gauravgandhi2411/AetherArt

AetherArt


This Space runs on CPU (~8–15 min per image). It is an architecture demo — use it to explore the UI and inspect the generated samples. For real-time generation, clone the repo and run locally with a CUDA GPU.

AetherArt is a personal research project exploring diffusion model inference on consumer hardware. It wires together SD 2.1, a custom Ukiyo-e LoRA trained on an RTX 3070, ControlNet (Canny + Depth), and three speed tiers (Standard / LCM 4-step / SDXL Turbo 1-step). Every output embeds its full generation parameters as PNG metadata — drag any prior output back into the UI to reproduce it exactly.

[Full documentation, benchmark results, and source code →](https://github.com/gaurav-gandhi-2411/AetherArt) [360-run scheduler benchmark findings →](https://github.com/gaurav-gandhi-2411/AetherArt/blob/main/reports/findings.md)

Features

• —3 speed tiers: Standard 30-step · LCM 4-step (5.3× faster) · SDXL Turbo 1-step
• —3 memory modes: fp16 (3.1 GB) · 8-bit INT8 (2.2 GB) · 4-bit NF4 (2.8 GB peak)
• —Ukiyo-e LoRA style adapter (rank-8, 80 WikiArt images, checkpoint-1000)
• —ControlNet conditioning: Canny edge + Depth map (combinable with LoRA)
• —DPM-Solver++ scheduler (best CLIP/latency in 360-run PartiPrompts benchmark)
• —Recreate from PNG — every output embeds full generation parameters; drag any prior output into the UI to restore exact settings

Speed tiers (RTX 3070, measured locally)

Running locally

git clone https://github.com/gaurav-gandhi-2411/AetherArt.git
cd AetherArt
pip install -r requirements.txt
python app.py

See CONTRIBUTING.md for full setup including CUDA torch installation.