Fréchet Distance optimized as FD-loss in representation space by decoupling population size from batch size improves generator quality, enables one-step generation from multi-step models, and motivates a multi-representation metric FDr^k.
hub Canonical reference
BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Dataset
Canonical reference. 71% of citing Pith papers cite this work as background.
73
Pith papers citing it
Background
71% of classified citations
abstract
Unifying image understanding and generation has gained growing attention in recent research on multimodal models. Although design choices for image understanding have been extensively studied, the optimal model architecture and training recipe for a unified framework with image generation remain underexplored. Motivated by the strong potential of autoregressive and diffusion models for high-quality generation and scalability, we conduct a comprehensive study of their use in unified multimodal settings, with emphasis on image representations, modeling objectives, and training strategies. Grounded in these investigations, we introduce a novel approach that employs a diffusion transformer to generate semantically rich CLIP image features, in contrast to conventional VAE-based representations. This design yields both higher training efficiency and improved generative quality. Furthermore, we demonstrate that a sequential pretraining strategy for unified models-first training on image understanding and subsequently on image generation-offers practical advantages by preserving image understanding capability while developing strong image generation ability. Finally, we carefully curate a high-quality instruction-tuning dataset BLIP3o-60k for image generation by prompting GPT-4o with a diverse set of captions covering various scenes, objects, human gestures, and more. Building on our innovative model design, training recipe, and datasets, we develop BLIP3-o, a suite of state-of-the-art unified multimodal models. BLIP3-o achieves superior performance across most of the popular benchmarks spanning both image understanding and generation tasks. To facilitate future research, we fully open-source our models, including code, model weights, training scripts, and pretraining and instruction tuning datasets.
hub tools
citation-role summary
background 24
baseline 4
dataset 2
other 1
citation-polarity summary
claims ledger
- abstract Unifying image understanding and generation has gained growing attention in recent research on multimodal models. Although design choices for image understanding have been extensively studied, the optimal model architecture and training recipe for a unified framework with image generation remain underexplored. Motivated by the strong potential of autoregressive and diffusion models for high-quality generation and scalability, we conduct a comprehensive study of their use in unified multimodal settings, with emphasis on image representations, modeling objectives, and training strategies. Ground
co-cited works
representative citing papers
RankE co-evolves AR policy and decoder via alternating ranking optimization, improving both FID and CLIP scores on LlamaGen-XL and Janus-Pro where policy-only RL degrades FID.
INSET embeds images as native tokens in interleaved instructions, outperforming prior methods on multi-image consistency and text alignment as complexity grows.
UniPath adaptively models coordination-path diversity in unified multimodal models by training a path-conditioned executor and using a lightweight planner for input-dependent selection, improving performance over fixed strategies.
Do-Undo Bench is a new evaluation task and dataset that forces models to simulate forward action effects and then undo them to measure genuine action understanding in image generation.
AIA loss teaches unified multimodal models task-specific cross-modal attention patterns to reduce conflicts between image understanding and generation without architecture decoupling.
XTC-Bench reveals that strong performance on generation or understanding tasks in unified multimodal models does not guarantee cross-task semantic consistency, which instead depends on how tightly coupled the learning objectives are across modalities.
Fine-tuning multimodal models on a new synthetic spatial benchmark improves generative spatial compliance on real and synthetic tasks and transfers to better spatial understanding.
Unified multimodal models exhibit pseudo-unification due to modality-asymmetric entropy encoding and pattern-split responses between text and image generation.
UniTac is the first unified multimodal model for cross-sensor tactile understanding and generation, using dual-level representations, two new understanding tasks, and a two-stage training paradigm with sensor-prior sampling to achieve SOTA understanding and realistic cross-sensor generation.
A masked discrete diffusion model adds token editing at inference and grouped cross-entropy training to reach 0.90 GenEval, 86.9 DPG, and 10.76 HPSv3 scores.
CSFlow derives inference-time timestep weights for flow matching by matching per-step frequency content to human CSF, yielding 4.7% FID reduction and smaller gains on IS and GenEval.
Introduces ProductWebGen benchmark for multimodal product webpage generation, compares editing-based vs unified-model workflows on 500 samples, and releases ProductWebGen-1k SFT dataset.
Representation Forcing enables end-to-end pixel-space unified multimodal models by making visual representation prediction a native autoregressive generation target that guides subsequent pixel diffusion in the same backbone.
Lumos-Nexus is a training-efficient video generation framework using two-stage alignment of a lightweight model followed by progressive frequency bridging to a high-fidelity generator in homogeneous latent space, plus the new VR-Bench for reasoning evaluation.
GenClaw introduces a three-stage code-driven workflow for agentic image generation that inserts programmatic sketches between linguistic reasoning and pixel synthesis.
CVQ replaces patch-wise vector quantization with channel-wise quantization of feature maps, enabling a next-channel autoregressive model that reports 100% codebook utilization and text-to-image scores of DPG 86.7 and GenEval 0.79.
DIVA factorizes visual representations in unified multimodal models into shared and unique components via complementary information flows and mutual information estimation to convert representation divergence into mutual reinforcement between understanding and generation branches.
Introduces TRACE-Edit dataset and evaluation protocol demonstrating semantic degradation of structural variables during VLM-to-DiT alignment in flow-matching video editors.
Lance presents a dual-stream mixture-of-experts model with modality-aware positional encoding and staged multi-task training that outperforms prior open-source unified models on image and video generation while keeping strong understanding performance.
RAE v2 reaches gFID 1.06 on ImageNet-256 in 80 epochs by combining multi-layer encoder sums, complementary REPA targets, and free guidance via output reparameterization.
LatentUMM proposes dual latent alignment at modality and capacity levels plus latent dynamics stabilization to reduce semantic drift and improve consistency in unified multimodal models.
RaPD enables resolution-agnostic image generation by diffusing in a semantics-enriched continuous Neural Image Field latent space using semantic guidance and a coordinate-queried attention renderer.
A unified visual conditioning approach fuses semantic and appearance features before VLM processing, with two-stage training and slot-wise regularization, to improve consistency in multi-reference image generation.
citing papers explorer
-
RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution
RankE co-evolves AR policy and decoder via alternating ranking optimization, improving both FID and CLIP scores on LlamaGen-XL and Janus-Pro where policy-only RL degrades FID.
-
Improved Baselines with Representation Autoencoders
RAE v2 reaches gFID 1.06 on ImageNet-256 in 80 epochs by combining multi-layer encoder sums, complementary REPA targets, and free guidance via output reparameterization.
-
DeCo: Frequency-Decoupled Pixel Diffusion for End-to-End Image Generation
DeCo decouples high- and low-frequency generation in pixel diffusion via a DiT plus lightweight decoder and a frequency-aware flow-matching loss, reaching FID 1.62 at 256x256 and 2.22 at 512x512 on ImageNet while closing the gap to latent diffusion methods.
-
Semantic-guided Gaussian Splatting for High-Fidelity Underwater Scene Reconstruction
SWAGSplatting augments 3D Gaussian Splatting with CLIP-derived semantic features, a semantic consistency loss, and adaptive primitive reallocation to achieve higher-fidelity reconstruction in low-visibility underwater scenes.