Stable Diffusion 3 can leverages the same $3 \times 3$ LatentUnfold mosaic—no extra data, no fine-tuning—yet still effectively preserves subject identity.
Flux Already Knows
Activating Subject-Driven Image Generation without Training
Introduction
We propose a simple yet effective zero-shot framework for subject-driven image generation using a vanilla Flux model. By framing the task as grid-based image completion and replicating the subject image(s) in a mosaic layout, we activate strong identity-preserving capabilities—without any additional data, training, or inference-time fine-tuning. This “free lunch” approach is further enhanced by a cascade attention design and meta prompting technique, boosting fidelity and versatility.
Experimental results show that our Latent Unfoldd outperforms baselines across multiple metrics in benchmarks and human preference studies, with some trade-offs. It supports diverse edits such as logo insertion, virtual try-on, and subject replacement or insertion, demonstrating that a pre-trained text-to-image model can deliver high-quality, resource-efficient customization for downstream applications.
Method
Visual overview of our mosaic-based generation.
Our Approach: Latent Unfold
Our method, called Latent Unfold, constructs a mosaic-formatted $M \times N$ grid (e.g., $3\times3$) for zero-shot subject-driven generation. One panel is left blank as the target area, while the rest are filled with repeated subject images (as shown conceptually in the teaser and pipeline figures). This design enhances identity consistency.
The reference image is first encoded into a latent code $\mathbf{L}_{r}$. This code is then tiled into the mosaic latent $\mathbf{L}$, where the target panel (e.g., top-left) can be initialized with zeros: $$ \mathbf{L} \;=\; \begin{pmatrix} \mathbf{0} & \mathbf{L}_r & \cdots & \mathbf{L}_r \\ \vdots & \vdots & \ddots & \vdots \\ \mathbf{L}_r & \mathbf{L}_r & \cdots & \mathbf{L}_r \end{pmatrix} $$ This Latent Unfold approach is highly flexible, supporting both single-view and multi-view input subject images by incorporating different perspectives into the mosaic.
Key Contributions
⚡ Training-Free
No data, no fine-tuning, no inference-time tricks.
No data, no fine-tuning, no inference-time tricks.
🧩 Mosaic Completion
Subject identity emerges from repeated references.
Subject identity emerges from repeated references.
🔍 Cascade Attention
Multi-scale refinement boosts consistency.
Multi-scale refinement boosts consistency.
🧠 Meta Prompting
Uses LLMs to improve instruction quality.
Uses LLMs to improve instruction quality.
🧪 Diverse Editing
Subject & logo insertion, virtual try-on, style etc.
Subject & logo insertion, virtual try-on, style etc.
📈 Outperforms Baselines
Validated in benchmarks and human studies.
Validated in benchmarks and human studies.
Mosaic Image Completion
Our goal is to edit the target panel within the mosaic while preserving the reference subject images in other panels. We adapt a denoiser-based completion pipeline. The key steps are:
- Noise Level Consistency: To match the denoiser's expectations, we blend noise into $\mathbf{L}$ using the FLUX model's noise schedule: $\mathbf{L}_t = (1 - t)\mathbf{x}_0 + t\mathbf{L}$.
- Masking: At each denoising step $t$, the partially denoised latent $\mathbf{x}_t$ is combined with the noised mosaic $\mathbf{L}_t$ using a mask $\mathbf{m}$: $\widehat{\mathbf{x}}_t = \mathbf{m} \odot \mathbf{x}_t + (\mathbf{1} - \mathbf{m}) \odot \mathbf{L}_t$. This ensures only the target panel is updated with new content.
- Denoising: The combined latent $\widehat{\mathbf{x}}_t$, along with the text prompt $\mathbf{p}$, is fed into the T2I denoiser $\mathcal{D}$ to produce $\mathbf{x}_{t+\Delta t}$.
After $T$ steps, decoding $\mathbf{x}_T$ yields the mosaic image with the target panel newly generated and other panels faithful to the subject reference. This process is summarized below:
Latent Unfold Pipeline
