🚀 Excited to share our CVPR 2026 paper:
RewardFlow: Generate Images by Optimizing What You Reward
🖼️ RewardFlow is an inversion-free framework for image editing and compositional generation that steers pretrained diffusion and flow-matching models at inference.
✨ Key contributions
🔹 A unified reward-guided sampling framework that combines signals for semantic alignment perceptual fidelity localized grounding object consistency and human preference
🔹 A differentiable VQA-based reward for fine-grained attribute-level supervision
📢 We invite submissions to the @icmlconf 2026 Workshop From Frames to Stories (F2S): Toward Reliable, Controllable, and Trustworthy Long-Horizon Video Generation
✨The workshop brings together researchers advancing long-horizon video generation, with a focus on long-context modeling, state representation, interactive control, and trustworthy evaluation.
🎬 We welcome submissions of recent work in related areas, including both published and unpublished papers.
📅 Submission deadline: April 30, 2026 (AoE)
🌐 Website: https://t.co/IduSx5sP1f
#VideoGeneration #GenerativeAI #ICML2026 #ICML
PLAN Lab #CVPR2026 🎉
7 papers accepted across 3D/2D grounding and hallucination mitigation, long-horizon robotics, 3D articulation, and video/image generation.
Follow us on 🤗for data + model releases
https://t.co/i6SwwVvmHc | https://t.co/4vfEsjEClj
🔥Congratulations to all co-authors! @adheeshjuvekar@Eman__Hussein@onkar_s_13@yu_tianjiao@xinzhuo_li@YingShen_ys Yuanzhe and Makanjuola 👏👏🏽👏🏿
🚀Introducing UniDFlow, a unified discrete flow-matching model for multimodal understanding, generation, and instruction-based editing under a single probabilistic interface.
🤗: https://t.co/9cZvFJgBo8
🌐: https://t.co/kYt09h3qgk
#Multimodal#Diffusion#ImageGeneration#VLM
@ylecun@giffmana Why so are we defined that process like that. I think if its multi Mode gaussian it will not be true for all the identity… Because each of the Eigan vectors can be point in multiple directions.
@TongPetersb yes workings on the similar concept where the features can be used for both generation and understanding purpose but in quantise latents space
@iScienceLuvr I am also working on something similar to kinda thing, just for Temporal Cohearance.
Also, in this paper if they used LFQ instead of VQ that would make more sense to me because they can increase the vocab size up to 100K.
Which can be easy to search in the latent space.
@lilyxu0@brwilder@lyeskhalil@MilindTambe_AI This is a nice idea.
I will be there at ICLR 25, even though I have a similar idea to work on for combinatorial diffusion molding through actor-critic sampling.
Let us connect at #ICLR25.