Open source

This week's picks center on making image and video generation more deployable: faster inference, split weights, and hardware-agnostic runtimes.

• 1

  SCAIL 2HF Models

  zai-org· 180 likes

  SCAIL-2 is an image-to-video model that animates a reference character using a driving video without relying on intermediate skeleton or pose representations, and supports character replacement and multi-character scenarios end-to-end. Builders working on character animation pipelines may find it worth evaluating specifically because it removes the skeleton extraction step that typically breaks down under complex or non-human motion.

  diffuserscharacter-animationvideo-generationpose-drivendiffusionimage-to-video
• 2

  FLUX.1 DevHF Models

  black-forest-labs· 13,202 likes

  FLUX.1-dev is a 12-billion-parameter rectified flow transformer from Black Forest Labs that generates images from text prompts, offering a balance between the full FLUX.1-pro capability and faster, more accessible inference. Builders who want a strong open-weights baseline for fine-tuning or custom pipeline development will find it well-supported in the diffusers library with native FluxPipeline integration.

  diffuserssafetensorstext-to-imageimage-generationfluxen
• 3

  JoyAI EchoHF Models

  jdopensource· 139 likes

  JoyAI-Echo is a text-to-video model built on the LTX-Video library that generates multi-shot, audio-video outputs up to five minutes long, maintaining consistent character appearance and voice timbre across shots via a cross-modal audio-visual memory bank. Builders working on narrative or cinematic generation pipelines may find the distilled DMD generator worth evaluating for its combination of long-form coherence and accelerated inference.

  ltx-videoJoyAI-Echotext-to-videovideo-generationaudio-video-generationlong-video
• 4

  openvinoGitHub

  openvinotoolkit· 10,371 stars

  OpenVINO is an open source C++ toolkit from Intel for optimizing and deploying AI inference across hardware targets, with built-in support for diffusion models and LLM inference. Builders working on image or video generation pipelines can use it to run models like Stable Diffusion efficiently on CPU, GPU, or NPU without being locked into a single vendor's runtime.

  aicomputer-visiondeep-learningdeploy-aidiffusion-modelsgenerative-ai
• 5

  LTX 2.3 WorkflowsHF Models

  RuneXX· 711 likes

  A collection of ready-to-run ComfyUI workflows for LTX-2.3 image-to-video generation, built around split model files extracted from Kijai's LTX2.3 comfy repository and supporting both safetensor and GGUF quantized variants of the Gemma 3 12B text encoder. Builders who want to run LTX-2.3 locally without loading a monolithic checkpoint will find the split-file approach easier to manage within memory constraints.

  ltxltx-2comfyuicomfyggufltx-video