FlexQ: Quadratic Latent Injection

• FlexQ is a latent-space regularization framework that employs a quadratic penalty to couple physical data fidelity with latent priors.
• It decouples PDE-constrained inversion from latent projection, enabling alternating physical and latent updates for stable optimization.
• Empirical evaluations show methods akin to FlexQ (such as DLO) outperform traditional regularizers in seismic inversion metrics like MAE, RMSE, and SSIM.

Searching arXiv for “FlexQ” to identify the topic and relevant papers. FlexQ is not defined as a named method in the provided arXiv corpus. The attested material instead centers on a family of latent-space optimization and injection procedures for inverse problems, conditional reconstruction, editing, privacy, and security. The closest explicitly specified quadratic latent-coupling formulation is Decoupled Latent Optimization (DLO) for full waveform inversion, which introduces a physical variable $x$, a latent variable $z$, a forward operator $F(x)$, and a pretrained diffusion decoder $G(z)$, and optimizes a joint objective that couples data fidelity with a quadratic penalty toward the learned prior manifold (Min et al., 12 Jun 2026). In that sense, any use of “FlexQ” in this context is best interpreted cautiously, with DLO providing the nearest precise formulation available in the source material.

1. Terminological status and nearest documented formulation

Within the cited sources, the most explicit latent-variable construction is the DLO objective

$$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$

where $x\in\mathbb{R}^n$ is the discretized velocity model, $z\in\mathbb{R}^d$ is a latent vector drawn from a standard Gaussian prior, $F(x)$ is the PDE forward operator, and $G(z)\equiv R_\theta(z)$ is the pretrained diffusion-model decoder (Min et al., 12 Jun 2026). The formulation relaxes the equality constraint $x=G(z)$ through a quadratic penalty rather than enforcing exact equality.

This construction is technically distinctive because no additional diffusion log-likelihood term appears explicitly. All prior information enters through the decoded sample $z$0, while the PDE-constrained misfit remains the standard

$z$1

The resulting framework therefore separates physical-space optimization from latent-space prior projection in a manner that is explicit, implementable, and benchmarked in the source literature (Min et al., 12 Jun 2026).

2. Quadratic coupling, alternating minimization, and implementation details

DLO alternates between an $z$2-step in physical space and a $z$3-step in latent space. With $z$4 fixed, the physical update minimizes

$z$5

whose gradient is

$z$6

The update is a descent step

$z$7

With $z$8 fixed, the latent update minimizes

$z$9

with gradient

$F(x)$0

computed by back-propagating through an unrolled DDIM chain, typically with $F(x)$1 steps for efficiency (Min et al., 12 Jun 2026).

The initialization and continuation strategy are equally specific. The physical initialization $F(x)$2 is a Gaussian-smoothed background or true velocity model, preserving the standard smooth-start basin of attraction of classical FWI. The latent initialization $F(x)$3 is sampled once from $F(x)$4. The penalty weight is held fixed, with $F(x)$5 given as an example of a moderately large value, and the pretrained diffusion prior uses $F(x)$6 VP steps while DLO itself employs a short $F(x)$7 DDIM sampler for $F(x)$8 (Min et al., 12 Jun 2026).

A central property of this design is the decoupling of forces: the data-fidelity gradient acts only on $F(x)$9, while the prior acts only through the decoded sample $G(z)$0. The source describes this as numerically stable and emphasizes that the method avoids backpropagation through the PDE solver during the latent update (Min et al., 12 Jun 2026).

3. Empirical behavior in seismic inversion

On OpenFWI, DLO is reported to uniformly outperform Tikhonov, TV, DiffusionFWI, and RED-DiffEq in MAE, RMSE, and SSIM under clean, noisy, and missing-trace settings, including noise levels up to $G(z)$1 and missing-trace rates up to $G(z)$2 (Min et al., 12 Jun 2026). The method is also evaluated out of distribution on Marmousi and Overthrust. Using a CF-B prior trained on $G(z)$3 OpenFWI models, it transfers directly to $G(z)$4 domains and recovers dipping layers and fault detail while showing minimal sensitivity to initialization smoothing and measurement noise (Min et al., 12 Jun 2026).

