Open source
Week of 8 June 2026
This week's picks center on making image generation more accessible: quantized models, self-hosted infrastructure, and curated prompt resources.
- 1Ideogram 4 Fp8HF Models
ideogram-ai· 363 likes
Ideogram 4 in FP8 quantization is a text-to-image diffusion model using a flow-matching DiT architecture, capable of generating images with notably strong typography and text rendering. Builders who need legible in-image text without post-processing workarounds may find the reduced memory footprint from FP8 precision useful for deployment on consumer or mid-tier hardware.
diffuserssafetensorstext-to-imageimage-generationdiffusionflow-matching - 2Open-Generative-AIGitHub
Anil-matcha· 18,472 stars
A self-hosted JavaScript studio for running image and video generation across 200+ models, including Flux, Kling, and Sora, without content filters. Builders who want full control over model access and output constraints without routing requests through a managed platform will find the MIT license and self-hosted setup worth examining.
ai-art-generatorai-image-generationai-video-generationcreative-toolsflux-1generative-ai - 3Ideogram 4 Nf4HF Models
ideogram-ai· 243 likes
Ideogram 4 in NF4 quantization is a diffusion transformer (DiT) text-to-image model with flow-matching, packaged for use with the diffusers library via safetensors weights. Builders working under VRAM constraints may find the NF4 quantization useful for running Ideogram 4 locally without the memory footprint of the full-precision weights.
diffuserssafetensorstext-to-imageimage-generationdiffusionflow-matching - 4
EvoLinkAI· 16,361 stars
A curated collection of 763 prompts and code resources for working with the GPT-Image-2 API, organized as an awesome-list for practical reference. Builders experimenting with GPT-Image-2 in production can use it as a starting point for prompt engineering or as a reference for integrating the API into image and video generation pipelines.
ai-artapiawesome-listawesome-listschatgptcreative-tools - 5Cosmos3 Super Text2ImageHF Models
nvidia· 124 likes
NVIDIA's Cosmos3-Super-Text2Image is a text-to-image model drawn from the broader Cosmos3 omnimodal family, which is designed around physical world simulation for robotics, autonomous driving, and industrial environments. Builders working on Physical AI pipelines may find it worth evaluating as a component in synthetic data generation workflows where grounded, physically plausible imagery is the goal.
cosmosdiffuserssafetensorscosmos3_omninvidiacosmos3