MultiCrafter: High-Fidelity Multi-Subject Generation

• MultiCrafter is a framework for high-fidelity multi-subject generation that ensures subject fidelity, prompt consistency, and preference alignment.
• It employs explicit positional supervision, a Mixture-of-Experts LoRA architecture, and online reinforcement learning to mitigate attribute leakage and enhance identity retention.
• Empirical results show a 28.3% improvement in subject fidelity along with competitive performance across various metrics such as CLIP-I, DINO-I, and CLIP-T.

MultiCrafter is a framework for high-fidelity multi-subject image generation that aims to synthesize user-provided subjects in a single image while preserving subject fidelity, ensuring prompt consistency, and aligning with human aesthetic preferences. It is positioned against existing methods built on the In-Context-Learning paradigm, which are described as relying on simple reconstruction-based objectives and therefore exhibiting severe attribute leakage and weak alignment with nuanced human preferences. MultiCrafter addresses these limitations through three coupled components: explicit positional supervision for spatially disentangled attention, a Mixture-of-Experts architecture for scenario-dependent capacity allocation, and an online reinforcement learning framework with identity-aware scoring and a stable training strategy tailored for MoE models (Wu et al., 26 Sep 2025).

1. Problem setting and failure mode

Multi-subject image generation requires the joint satisfaction of three criteria: subject fidelity, prompt consistency, and preference alignment. In the formulation adopted by MultiCrafter, the central failure mode of prior ICL-based approaches is attribute leakage, also described as “attention bleeding,” in which distinctive attributes of one subject leak into another and degrade fidelity. The framework reports that the root cause of this failure is a significant entanglement of attention between different subjects during the generation process, particularly in spatially critical regions such as the “double blocks” of the Flux/DiT architecture (Wu et al., 26 Sep 2025).

This diagnosis shifts the problem away from a purely reconstruction-centered view. A common misconception is that simple reconstruction losses are sufficient if the reference conditioning is strong enough. MultiCrafter rejects that premise by treating subject interference as an internal attention-allocation problem rather than merely a data or prompt problem. The resulting design therefore targets the geometry of attention, the capacity of the adaptation mechanism, and the reward structure used for post-training alignment.

2. Spatially disentangled attention

The first component is a spatially disentangled attention mechanism implemented through explicit positional supervision. MultiCrafter introduces “Identity-Disentangled Attention Regularization” (IDAR), which supervises attention maps so that each subject attends to its own spatial region and not to overlapping regions associated with other subjects. Ground-truth masks for each subject region are used during training, while inference remains unchanged in efficiency because the supervision is training-only (Wu et al., 26 Sep 2025).

For subject $i$ at double block layer $k$, the attention map is defined as

$$m^i_k = \mathrm{Softmax}\left(\frac{\mathbf{Q}_{k,i}\mathbf{K}^T}{\sqrt{d}}\right).$$

These per-block maps are averaged across $K$ double blocks to obtain a mean attention map $\hat{M}_i$. The regularization is then applied with a Dice loss against the ground-truth binary mask $M_i$:

$$\mathcal{L}_{attn} = \sum_{i=1}^{N} \left( 1 - \frac{2 \sum_{j} (\hat{M}_{i,j} \cdot M_{i,j}) + \epsilon}{\sum_{j} \hat{M}_{i,j} + \sum_{j} M_{i,j} + \epsilon} \right).$$

The training objective combines diffusion and attention losses:

$L = L_{diff} + \lambda \cdot L_{attn}.$

The stated effect is to force spatially disentangled attention for each subject and to align attention regions directly with subject positions, thereby eliminating attribute leakage and improving subject-specific fidelity. This suggests that MultiCrafter treats localization not as an auxiliary signal but as the primary structural constraint needed to preserve identity under multi-subject composition.

3. Mixture-of-Experts adaptation

The second component is a Mixture-of-Experts attention architecture introduced to address the limited capacity of simple LoRA-based fine-tuning. MultiCrafter argues that standard LoRA lacks the capacity to handle the variety of multi-subject layouts, appearances, and scene complexities encountered in practice. Its solution is a Mixture-of-Experts LoRA mechanism integrated into the FFN layers of the Flux backbone, while the remaining layers retain standard LoRA for efficiency (Wu et al., 26 Sep 2025).

