Chroma

Chroma is an 8.9B parameter text-to-image model trained by Lodestone Rock on top of the FLUX.1-schnell architecture and released fully under Apache 2.0, sidestepping the noncommercial restrictions attached to FLUX.1-dev derivatives. Training used a 5 million sample dataset curated from a 20 million image pool spanning artistic, photographic, and niche styles, and the model is explicitly positioned as a neutral base for further fine-tuning rather than a polished end-user checkpoint. The current flagship release, Chroma1-HD, is distributed through Hugging Face, where it draws tens of thousands of downloads per month and has already accumulated adapters, finetunes, merges, and quantized variants; the training codebase, a flow matching toolkit with streaming dataloaders and multi-GPU support, is public in the lodestone-rock/flow repository under Apache 2.0 as well. In practice most people run it inside ComfyUI or other FLUX-compatible frontends on a consumer GPU, and its permissive license and broad stylistic range have made it one of the most popular community-trained open base models.