gen‑ai.news

Open source

Week of 31 August 2026

This week's picks center on cutting video generation inference steps and adding structured control without retraining from scratch.

  • 1

    FastVideo· 213 likes

    FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree generates synchronized video and audio from a text prompt in just four transformer forward passes, using a data-free DMD2 distillation approach with VSA-H3 attention at 90% sparsity. Builders targeting low-latency inference pipelines may find the data-free distillation path useful, since it removes the dependency on curated video datasets during the distillation stage.

    fastvideodiffuserssafetensorstext-to-videovideoaudio
  • 2

    alibaba-pai· 162 likes

    Acceleration LoRAs for MiniMax-H3 that apply Parallel Decoding Distillation to reduce inference to roughly 8 steps, covering both the FL2VA and Ref2VA variants of the base model. Builders looking to cut generation time on MiniMax-H3 without retraining from scratch have a ready-to-use starting point in BF16 at rank 64.

    videox_funtext-to-videobase_model:MiniMaxAI/MiniMax-H3base_model:finetune:MiniMaxAI/MiniMax-H3
  • 3
    civitaiGitHub

    civitai· 7,240 stars

    Civitai is an open-source platform for hosting and sharing Stable Diffusion models, textual inversions, and related assets, built with TypeScript. Builders working on image generation workflows may find the codebase useful as a reference for designing a social, community-driven model repository with discovery and sharing features.

    aisocial-networkstable-diffusion
  • 4

    alibaba-pai· 166 likes

    A single 6.8 GB ControlNet-Union checkpoint conditions the MiniMax-H3 video generator on five control modalities (Canny, Depth, HED, MLSD, and Pose) as well as video inpainting, with no per-condition checkpoint switching required. Builders who want structured spatial control over video generation without the overhead of classifier-free guidance will find the guidance-distilled design useful, since it runs at `guidance_scale = 1.0` with one forward pass per step.

    videox_funcontrolnetvideo-to-videotext-to-videoimage-text-to-video
  • 5

    Vincentwei1021· 6,893 stars

    A TypeScript toolkit that plugs into Claude Code and Codex as an agent skill, generating cinematic product videos through Remotion using a library of 152 shot recipe cards and 209 motion previews. Builders who want a structured, catalogued starting point for AI-driven video production rather than building shot logic from scratch will find the pre-built template and recipe system worth examining.

    agent-skillsai-agentsai-videoclaude-codeclaude-code-skillsclaude-skills