GuidedBridge: Training-freely Improving Bridge Models with Prior Guidance

Conference: ICML 2026
arXiv: 2606.03119
Code: To be confirmed
Area: Image Generation / Diffusion Bridge / Guidance
Keywords: bridge model, prior guidance, frequency modulation, image translation, training-free

TL;DR

Addressing diffusion bridge models (data-to-data generation), the paper proposes training-free Prior Guidance (PG): by perturbing a clean prior to construct a "weak prior," the model extrapolates between the denoising results of strong and weak priors to amplify prior utilization. Further incorporating a U-shaped Frequency-modulated PG (FMPG) and a CFG-FMPG cascaded framework, the method consistently improves the FID of pre-trained bridge models such as DDBM / DBIM on tasks like Edges→Handbags, DIODE, and ImageNet inpainting without additional training or increased NFE.

Background & Motivation

Background: Diffusion guidance already has two mature paradigms: CFG uses "conditional vs. unconditional" denoising results for extrapolation to strengthen conditional alignment, while AG uses "well-trained vs. under-trained" denoising results for extrapolation to enhance score accuracy. Both essentially "create a quality gap between two denoising steps and extrapolate towards the higher-quality direction." Meanwhile, bridge models (DDBM, DBIM, I2SB, etc.) transform the generation process into data-to-data by conditioning on a clean prior \(\bm{x}_T\), which is more efficient for tasks with strong priors like image-to-image translation and restoration than diffusion starting from pure Gaussian noise.

Limitations of Prior Work: While CFG/AG can be directly ported to bridge models, they fail to utilize the actual key difference between bridge and diffusion: prior exploitation (leveraging the clean prior provided by \(\bm{x}_T\)). In other words, ported guidance only "strengthens conditions" or "corrects score errors" without specifically reinforcing whether the model truly utilizes the prior effectively. Additionally, AG requires training an extra under-trained network, which incurs significant costs.

Key Challenge: The greatest advantage of bridge models compared to diffusion lacks a corresponding guidance design. Furthermore, the SNR of bridge models follows a U-shaped trajectory (clean at both ends, noisiest in the middle), unlike the monotonically increasing SNR of diffusion. This makes using a constant guidance scale for all timesteps and frequency bands wasteful of the bridge's geometric structure.

Goal: (1) Design the first training-free guidance for bridge models where the guidance signal directly corresponds to "prior exploitation"; (2) Match the guidance intensity to the U-shaped SNR and frequency-domain behavior of the bridge; (3) Ensure the method remains applicable even when the prior itself is weak (e.g., masked regions in inpainting).

Key Insight: Since both CFG and AG follow the "create a worse denoising and extrapolate" logic, the most natural way to "create a gap" for a bridge model is to destroy the prior available to the model. Because the bridge has not seen a perturbed prior, it will yield a worse result. This "poor performance" exactly corresponds to "insufficient prior utilization," and the extrapolation direction naturally leads to "better prior utilization."

Core Idea: Training-freely apply a degradation operator \(\mathcal{H}\) (e.g., adding Gaussian noise, blurring, or JPEG compression) to \(\bm{x}_T\) and \(\bm{x}_t\) at each step. Use the difference between \(D_{\bm{\theta}}(\bm{x}_t,t,\bm{x}_T)\) and \(D_{\bm{\theta}}(\mathcal{H}(\bm{x}_t),t,\bm{x}_T)\) for guidance extrapolation, and schedule the guidance scale separately for high/low-frequency bands according to the U-shaped SNR.

Method

Overall Architecture

The input consists of a pre-trained bridge model (DDBM / DBIM) and a clean prior \(\bm{x}_T\) (e.g., an image to be translated or restored). During generation, instead of calling a single \(D_{\bm{\theta}}(\bm{x}_t,t,\bm{x}_T)\) to predict \(\bm{x}_0\), the following steps are performed at each sampling step:

1. Use a degradation operator \(\mathcal{H}\) to perturb the current noisy latent \(\bm{x}_t\) into a "weak prior" \(\mathcal{H}(\bm{x}_t)\) (while \(\bm{x}_T\) remains a clean conditional signal);
2. Simultaneously call \(D_{\bm{\theta}}(\bm{x}_t,t,\bm{x}_T)\) and \(D_{\bm{\theta}}(\mathcal{H}(\bm{x}_t),t,\bm{x}_T)\) to obtain strong and weak denoising results;
3. Replace the original denoising output with the extrapolation \(D_{\text{PG}} = D_{\text{weak}} + w_{\text{PG}}(D_{\text{strong}} - D_{\text{weak}})\);
4. FMPG further splits this extrapolation into low-frequency (LF) and high-frequency (HF) bands with \(w^{\text{LF}}_{\text{PG}}\) and \(w^{\text{HF}}_{\text{PG}}\) respectively;
5. For weak-prior tasks like inpainting, CFG is used for \(t\in[T,t_s)\) to generate coarse structures, then switched to FMPG for \(t\in[t_s,0)\) for detail refinement (CFG-FMPG cascade).

