Indie dev open-sources dflow_2b, a 2B text-to-image model trained on a single GPU

quantiler · reddit · 2026-11-21

An indie developer previewed dflow2b on Hugging Face: a 2B-parameter text-to-image model with a small text encoder (t5gemma 2), with both the model and VAE designed and trained by the author on a single GPU. It targets photography-oriented output, accepts natural captions and JSON layouts (MOSS-compatible format), and ships an example gallery with complete ComfyUI workflows for each image. Still in training but already usable.

Original post →

More from Multimodal

Multimodal channel →