AetherArt — Production-Grade Diffusion on Consumer GPUs

ℹ️ Free CPU tier — generation works but is slow (~12–15 min at 512×512).

SD 2.1 was trained at 768×768; defaults here are set to 512×512 (minimum for coherent output). Higher resolution = better quality but proportionally longer wait.

See the Sample Outputs tab for instant viewing of 20 results from a local RTX 3070, or clone the repo for full GPU acceleration.
Model
10 60
1 15
512 768
512 768

Ready.

Upload a previously generated PNG to restore its prompt, seed, and settings.

Upload a conditioning image to guide the generation. Canny extracts edges; Depth estimates a depth map. Leave type as none to use standard generation.
Conditioning Type
0.1 2
50 200
100 300

Apply a fine-tuned style adapter. The trigger token and negative prompt are managed automatically when Auto-prepend trigger token is checked. Not supported in Turbo mode.
Style adapter

ukiyo-e: Japanese woodblock print style (rank-8, SD 2.1, 80 images)
0.1 1.5

Adds 'ukyowood' to your prompt and the LoRA's default negative

fp16 (default) — full precision U-Net, ~4.5 GB VRAM peak.
8-bit INT8 — bitsandbytes quantized U-Net, ~2.5 GB VRAM. GPU only.
4-bit NF4 — aggressively quantized, ~1.5 GB VRAM. GPU only.
Quantized pipelines are cached after first load. LoRA is disabled in quantized mode.

8-bit and 4-bit modes are not available on this CPU Space. Run locally with a CUDA GPU to use them.
U-Net precision

fp16: default quality (8bit/4bit require GPU)

Standard — 30-step DPM-Solver++, best quality (~12–15 min at 512×512 on CPU).
LCM Fast and SDXL Turbo modes available locally only — see the Sample Outputs tab for examples generated on RTX 3070.
Speed mode

standard: 30-step DPM++ · fast_lcm: 4-step LCM