38 / Sep 15 – 21, 2025 · 11 RELEASES

AI News Week 38, 2025

Week 38 2025 AI releases.

Vision & Image (4)

FLUX.1-Kontext
Black Forest Labs · Text-to-Image
A context-aware image generation model that extends FLUX.1 capabilities with improved understanding of spatial relationships and compositional control. Designed for developers to generate images with enhanced semantic coherence and layout precision.
Flux
Black Forest Labs · Text-to-Image
Open-source text-to-image generation model offering fast, high-quality image synthesis with efficient inference capabilities. Available in multiple variants including a distilled version for resource-constrained deployments.
OmniGen2
VectorSpaceLab · Text-to-Image
OmniGen2 is a generative model framework designed for multi-modal content generation, building upon the original OmniGen architecture. The system appears to focus on unified generation capabilities across different content types.
DRA-Ctrl
Zhejiang University, Kunbyte AI, Ant Group, Hangzhou Normal University · Text-to-Image
Dimension-Reduction Attack (DRA-Ctrl) repurposes pretrained video generative models for controllable image synthesis tasks by treating condition-target pairs as two-frame videos, using mixup-based transitions and attention masking to bridge the gap between continuous video and discrete image generation.

InterActHuman
CUHK MMLab & ByteDance · Text-to-Video
A diffusion transformer framework for audio-driven multi-person video generation that uses iterative mask prediction to localize and align multi-modal inputs for each subject, enabling precise lip-sync and identity-specific audio conditioning. Supports dialogue videos, human-object interactions, and multi-concept customization across diverse domains including cartoons and animals.
VMem: Surfel-Indexed View Memory for Video Generation
University of Oxford · Text-to-Video
A plug-and-play memory module for video models that enables consistent autoregressive scene generation by anchoring past views to surface elements (surfels), improving long-term coherence and reducing computational cost compared to limited context window approaches.

SongBloom
Academic Research · Music Generation
SongBloom is a framework for full-length coherent song generation that interleaves autoregressive sketching with diffusion-based refinement, combining the fidelity of diffusion models with the scalability of language models to progressively extend and refine musical compositions.

AnimaX
Beihang University, Tsinghua University, University of Hong Kong, VAST · 3D Generation
Feed-forward 3D animation framework that uses joint video-pose diffusion models to animate arbitrary articulated 3D meshes from text descriptions, transferring motion priors from video diffusion models to skeleton-based animation via multi-view pose estimation and inverse kinematics.
AnimateAnyMesh
Alibaba DAMO Academy & Huazhong University of Science and Technology · 3D Generation
A feed-forward 4D foundation model that generates text-driven animations for arbitrary 3D meshes in seconds using a novel DyMeshVAE architecture that disentangles spatial-temporal features and a Rectified Flow-based generation strategy, trained on a new DyMesh Dataset of 4M+ annotated dynamic mesh sequences.

ComfyUI-OmniGen2
Yuan-ManX · AI Infrastructure
ComfyUI extension integrating OmniGen2, a multimodal generative model capable of image and text generation from diverse inputs. Extends ComfyUI workflow system with OmniGen2's unified generation capabilities.

Align Your Flow
NVIDIA Research (Toronto AI Lab) · Research Paper
Proposes flow map distillation method that unifies consistency models and flow matching, achieving state-of-the-art few-step image generation on ImageNet while maintaining quality across variable step counts. Analytically proves consistency models suffer from error accumulation in multi-step sampling and introduces new training objectives with autoguidance and adversarial finetuning techniques.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.