Triple-Flow Fused PerGeneration in CusEnhancer

• The paper demonstrates that triple-flow fused PerGeneration integrates DDIM inversion, backward ResInversion, and forward customized-generation to enrich scene, identity, and control signals without retraining.
• Triple-Flow Fused PerGeneration is a mechanism that merges reconstruction and customized generation by leveraging three distinct diffusion flows, ensuring seamless integration of auxiliary representations.
• The method addresses key challenges like scene degradation and control incompatibility by employing a pivot-centered fusion that aligns both backward and forward latent spaces for enhanced performance.

Triple-flow fused PerGeneration is the core mechanism behind BiMD in the CusEnhancer / CustomEnhancer framework, introduced in “CusEnhancer: A Zero-Shot Scene and Controllability Enhancement Method for Photo Customization via ResInversion” (Ren et al., 25 Sep 2025). It is a diffusion-based fusion procedure for a personalized diffusion model $G_p$ that is already specialized for identity customization, with the specific purpose of injecting additional scene, identity, and control information into its generation process without retraining. The method is termed “triple-flow fused” because it unifies three diffusion trajectories—one pivot trajectory, one backward reconstruction trajectory, and one forward customized-generation trajectory—into a single generation procedure. In the supplied literature, this mechanism is distinct from “Transition Flow Matching,” which learns a transition flow $X^\theta(x_t,t,r)$ as a global state-to-state map and does not use the phrase “triple-flow fused PerGeneration” (Ma, 16 Mar 2026).

1. Position within the CusEnhancer / CustomEnhancer framework

CusEnhancer / CustomEnhancer is presented as a zero-shot enhancement pipeline for existing identity customization models. Its stated objective is to augment current personalized diffusion systems that otherwise exhibit degraded scenes, insufficient control, and suboptimal perceptual identity. The framework leverages face swapping techniques and a pretrained diffusion model to obtain additional representations in a zeroshot manner for encoding into personalized models, and it uses triple-flow fused PerGeneration as the mechanism that integrates those representations into the denoising process (Ren et al., 25 Sep 2025).

The motivation is organized around three failures of current personalized diffusion models such as PhotoMaker and InstantID. First, scene degradation arises because large-scale fine-tuning on face-centric data makes the model over-focus on the face and lose general scene generation ability. Second, control incompatibility occurs because these personalized models are not naturally compatible with control modules like ControlNet trained for base SDXL, so controllability is lost unless one retrains control modules for each personalized model. Third, an identity gap remains because the identity encoder used by these methods captures mostly high-level semantic identity, but misses concrete perceptual facial details.

Within that problem setting, triple-flow fused PerGeneration serves as the plug-in enhancement method that borrows scene priors from pretrained SDXL, borrows perceptual identity details from face swapping, and injects both into the generation process of a personalized model without training. The central difficulty is not merely acquiring these auxiliary representations, but integrating them without destroying the personalized model’s own identity generation ability. This establishes the methodological role of the triple-flow construction: it is an integration mechanism rather than a standalone personalization model.

2. Constituent flows and the fused denoising update

The BiMD / triple-flow framework consists of three synergistic flows in the personalized model’s denoising process. The first is the pivot flow, described as the reference trajectory obtained by standard DDIM inversion. It provides the anchor sequence

$Z_0^*, Z_1^*, \dots, Z_T^*,$

which defines the “pivotal” latent space into which the other flows are aligned. The second is the backward ResInversion flow, which reconstructs the input image $I_P$ from the pivot anchor in a customization-free / partial-functional space rather than fully overwriting the personalized model. The third is the forward customized-generation flow, namely the personalized model’s own identity-customization trajectory that produces the customized person-specific generation (Ren et al., 25 Sep 2025).

