June 28, 2026

DanceOPD treats each image generation capability (text-to-image, local editing, global editing) as a velocity field and distills them into a unified student flow-matching model via on-policy sampling. For each training sample, the student routes to one frozen capability field, queries it at a low-noise on-policy state, and matches the resulting velocity with a local MSE loss. This avoids capability interference. Editing scores improve by up to 21.9% in specific categories while text-to-image metrics are preserved or improved by up to 2.0%. 64 upvotes on HF Daily Papers.

Qwen-Image-Agent addresses the context gap in text-to-image generation: user prompts are often underspecified, implicit, or require up-to-date knowledge. The framework iteratively constructs the full generation context via two modules: Context-Aware Planning (identifying missing context) and Context Grounding (gathering it via reasoning, web search, memory, and user feedback). The system achieves state-of-the-art on IA-Bench (45.4%), WISE-Verified (0.9020), and MindBench (0.42). 41 upvotes on HF Daily Papers.

Hansen and Wang reframe hallucination in visual world models as a data coverage problem rather than a model capacity problem. Three failure modes are identified: perceptual, action-marginalized, and scene-diverging. Three model-internal signals are derived that predict hallucination with approximately -0.80 Spearman correlation. They introduce MMBench2, a 427-hour 210-task dataset with ground-truth actions and rewards. Coverage-aware training and curiosity-reward fine-tuning enable adaptation to new environments with as few as 50 trajectories. 41 upvotes on HF Daily Papers.

This Qwen team paper challenges the assumption that verification is the easy half of generate-then-verify for coding agents. Studying four reward constructions across general coding, frontend, and long-horizon tasks, it finds no static reward function remains effective as policy capability grows. Verification must co-evolve with the generator, characterized across three axes: scalability, faithfulness, and robustness.

Tencent's Hunyuan team released UniRL, an open-source framework for unified RL post-training across LLMs, vision-language models, and diffusion/flow-matching models. It implements a single generate-score-advantage-update-sync loop usable across heterogeneous model families. Two algorithms ship with it: Flow-DPPO for diffusion/flow models using trust-region masks based on exact divergence, and DRPO for LLMs with a smoothed advantage-weighted quadratic regularizer.