Research map / Vision-language and multimodal / Multimodal LLMs

EVLF: Early Vision-Language Fusion for Generative Dataset Distillation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Bias-aware multi-image robustness · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EVLF, a lightweight cross-attention module that fuses text class embeddings with VAE image latents before diffusion, as a plug-and-play addition to diffusion-based dataset distillation pipelines (D^4M, MGD^3).

Problem

Diffusion-based dataset distillation injects class semantics via late-stage cross-attention during denoising, so textual prompts dominate and visual latents are diminished, yielding over-corrected, visually distorted synthetic samples.

Approach

Move vision-language fusion to the encoder-to-backbone interface before denoising: a lightweight cross-attention (image tokens as queries, text tokens as keys/values) produces a fused initial latent trained with an MSE fidelity loss to the image latent plus an InfoNCE semantic alignment loss; the denoiser is optionally fine-tuned on fused latents.

Key results

• D^4M+EVLF improves CIFAR-10 at IPC 10 from 37.6% to 45.7%, an 8.1-point gain over D^4M.
• On ImageIDC at IPC 10 with ResNetAP-10, D^4M+EVLF reaches 57.3% vs 47.7% for D^4M (+9.6).
• On ImageNet-1K at IPC 50, MGD^3+EVLF reaches 61.9% vs 60.3% for MGD^3.

Surprising findings

• Denoiser fine-tuning on fused embeddings is reported to help, but on its own the cross-attention variant without fine-tuning underperforms the fine-tuning-only variant on ImageIDC at IPC 10 (51.1% vs 54.1%).

Limitations

• Focuses on class-level conditioning; does not address instance-level or multi-label scenarios.
• Denoiser fine-tuning is optional and needed for some pipelines to align the fused latent distribution.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 ImageNet-1K ImageNette ImageWoof ImageIDC Tiny-ImageNet Dogs Flowers

Benchmarks: CIFAR-10 CIFAR-100 ImageNette ImageWoof ImageIDC Tiny-ImageNet ImageNet-1K

Metrics: top-1 accuracy coverage

Baselines: D^4M MGD^3 MinimaxDiffusion DiT DM IDC-1 Herding Random SRe^2L RDED KRR-ST FRePo

Builds on: D^4M MGD^3 MinimaxDiffusion Latent Diffusion Models Diffusion Transformers

Closest papers by meaning

• Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.87)
• Florence-VL: Enhancing Vision-Language Models with Generative Vision Encoder and Depth-Breadth Fusion (CVPR 2025 · similarity 0.86)
• Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.86)
• Beyond Words: Augmenting Discriminative Richness via Diffusions in Unsupervised Prompt Learning (CVPR 2025 · similarity 0.85)
• Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.85)
• Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.85)