The interaction among these flows is explicitly described as pivot-centered fusion. The pivot flow is the shared anchor; the backward flow contributes reconstructed residual information from ResInversion; and the forward flow contributes identity customization information from $G_p$. The paper writes the BiMD sampling update as

$$Z_{t-1}=\frac{1}{\sqrt{\alpha_t}}\bigg\{Z_t-\frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}_t}} \Big\{\epsilon_\theta(Z_t,t)+\epsilon_{Res}(Z_t,t)-\epsilon_{G_{p,cus}}(Z_t,t)\Big\}\bigg\},$$

where $\alpha_t$ is the noise schedule parameter, $\epsilon_\theta$ is the model’s predicted noise, $\epsilon_{Res}$ is the noise from the ResInversion backward flow, and $\epsilon_{G_{p,cus}}$ is the noise corresponding to the personalized model under customization conditions.

The significance of this update is that the latent trajectory is no longer driven by the base model alone, but by a composite noise field that integrates reconstruction and customization. A common misunderstanding is to interpret the backward branch as a full inversion that recreates the entire behavior of the personalized model. The paper explicitly rejects that interpretation: the backward flow operates in a space where the model generates non-specific identity, thereby preserving the model’s own customized generation ability from being overwritten.

3. Compatible counter-directional latent spaces

A central conceptual claim is that the method identifies two compatible counter-directional latent spaces. The first is the customization space, defined as the forward latent space in which the personalized model $X^\theta(x_t,t,r)$0 performs identity-customized generation. The second is the customization-free space, defined as the backward or reconstruction latent space used by ResInversion, where the model reconstructs the input without imposing or overwriting the personalized identity behavior (Ren et al., 25 Sep 2025).

These spaces are termed counter-directional because one corresponds to forward generation and the other to backward inversion or reconstruction. They are nevertheless compatible because both are organized around the same pivot trajectory $X^\theta(x_t,t,r)$1. The paper’s key point is that the backward inversion does not attempt to reconstruct the full generative mechanism; instead, it reconstructs within a space that remains compatible with the forward customized-generation flow. Their combination occurs by intervening at the pivot flow: the pivot latent serves as the shared carrier, and the backward and forward information are fused as complementary corrections to that latent.

The pivotal latent space itself is the anchor latent trajectory $X^\theta(x_t,t,r)$2 produced by DDIM inversion. Its role is to provide a stable backbone around which both reconstruction and generation can be aligned. The method manipulates this pivotal latent space through two interventions. Backward intervention is performed by ResInversion, which reconstructs the input image and encodes its information into latent residuals. Forward intervention is performed by the personalized model’s customized generation, which injects identity-conditioned information. Rather than replacing the pivot, BiMD modifies the denoising update at the pivot using the combined noise terms. This suggests that the pivotal space functions as a compatibility layer between reconstruction-side and customization-side dynamics.

4. Unification of generation, reconstruction, and control injection

One of the main conceptual contributions is the claim that triple-flow fused PerGeneration unifies generation and reconstruction by transferring information through both the forward and inverse diffusion processes. In ordinary DDIM inversion or Null-text inversion, reconstruction is used mainly to recover an input image, while generation happens separately. In this framework, reconstruction itself becomes part of generation. The procedure consists of obtaining a pivot trajectory via DDIM inversion, reconstructing residual information via ResInversion, and then feeding both the reconstructed information and the personalized generation signal into the same denoising chain (Ren et al., 25 Sep 2025).

The final sample is therefore not produced by a separate edit-after-reconstruction step. Instead, it emerges from a single unified diffusion process that simultaneously reconstructs image content, preserves personalized identity, and incorporates extra scene and control representations. This is the paper’s operational meaning of “triple-flow fused PerGeneration”: three flows are fused into a single per-sample generation procedure.

The same fused design is used to enable zero-shot and training-free controllability for personalized models. The framework exploits the fact that pretrained SDXL already has control modules such as OpenPose and Canny-based ControlNet. Instead of retraining these modules for each personalized model, the method injects the control capability through the SDXL branch and then propagates it into the personalized generation pipeline. The resulting representation is written as

$X^\theta(x_t,t,r)$3

where $X^\theta(x_t,t,r)$4 is the scene representation from pretrained SDXL, $X^\theta(x_t,t,r)$5 is the perceptual facial representation from face swapping, and $X^\theta(x_t,t,r)$6 denotes arbitrary control modules trained for SDXL.

