gen‑ai.news

Open source

Week of 1 June 2026

This week's picks focus on making diffusion models faster, smaller, or easier to integrate into real pipelines.

  • 1
    PiDHF Models

    nvidia· 226 likes

    NVIDIA's PiD replaces the standard VAE decoder with a 4-step distilled diffusion model that denoises directly in high-resolution pixel space, combining decoding and upsampling into a single forward pass. Builders looking to improve output sharpness from latent diffusion pipelines like FLUX or SD3 can drop PiD in as a decoder replacement without retraining the upstream model.

    pytorchdiffuserssafetensorssuper-resolutiondiffusionpixel-diffusion-decoder
  • 2
    diffusersGitHub

    huggingface· 33,746 stars

    Hugging Face's Diffusers library provides PyTorch implementations of state-of-the-art diffusion models for image, video, and audio generation, covering workflows from text-to-image to image-to-video. Builders working with models like Flux or latent diffusion architectures will find a unified API and active maintenance that reduces the overhead of integrating new generative models into their pipelines.

    deep-learningdiffusionfluximage-generationimage2imageimage2video
  • 3

    prism-ml· 80 likes

    A ternary-weight (1.58-bit) text-to-image diffusion transformer that compresses the 4B FLUX.2 Klein architecture from 7.75 GB down to a 1.21 GB transformer, generating 1024x1024 images in roughly 4.5 seconds on an RTX 3080. Builders working under tight VRAM budgets may find it worth evaluating as a drop-in diffusers-compatible pipeline that fits comfortably on mid-range consumer GPUs without sacrificing full resolution output.

    diffuserssafetensorsternary1.58-bitgemlitehqq
  • 4
    InvokeAIGitHub

    invoke-ai· 27,310 stars

    InvokeAI is a WebUI and toolkit for running Stable Diffusion models locally, supporting workflows like img2img, inpainting, and outpainting across Linux and macOS. Builders looking for a well-maintained open-source foundation to build image generation products on top of will find it worth studying, given its stated role as the base for multiple commercial deployments.

    ai-artartificial-intelligencegenerative-artimage-generationimg2imginpainting
  • 5
    LensHF Models

    microsoft· 149 likes

    Microsoft's Lens is a text-to-image diffusion model built around rethinking training efficiency for foundational generative models, compatible with the diffusers library and distributed as safetensors. Builders exploring how to reduce the computational cost of training large-scale text-to-image systems from scratch may find the accompanying research and architecture worth examining alongside the weights.

    diffuserssafetensorstext-to-imageen