Open source
Week of 14 September 2026
This week's picks center on reducing friction in diffusion pipelines, from faster inference to integrated tooling and metadata control.
- 1Krea2 Turbo Distill 4step LoRAHF Models
lvladikov· 143 likes
A LoRA adapter for Krea 2 Turbo that cuts the minimum denoising steps from 8 to 4, running inference roughly 1.6 times faster while recovering fine texture detail that standard 4-step Turbo loses. Builders targeting latency-sensitive pipelines can drop it onto the existing Krea 2 Turbo checkpoint via diffusers without retraining a base model.
diffusersloratext-to-imagedistillationstep-distillationkrea-2 - 2ODSGitHub
Osmantic· 6,475 stars
Osmantic/ODS packages a self-hosted AI stack for personal hardware, combining LLM inference, a chat UI, voice, agents, RAG, and image generation via ComfyUI into a Docker-based setup that runs on Nvidia or AMD GPUs. Builders already using ComfyUI for image workflows may find value in having it pre-integrated alongside orchestration tools like n8n and open-webui rather than managing those connections manually.
ai-agentsamdcomfyuidockerllama-cppllm - 3remove-ai-watermarksGitHub
wiltodelta· 5,544 stars
`wiltodelta/remove-ai-watermarks` strips both visible and invisible watermarks from images and video, covering formats like SynthID, C2PA, EXIF, IPTC, and XMP through a Python library and CLI. Builders working with Flux, ComfyUI, or other diffusion pipelines may find it useful for scrubbing provenance metadata before distributing or fine-tuning on generated assets.
ai-watermarkc2paclicomfyuicomputer-visioncontent-credentials - 4FLUX.1 SchnellHF Models
black-forest-labs· 5,853 likes
FLUX.1-schnell is a distilled, fast-inference text-to-image model from Black Forest Labs built on a rectified flow transformer architecture, designed to generate high-quality images in as few as 1 to 4 sampling steps. Builders looking to reduce latency or inference cost in production pipelines will find it a practical starting point, as it runs via the standard Hugging Face diffusers FluxPipeline and is compatible with hosted endpoints.
diffuserssafetensorstext-to-imageimage-generationfluxen - 5ComfyUI-CopilotGitHub
ATH-MaaS· 5,514 stars
ComfyUI-Copilot is a custom node for ComfyUI that brings AI-powered assistance directly into your workflow, supporting models like GPT-4 and DeepSeek-V3 to help automate and navigate node-based pipelines. Builders already working in ComfyUI with Flux or similar generation stacks may find it useful for reducing the manual overhead of wiring together complex graphs.
agentaicomfy-uicomfyuicomfyui-nodescopilot