SDXL + ControlNet scales to 3.77x on 4x RTX 5090, but synthetic airport images still fail detector training
RuntimeCoffee · reddit · 2026-10-14
The author tested whether AI-generated satellite imagery could replace collected and annotated data for aircraft detection:
- Pipeline: SDXL Base + ControlNet Canny, generating image-to-image variants from 30 annotated RarePlanes tiles containing 91 aircraft (512x512, FP16, 25 steps, strength 0.25, ControlNet scale 0.6, Canny thresholds 75/150), on PyTorch 2.8 + CUDA 12.8 + Diffusers 0.41.
- GPU scaling: one RTX 5090 hit 1.69 img/s (240 images in 142s, 9.9 GiB peak); 4 GPUs reached 6.36 img/s — a 3.77x speedup with 94.2% scaling efficiency.
- Verdict: throughput scaled well but image quality didn't — aircraft stayed recognizable while airport surfaces developed strange rocky textures, unusable for confident detector training.
- Validation came from unseen locations to test generalization.
More from Multimodal
- LightVAE speeds up video VAE decoding up to 2.24x on H100 with minimal quality loss — Shiqiao Gu · 2026-10-14
- Midjourney v8.2 cinematic portraits: weighted multi-sref prompt recipe shared — michaelrabone · 2026-10-14
- Raising CFG on distilled video models like LTX-2.5 makes output worse, not better — Gen-Talks · 2026-10-14
- Two Gemini instances collaborate on striking black-and-white pixel art — liminal_bardo · 2026-10-14
- Cinematic AI short film made with Seedance 2.5 turns heads — Aiden_Tech_Ai · 2026-10-14
- Maya-2 open-weight Indic voice model claims wins over ElevenLabs and Cartesia in Hindi TTS — DevvMandal · 2026-10-14