The entire process requires no training and no extra parameters. It simply replaces the denoising calls during sampling, and the NFE is strictly aligned with the original baseline (compensating for the extra forward pass per step by using fewer sampling steps).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Pre-trained Bridge Model + Clean Prior x_T"] --> B["Current Noisy Latent x_t"]
    B --> C["Prior Guidance (PG): Degradation operator H creates weak prior<br/>Extrapolation D_PG from Strong/Weak Denoising"]
    C --> D["FMPG: LF/HF Frequency Splitting<br/>Inverted U-shape for LF, U-shape for HF Scheduling"]
    D -->|"Prior is sufficient"| E["Update x_t, Iterative Sampling"]
    D -->|"Prior is weak (Inpainting)"| F["CFG-FMPG Cascade<br/>Early CFG for Structure → Late FMPG for Details"]
    F --> E
    E --> G["Output Sample x_0"]

Key Designs

1. Prior Guidance (PG): Creating Worse Denoising via Degraded Priors

PG fills the gap where bridge models lack guidance corresponding to their core advantage. It follows the "gap creation \(\to\) extrapolation" logic of CFG/AG but changes the source of the "gap" to the destroyed prior. Specifically, a training-free degradation operator \(\mathcal{H}\) (defaulting to adding Gaussian noise \(\bm{\epsilon}\sim\mathcal{N}(\bm{0},\bm{I})\), though blur or JPEG also work) is applied to each intermediate latent \(\bm{x}_t\) to obtain a weak version \(\mathcal{H}(\bm{x}_t)\), while the conditional signal \(\bm{x}_T\) stays clean. Since the bridge was never exposed to degraded priors during training, \(D_{\bm{\theta}}(\mathcal{H}(\bm{x}_t),t,\bm{x}_T)\) is inevitably worse than \(D_{\bm{\theta}}(\bm{x}_t,t,\bm{x}_T)\). The extrapolation:

\[D_{\text{PG}} = D_{\bm{\theta}}(\mathcal{H}(\bm{x}_t),t,\bm{x}_T) + w_{\text{PG}}\,\big(D_{\bm{\theta}}(\bm{x}_t,t,\bm{x}_T) - D_{\bm{\theta}}(\mathcal{H}(\bm{x}_t),t,\bm{x}_T)\big)\]

naturally points toward "more thorough prior utilization." Keeping \(\bm{x}_T\) clean in the weak term ensures the difference between the two results only stems from prior degradation rather than condition changes. Unlike AG, which requires training an under-trained network as a reference, PG is entirely training-free. Unlike CFG, which only strengthens "conditional alignment," PG directly addresses prior exploitation.

2. Frequency-modulated PG (FMPG): Scheduling Guidance by U-shaped SNR

Applying a constant \(w_{\text{PG}}\) uniformly across all timesteps and frequencies wastes the bridge's geometric structure. FMPG addresses this. By tracing the frequency-domain energy transfer of \(\Delta\bm{x}_t \to \Delta\bm{x}_0\), the authors found that the bridge's SNR is U-shaped. Consequently, at intermediate steps, high-frequency (HF) priors are heavily submerged in noise, and hard guidance here only amplifies noise. Conversely, low-frequency (LF) priors remain readable throughout the trajectory, making intermediate steps the ideal time to strengthen LF. FMPG splits PG into two paths using filters: \(I^{\text{LF}}\) for low frequencies with an inverted U-shaped \(w^{\text{LF}}_{\text{PG}}\) (amplified in the middle), and \(I^{\text{HF}}\) for high frequencies with a U-shaped \(w^{\text{HF}}_{\text{PG}}\) (compressed in the middle, amplified at ends). The LF path is formulated as:

\[D^{\text{LF}}_{\text{FMPG}} = I^{\text{LF}}[D_{\text{weak}}] + w^{\text{LF}}_{\text{PG}}\,\big(I^{\text{LF}}[D_{\text{strong}}] - I^{\text{LF}}[D_{\text{weak}}]\big),\]

with a similar form for \(D^{\text{HF}}_{\text{FMPG}}\). This encodes the U-shaped SNR of data-to-data geometry directly into the guidance schedule, aligning guidance intensity with generation dynamics—a concept that does not exist in diffusion's monotonic SNR.

3. CFG-FMPG Cascade: Saving Tasks with Weak Priors

When a prior is too weak for PG to create a meaningful difference—such as inpainting where a central 128×128 region is masked—other guidance is needed to establish the coarse structure first. CFG-FMPG splits the sampling trajectory into two segments: early \(t\in[T,t_s)\) uses CFG on class labels \(\bm{l}\) for conditional alignment to fill the masked area with semantic structure; subsequently \(t\in[t_s,0)\) switches to FMPG to exploit this CFG-produced result as a "strong enough prior" to restore high-frequency textures. Both segments share the same bridge network and trajectory, so NFE does not increase. This arrangement compensates for PG's inability to create a gap in empty masks and CFG's inability to restore high-frequency details.

