces energy spikes and visible artifacts in high-resolution latent diffusion (Sanjyal, 13 Jul 2025). Large $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$5 in flow-based CFG causes off-manifold drift, text misalignment, and brittle behavior (Saini et al., 9 Oct 2025). In spin–orbitronic FMR, the orientation of $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$6 can amplify rectified voltages that mimic genuine spin–charge conversion rather than improving measurement fidelity (Yactayo et al., 16 Feb 2026).

The reported limitations are domain-specific but structurally similar. RectifiedHR is restricted to SD 1.5 at $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$7, uses a simple latent norm as energy, and offers no theoretical guarantees (Sanjyal, 13 Jul 2025). REG relies on a diagonal-dominance-style Jacobian approximation, incurs extra backward computation during sampling, and does not implement the theoretically modified initial noise distribution (Gao et al., 31 Jan 2025). Rectified-CFG++ cannot repair deficiencies in the underlying RF model and can eventually over-constrain at extremely large guidance scales (Saini et al., 9 Oct 2025). FlowAWR depends on finite group-size estimates of advantage, adaptive $$\bar{p}_\theta(\mathbf{x}_{0:T}\mid \mathbf{y}) \propto p_\theta(\mathbf{x}_{0:T}\mid \mathbf{y})\,R_0(\mathbf{x}_0,\mathbf{y}),$$8, EMA updates of the reference policy, and a ReFlow-type derivation (Fu et al., 29 Jun 2026). The FMR literature requires careful geometry control, FM-only references, thickness series, and use of the total AMR of the full stack in RF-field calibration (Yactayo et al., 16 Feb 2026).

These works collectively suggest that guidance field rectification is best understood as a design principle rather than a single algorithm. The principle is to replace a nominally useful but dynamically or experimentally inconsistent guidance signal with a field whose magnitude, geometry, temporal schedule, or reward dependence is explicitly matched to the system it steers.