Under this formulation, the personalized model receives scene priors, perceptual identity details, and control guidance in a zero-shot and training-free way. The stated consequence is that the control module is not forced to adapt to the personalized model; instead, the personalized model is enhanced through the SDXL control pathway. The framework is therefore said to support controllability over the human subject and over non-primary elements such as the background or environment.

5. ResInversion as the practical backward flow

ResInversion is introduced to address the high time complexity of null-text inversion. Its role is auxiliary with respect to the main conceptual novelty, but critical for making the backward branch practical in SDXL-scale personalized models. The method uses DDIM inversion plus a pre-diffusion residual-noise rectification,

$X^\theta(x_t,t,r)$7

where $X^\theta(x_t,t,r)$8 is the removed noise from DDIM inversion and $X^\theta(x_t,t,r)$9 is the denoised component from the pre-diffusion process (Ren et al., 25 Sep 2025).

This construction avoids iterative optimization and is described as much faster while maintaining reconstruction fidelity. The paper reports up to $Z_0^*, Z_1^*, \dots, Z_T^*,$0 speedup over NTI, and states that the full pipeline becomes $Z_0^*, Z_1^*, \dots, Z_T^*,$1–$Z_0^*, Z_1^*, \dots, Z_T^*,$2 faster than NTI-based implementations. It also states that experiments demonstrate that CustomEnhancer reach SOTA results at scene diversity, identity fidelity, training-free controls, while also showing the efficiency of ResInversion over NTI.

The practical significance of ResInversion is not simply acceleration in isolation. Its broader function is to supply a fast, faithful backward flow that provides the reconstruction-side latent information needed by triple-flow fusion. A plausible implication is that, without such an inversion mechanism, the backward half of the unified generation-and-reconstruction design would remain too slow for routine use in personalized SDXL pipelines.

6. Relation to prior personalization methods and adjacent flow paradigms

Relative to earlier personalization methods, the framework is positioned as a plug-in enhancement method rather than a replacement for identity customization models. Compared with DreamBooth, Textual Inversion, and LoRA, the supplied description emphasizes that those methods require test-time fine-tuning, are slow and data-dependent, and do not inherently solve scene diversity or control compatibility. Compared with PhotoMaker and InstantID, the description emphasizes degraded scene generation, poor compatibility with control modules, and identity gaps due to feature extraction limitations. CustomEnhancer is said to improve those systems by combining scene capability from SDXL, perceptual identity from face swapping, reconstruction and generation fused via BiMD, and controllability inherited from SDXL control modules (Ren et al., 25 Sep 2025).

The phrase “Triple-Flow Fused PerGeneration” may also invite confusion with more general flow-based generative modeling. The most relevant adjacent terminology in the supplied literature is “Transition Flow Matching,” which proposes learning a transition flow rather than a local velocity field. That paper defines a mapping

$Z_0^*, Z_1^*, \dots, Z_T^*,$3

so that a current state at time $Z_0^*, Z_1^*, \dots, Z_T^*,$4 can be mapped directly to a future state at time $Z_0^*, Z_1^*, \dots, Z_T^*,$5, enabling single-step generation with $Z_0^*, Z_1^*, \dots, Z_T^*,$6 or multi-step generation on any chosen time grid (Ma, 16 Mar 2026).

The relationship, however, is conceptual rather than terminological. “Transition Flow Matching” is described as a global-flow method that unifies local-flow and mean-velocity perspectives through the Transition Flow Identity, and the paper explicitly notes that the exact phrase “Triple-Flow Fused PerGeneration” does not appear there. Triple-flow fused PerGeneration is instead a pivot-centered fusion mechanism for personalized diffusion, in which a pivot flow, a backward reconstruction flow, and a forward customized-generation flow are combined to inject scene priors, perceptual identity features, and control signals into an existing personalized model without retraining. This distinction is important because the two frameworks address different technical problems: one concerns plug-in enhancement of personalized diffusion models, while the other concerns direct state-to-state generation over arbitrary future horizons.