Open source: what people are building this week

This week's picks focus on reducing compute overhead and simplifying multi-step pipelines for image and video generation.

• 1

  OpenMontageGitHub

  OpenMontageInc· 79,028 stars

  OpenMontage is an open-source framework that structures AI coding assistants like Cursor or Copilot into a video production system, providing 12 production pipelines and over 100 tools spanning generation, editing, and audio via integrations with Flux, FFmpeg, and ElevenLabs. Builders looking to automate multi-step video workflows without assembling a custom agent stack from scratch have a ready-made starting point here.

  agentagentic-aiaiclaudecopilotcursor
• 2

  P H3 Image ComfyHF Models

  PrunaAI· 263 likes

  P-H3-Image-Comfy is a compressed, ComfyUI-compatible text-to-image diffusion model from PrunaAI that also supports image-to-image editing, prompt upsampling, and LoRA adapters. Builders looking to run capable image generation pipelines with reduced compute overhead will find it worth evaluating, given PrunaAI's focus on model compression without sacrificing output quality.

  diffusion-single-filecomfyuitext-to-imageimage-to-imageimage-generationimage-editing
• 3

  P H3 ImageHF Models

  PrunaAI· 183 likes

  P-H3-Image is a text-to-image diffusers model from PrunaAI built on the MiniMax H3 architecture, supporting text-to-image generation, image-to-image translation, and prompt upsampling in a single pipeline. Builders working under compute constraints may find it worth evaluating given PrunaAI's focus on model compression and efficiency optimization.

  diffuserssafetensorstext-to-imageimage-to-imageimage-generationimage-editing
• 4

  Qwen Image 2.1 Turbo QuantizedHF Models

  abenzerps· 5,749 likes

  Qwen-Image-2.1-Turbo-Quantized is a 7B text-to-image model running just 8 denoising steps to produce 1024x1024 outputs in roughly 10-20 seconds on GPU, available in GGUF, FP8, INT8, NVFP4, and MLX formats. Builders targeting constrained hardware or local deployment have several quantization tiers to choose from, including 4-bit MLX variants, without needing to convert weights themselves.

  ggufqwenimage-generationcomfyuicomfyui-ggufmlx
• 5

  Pika H3 Flex Turbo LoRAHF Models

  Pika-Labs· 178 likes

  Pika-H3-Flex-Turbo-LoRA converts still images to video using Pika's motion style fused into a single standard LoRA that runs on stock MiniMax-H3 setups via ComfyUI or Diffusers without custom nodes or model patches. Builders who need flexible inference budgets can switch between 2-step draft, 4-step, and 8-step modes from the same file, making it practical to balance speed and quality within a single deployment.

  minimax-h3pikapika-h3-flexturbo4-step8-step