Last week in Image & Video Generation
Lightricks' LTX-2.3 hits 1080x1920 portrait mode, while PKU's Helios runs 14B video models in real-time on a single GPU.
The open-source AI landscape for image and video generation accelerated dramatically last week, headlined by Lightricks' LTX-2.3 model. This release features significantly better prompt following and a native portrait mode capable of generating images up to 1080x1920 resolution. The community response was immediate, with users quickly creating a free local video editor (LTX Desktop), porting it to Linux, and publishing detailed prompting guides and 12GB GGUF workflows for various generation tasks.
Parallel breakthroughs came from PKU-YuanGroup with Helios, a 14-billion-parameter video foundation model that can perform text-to-video (t2v), image-to-video (i2v), and video-to-video (v2v) generation for clips up to a minute long, all running in real-time on a single consumer GPU. Other notable releases include Kiwi-Edit for temporally consistent video editing via text prompts, TencentARC's CubeComposer for seamless 4K 360° video conversion, and the Spectrum method, which offers a 3x to 5x speedup for existing diffusion pipelines without requiring model retraining.
- Lightricks' LTX-2.3 model launched with enhanced prompt adherence and native 1080x1920 portrait image generation.
- PKU-YuanGroup's Helios is a 14B parameter video model performing real-time t2v, i2v, and v2v on a single GPU.
- The Spectrum method provides a 3–5x inference speed boost for diffusion models via Chebyshev polynomial step prediction.
Why It Matters
These tools democratize high-quality video generation and editing, moving advanced AI capabilities from cloud APIs to local, affordable hardware.