These results place the quadratic latent-coupling formulation in a broader inverse-problem context. Classical regularizers stabilize FWI but fail to reproduce realistic geological structures, whereas earlier diffusion-prior methods improve realism at the cost of a fragile trade-off between data fidelity and prior consistency. DLO’s reported advantage is that it preserves standard FWI initialization while biasing the iterate toward geologically plausible models through the penalty term $G(z)$5 (Min et al., 12 Jun 2026).

A plausible implication is that the source material treats latent injection less as a replacement for PDE-constrained inversion than as a structural prior that is grafted onto the existing optimization loop. In the provided corpus, this is the clearest example of a quadratic latent-coupling method that could plausibly motivate an ambiguous label such as “FlexQ,” although that name itself is not attested.

4. Diffusion inversion and latent-bias correction in image reconstruction

Several related works in the corpus replace quadratic coupling with per-step latent correction during diffusion inversion. “Latent Bias Alignment for High-Fidelity Diffusion Inversion in Real-World Image Reconstruction and Manipulation” introduces a learnable latent bias $G(z)$6 at each reverse diffusion step,

$G(z)$7

and learns these biases to reduce misalignment between inversion and generation trajectories (Chen et al., 25 Mar 2026). The same work adds Image Latent Boosting, which refines the VQAE latent $G(z)$8 by minimizing a combination of reconstruction consistency and a diffusion-compatibility regularizer before the inversion stage (Chen et al., 25 Mar 2026).

A different training-free line appears in “Latent Inversion with Timestep-aware Sampling for Training-free Non-rigid Editing,” which combines text optimization, null-text inversion, and timestep-aware text injection. Early sampling steps use the source prompt to preserve coarse geometry, and later steps use an interpolated target embedding to increase editability. The paper reports that a cutoff ratio $G(z)$9 gives the best compromise, and on TEdBench at mid-range LPIPS $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$0 it reports CLIP-Text $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$1 and Aesthetic $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$2, compared with $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$3 for Imagic and $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$4 for MasaCtrl (Jung et al., 2024).

“Structure-Preserving Zero-Shot Image Editing via Stage-Wise Latent Injection in Diffusion Models” makes the stage structure explicit. It uses timestep-specific null-text embeddings $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$5, source latents $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$6, and reference latents $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$7, with shape injection in early steps and attribute injection in later steps. The mixing coefficient may be binary or a linear ramp,

$$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$8

while self-attention on source latents preserves structure and cross-attention with reference latents transfers attributes (Jeong et al., 22 Apr 2025).

Across these papers, latent injection is not a single algorithm but a family of mechanisms: additive bias vectors, per-timestep null embeddings, stage-wise source/reference mixing, and interface correction between VQAE and diffusion latents. The common motif is controlled intervention in latent trajectories rather than direct retraining of the generative backbone.

5. Flow-based and rectified-flow steering formulations

In flow-based editing, the corpus documents a shift from repeated source-latent reinjection toward weaker, more targeted interventions. “ProEdit: Inversion-based Editing From Prompts Done Right” argues that direct latent blending overly re-imposes the source distribution and inhibits changes in color, pose, and number. Its Latents-Shift perturbs only the edited region of the initial inverted latent using an AdaIN-style transformation and a mask $$L(x,z) = \|F(x)-d_{\mathrm{obs}}\|^2 + \lambda\|x-G(z)\|^2,$$9, with $x\in\mathbb{R}^n$0, and applies the perturbation only once at $x\in\mathbb{R}^n$1. KV-mix then combines source and target key/value features with $x\in\mathbb{R}^n$2 to preserve background consistency while allowing the target prompt to dominate inside the edit region (Ouyang et al., 26 Dec 2025).

The same section of the corpus also includes a control-theoretic construction. “Prompt-Guided Dual Latent Steering for Inversion Problems” runs two rectified-flow inversion paths in parallel: a structural path under a null prompt and a semantic path under a text prompt. These are fused by an optimal controller derived from a Linear Quadratic Regulator, with time-varying gain

$x\in\mathbb{R}^n$3

and time-decaying steering

$x\in\mathbb{R}^n$4

The implementation uses Flux.1-dev with 28 Euler steps, RF-Inversion rectification $x\in\mathbb{R}^n$5, and reports on FFHQ-1K and ImageNet-1K improvements across super-resolution, deblurring, and inpainting tasks (Wu et al., 23 Sep 2025).