Loss & Training

Completely training-free. All bridge checkpoints use official versions (DDBM uses the hybrid SDE-ODE sampler, DBIM uses pure ODE mode with \(\eta=0.0\)). Practice involves two steps: first searching for a degradation operator \(\mathcal{H}\) among {Noise, Blur, JPEG, Pooling}, then tuning \(w_{\text{PG}}\) (including FMPG components). The tuning cost is comparable to CFG/AG.

Key Experimental Results

Main Results

NFE is strictly aligned with the baseline (using fewer sampling steps to offset the extra forward pass).

Dataset	Metric	DBIM (Baseline)	DBIM+FMPG (Ours)	NFE
Edges→Handbags 64×64	FID ↓	1.74	1.07	20
Edges→Handbags 64×64	FID ↓	0.91	0.78	100
DIODE-Outdoor 256×256	FID ↓	4.99	3.20	20
DIODE-Outdoor 256×256	FID ↓	2.58	2.06	100
DIODE	LPIPS ↓	0.201	0.199	20
Edges→Handbags	MSE ↓	0.005	0.005	20

Effectiveness on DDBM: DDBM+PG reduces FID from 1.30 / 0.65 to 1.23 / 0.59 at NFE=150 / 300 on Edges→Handbags.

Ablation Study

Comparison of PG variants and FMPG on DIODE (Baseline: DBIM):

Configuration	NFE=10 FID	NFE=20 FID	NFE=40 FID	Note
DBIM (No guidance)	7.99	4.99	3.35	Original baseline
ECSI (Concurrent work)	6.83	4.12	-	Fast sampler only
DBIM+PG (Blur)	7.33	3.89	2.64	Degradation via Blur
DBIM+PG (Noise)	6.25	3.77	2.96	Degradation via Noise
DBIM+FMPG (Noise)	5.28	3.20	2.62	Best with freq modulation

Key Findings

• FMPG > PG > Baseline: Splitting constant \(w_{\text{PG}}\) into U-shaped and inverted U-shaped frequency schedules further reduces FID across all NFE, proving that bridge models require specialized guidance scheduling.
• Greater gains at lower NFE: On DIODE at NFE=10, FID dropped from 7.99 to 5.28 (-34%), compared to a -20% drop at NFE=100. FMPG is particularly valuable for fast sampling as it extracts prior information that the bridge model fails to utilize in fewer steps.
• CFG-FMPG is the key to inpainting: In ImageNet 128×128 mask repair, CFG-FMPG outperforms both pure CFG and simple hybrids. Qualitative results show simultaneous recovery of semantic layout and high-frequency textures.

Highlights & Insights

• The "Third Solution" for the Gap Paradigm: CFG addresses "conditional vs. unconditional," AG addresses "well-trained vs. under-trained," and PG addresses "clean vs. degraded prior." These are orthogonal, suggesting multiple guidances can be stacked.
• Training-free yet Physically Interpretable: The frequency scheduling is derived from the visualization of energy transfer in \(\Delta\bm{x}_t \to \Delta\bm{x}_0\), whereas diffusion's SNR doesn't support such concepts.
• Honest NFE Alignment: By reducing sampling steps to match baseline NFE, the authors avoid the pitfall of swapping 2× compute for performance, making PG/FMPG a true "free lunch."

Limitations & Future Work

• Per-task Search for Degradation: Whether noise, blur, or JPEG works best remains empirical.
• Manual Heuristic for Schedule: While based on SNR, the exact peaks and amplitudes of the U-shapes still require tuning.
• Task Scope: Evaluation was limited to image-to-image tasks. Its applicability to audio or video bridges remains unverified.
• Switching Timestep \(t_s\): In CFG-FMPG, the transition point requires manual tuning.

Related Work & Insights

• vs. CFG: CFG strengthens semantic alignment; PG strengthens prior utilization. They are complementary.
• vs. AG: AG requires an extra under-trained network; PG uses instant degradation during the forward pass, lowering deployment costs.
• vs. DDBM / DBIM: These provide backbones; Ours serves as a universal enhancement plugin without altering parameters.
• vs. ECSI: ECSI optimizes the sampler; PG/FMPG optimizes guidance. They are decoupled and potentially stackable.

Rating

• Novelty: ⭐⭐⭐⭐ Porting the "gap-extrapolation" paradigm to bridge models and encoding U-shaped SNR into guidance scheduling.
• Experimental Thoroughness: ⭐⭐⭐⭐ Covers multiple backbones and tasks with complete NFE curves, though lacks audio/video verification.
• Writing Quality: ⭐⭐⭐⭐ Clear relationship mapping between CFG/AG/PG; good coordination between frequency analysis and SNR.
• Value: ⭐⭐⭐⭐ A truly training-free, NFE-neutral plugin for bridge models with high practical utility.

Related Papers

• [CVPR 2026] Understanding, Accelerating, and Improving MeanFlow Training
• [ICLR 2026] Improving Diffusion Models for Class-imbalanced Training Data via Capacity Manipulation
• [CVPR 2026] Improving Diffusion Generalization with Weak-to-Strong Segmented Guidance
• [ICLR 2026] Stochastic Self-Guidance for Training-Free Enhancement of Diffusion Models
• [ICML 2026] Spectral Guidance for Flexible and Efficient Control of Diffusion Models