DFlow2 2B · 4 steps, text to image on your own GPU, nothing leaves this tab | checking WebGPU… | speed shown after the first image

Describe subject, setting, light and medium; "quoted text" is spelled letter by letter. Trained at about 1 MP, so 512² is only a preview. Seeds match dflow: load the LoRA card example (seed 4242, 1824×1024) and compare with the card image.

 
Download image (.png)
Storage, limits and credits

Nothing downloaded yet.

The model downloads once (about 2.7 GB with 8-bit weights, 1.7 GB with 4-bit) into this browser's storage, then works offline. It needs WebGPU (Chrome or Edge 121+, Android 12+, Safari 26) and roughly 3 GB of free GPU memory for 8-bit, 2 GB for 4-bit. On a phone, close other tabs first. If the tab reloads by itself, the GPU ran out of memory: pick a smaller size.

Model: dflow_2b by Data Archetype, 4-step eDMD + LADD LoRA by lvladikov, decoder dinac3_96, text encoder from Google's T5Gemma 2. ONNX conversion: model repo. This project uses dflow_2b by Data Archetype. Non-commercial use only (MG-BY-NC-2.0); Gemma is provided under and subject to the Gemma Terms of Use found at gemma.ai/resources/terms.