These works clarify an important distinction. In DLO-like formulations, the prior enters through a penalty toward a decoded latent sample. In flow-based editing and rectified-flow steering, the intervention is dynamic and trajectory-level: perturbation, attention mixing, or closed-form control is applied at specific timesteps to balance structure preservation against semantic change.

6. Security, privacy, and inversion as a threat model

The same latent-injection logic also appears in security and privacy settings. “Public Diffusion Models, Private Images: Key-Controlled Inversion for Conditional Reconstruction” introduces a key-dependent Gaussian perturbation $x\in\mathbb{R}^n$6 into each O-BELM inversion step. Because the same $x\in\mathbb{R}^n$7 is used during decryption and the O-BELM coefficients satisfy mutual-inverse constraints, correct-key reconstruction remains possible while wrong-key outputs collapse. The paper states IND-CPA security with exponentially small adversarial advantage in a tunable security parameter and reports, on ImageNet, correct-key PSNR $x\in\mathbb{R}^n$8, SSIM $x\in\mathbb{R}^n$9, and FID $z\in\mathbb{R}^d$0, versus wrong-key or no-key PSNR $z\in\mathbb{R}^d$1, SSIM $z\in\mathbb{R}^d$2, and FID $z\in\mathbb{R}^d$3 (Zhang et al., 22 Jun 2026).

“PRoADS: Provably Secure and Robust Audio Diffusion Steganography with latent optimization and backward Euler Inversion” uses latent optimization to reduce encoder–decoder mismatch,

$z\in\mathbb{R}^d$4

followed by backward-Euler inversion to minimize diffusion inversion error. The paper reports BER $z\in\mathbb{R}^d$5 under $z\in\mathbb{R}^d$6 kbps MP3 compression and gives detailed ablations showing that latent optimization and backward Euler each reduce BER, with the combined method performing best (Yan et al., 11 Mar 2026).

A privacy-preserving synthetic-data variant appears in “Latent Noise Injection for Private and Statistically Aligned Synthetic Data Generation,” which uses Masked Autoregressive Flows to map each real sample $z\in\mathbb{R}^d$7 to $z\in\mathbb{R}^d$8, perturbs it by Gaussian latent noise, and maps back via $z\in\mathbb{R}^d$9. The procedure is described as satisfying local $F(x)$0-differential privacy and preserving a one-to-one correspondence between real and synthetic points. The paper emphasizes high-dimensional robustness and reports that meta-analysis across $F(x)$1 independent studies restores classical efficiency (Shen et al., 19 Jun 2025).

The threat model is not confined to vision or audio. “An Invariant Latent Space Perspective on LLM Inversion” treats the LLM as an invariant decoder and injects a learned pseudo-representation $F(x)$2 directly into the decoder, bypassing its embedding layer. Across 9 datasets it reports an average $F(x)$3 BLEU improvement over baselines, and its analysis argues that prevalent defenses provide limited protection (Ye et al., 24 Nov 2025).

7. Conceptual synthesis

Taken together, the corpus supports a coherent encyclopedic characterization of latent injection methods even though it does not define “FlexQ” itself. The recurring pipeline is: encode or invert an observation into a latent representation; modify that latent object by a penalty term, an additive bias, a steering controller, a perturbation, or a region-specific shift; then decode or sample through a fixed pretrained model. The intervention may be static, as in quadratic coupling $F(x)$4 (Min et al., 12 Jun 2026), or dynamic, as in timestep-specific bias injection (Chen et al., 25 Mar 2026), stage-wise source/reference mixing (Jeong et al., 22 Apr 2025), LQR steering (Wu et al., 23 Sep 2025), or key-conditioned perturbation (Zhang et al., 22 Jun 2026).

A plausible interpretation is that the intended topic behind the label “FlexQ” belongs to this broader design space of latent-space regularization and injection. If the intended referent is specifically a quadratic latent-coupling formulation, DLO is the nearest fully specified method in the provided material: it preserves classical initialization, decouples physical and latent updates, avoids backpropagation through the PDE solver during the latent step, and achieves strong robustness and geological realism in full waveform inversion benchmarks (Min et al., 12 Jun 2026). If, instead, the intended referent is a more general latent-injection paradigm, the surrounding literature shows that the same organizing idea now spans diffusion inversion, zero-shot editing, rectified-flow control, steganography, language-model inversion, and differentially private synthetic data generation.