Open source
Week of 29 June 2026
This week's picks center on making image and video generation more modular, self-hostable, and fine-tunable for production pipelines.
- 1Krea 2 TurboHF Models
krea· 373 likes
Krea-2-Turbo is a 12-billion-parameter diffusion transformer for text-to-image generation, built to run in as few as 8 inference steps with guidance scale set to zero. Builders looking to reduce latency in production pipelines may find the turbo variant worth evaluating, particularly given its native diffusers integration via `Krea2Pipeline` and support for resolutions up to 2048x2048.
diffuserssafetensorstext-to-imageenbase_model:krea/Krea-2-Rawbase_model:finetune:krea/Krea-2-Raw - 2khojGitHub
khoj-ai· 35,388 stars
Khoj is a self-hostable AI assistant that connects to local or remote LLMs to answer questions from personal documents or the web, with support for scheduled automations, custom agents, and image generation. Builders working on multimodal pipelines may find its self-hosted image generation integration and LLM-agnostic architecture worth examining as a reference for wiring together autonomous workflows.
agentaiassistantchatchatgptemacs - 3Krea 2 RawHF Models
krea· 238 likes
Krea-2-Raw is a text-to-image checkpoint from Krea designed as a mid-training base for fine-tuning rather than direct inference, supporting workflows like LoRA training that can then be applied to the production-ready Krea-2-Turbo model. Builders working on domain-specific image generation may find it useful as a starting point, given that Krea themselves have released a collection of in-house LoRAs trained on this checkpoint.
diffuserssafetensorstext-to-imageendiffusers:Krea2Pipeline - 4OpenMontageGitHub
calesthio· 27,832 stars
OpenMontage is an open-source agentic video production system that exposes 12 pipelines and 52 tools, letting you drive video creation workflows directly from an AI coding assistant like Cursor or Copilot. Builders who want to wire together tools like FFmpeg, ElevenLabs, and Flux into reproducible, agent-controlled pipelines without building that orchestration layer from scratch will find it worth a look.
agentagentic-aiaiclaudecopilotcursor - 5LTX 2.3 3DREAL LoRAHF Models
fal· 103 likes
An IC-LoRA built on LTX-2.3 that converts grey 3D viewport renders (Blender blockouts, game-engine captures, low-poly CG) into photorealistic video while preserving the original composition, camera movement, and layout. Builders working with synthetic data pipelines or previsualization workflows can use it to skip manual compositing and get film-quality output directly from rough scene geometry.
loraic-loraltxltx-videoltx-2.3video