For an input vector $h$, gating logits over $N_e$ experts are produced by

$k$0

and the expert-augmented FFN output is

$k$1

Each expert is therefore a parameterized LoRA block, and only a sparse subset of experts is activated for a given instance.

The architectural claim is not merely increased parameter count. Different experts are intended to specialize in different scenarios, such as distinct spatial arrangements or subject types, while sparse activation preserves inference speed. A plausible implication is that the MoE component functions as a scenario-conditioned capacity allocator: subject disentanglement supplies structural supervision, and the experts supply the adaptive representational headroom needed to realize that supervision across heterogeneous compositions.

4. Online reinforcement learning and identity-aware scoring

The third component is an online reinforcement learning stage for preference alignment. MultiCrafter uses “Identity-Preserving Preference Optimization” through online RL to optimize aesthetic preference, text-image alignment, and multi-subject identity fidelity jointly. The framework is built on MixGRPO, described as using windowed SDEs on flow-matching models for stability and efficiency, and it introduces Group Sequence Policy Optimization (GSPO) to stabilize sequence-level policy ratios in the presence of MoE stochasticity (Wu et al., 26 Sep 2025).

The reward is defined as

$k$2

Here, $k$3 measures text-image alignment, $k$4 measures aesthetic quality, and $k$5 measures multi-subject identity consistency. For human subjects, facial embeddings are extracted from both reference images and all faces in the generated image, a similarity matrix is constructed with cosine similarity, and the Hungarian matching algorithm is used to obtain an optimal one-to-one correspondence. For objects, Florence-2 and SAM2 are used to locate and extract the target region, and DINOv2 embeddings are used for similarity. The paper states that this matching-based design prevents “reward hacking” in which the model blurs identity or duplicates faces.

The significance of this stage is twofold. First, it replaces purely reconstruction-oriented supervision with direct optimization against human-valued criteria. Second, it makes multi-subject fidelity measurable in a setting where naïve reward models are otherwise unreliable. The design implies that identity preservation is not treated as a side effect of generation quality; it is explicitly operationalized as a reward-bearing objective.

5. Empirical results and ablation evidence

The experimental results reported for MultiCrafter indicate that it considerably outperforms prior state-of-the-art methods, including UNO, DreamO, and OmniGen, in both multi-human and multi-object settings (Wu et al., 26 Sep 2025). The most specific headline result is that subject fidelity, measured by Face-Sim, improves by 28.3% over the closest competitor. The framework is also reported to achieve either the best or highly competitive performance on CLIP-I, DINO-I, CLIP-T, and HPS.

Qualitative findings are described in terms of reduced attribute leakage, improved retention of unique facial characteristics, stronger semantic alignment to prompts, and less “oily” artificial stylization. These observations are consistent with the three-part structure of the method: IDAR targets leakage directly, MoE-LoRA increases scenario coverage, and the RL stage adds preference alignment beyond pixel reconstruction.

Ablation studies are presented as evidence that each component is necessary. Removing any of IDAR, MoE, or IPPO degrades performance, especially subject fidelity. Adding IDAR is reported to massively improve subject fidelity, MoE-LoRA boosts both fidelity and prompt alignment, and IPPO further enhances both fidelity and alignment with preferences. Within the logic of the paper, this ablation pattern supports a modular interpretation: spatial disentanglement addresses the dominant failure mode, expert routing addresses capacity limitations, and RL refines the generator toward human-preferred outputs.

6. Relation to prior paradigms and broader significance

MultiCrafter is explicitly framed as a response to limitations of ICL-based multi-subject generation. Those earlier systems are described as being constrained by simple reconstruction-based objectives, which do not adequately regulate subject-specific attention and do not encode nuanced human preference signals. MultiCrafter therefore combines localized supervision, adaptive expert specialization, and post-training alignment in a single framework rather than relying on a single training signal to handle all three desiderata (Wu et al., 26 Sep 2025).

Its principal conceptual claim is that high-fidelity multi-subject generation requires coordinated control over where the model attends, how much specialized capacity it can allocate to diverse scenarios, and how output quality is judged after generation. The paper’s analysis of attribute leakage as an attention-entanglement problem is especially important because it redefines a long-standing failure mode in mechanistic rather than purely phenomenological terms. This suggests a broader methodological lesson for multi-subject generation: fidelity degradation may arise less from insufficient conditioning alone than from the absence of explicit structural constraints on how conditioned subjects compete for attention within the denoising process.