MOSS Image with 4-step LoRA runs on-device on iPhone: ~30s for 1024x1024
woadwarrior · reddit · 2026-11-12
A developer got MOSS Image running fully on-device on an iPhone with a 4-step LoRA: the DiT and text encoder are 4-bit quantization-aware distilled (QAD) and run on the GPU, while the VAE runs in fp16 on the Neural Engine.
- Requires an Apple device with at least 12GB RAM; 16GB devices can also run Krea 2 Turbo and several other QAD-distilled models
- MOSS Image SFT generates a 1024x1024 image in 30s and 512x512 in 10s
- A TestFlight build is available; the author welcomes feedback via Reddit DM or Discord
More from Infra
- Analyst: RASA bill unlikely to crash GPU rental prices but will hit 2027 compute demand — zephyr_z9 · 2026-11-12
- Leaked NVIDIA emails: board rallies employees, all-hands set for Nov 18 after CEO exit — RebeccaBellan · 2026-11-12
- Apex Neural Rendering claims up to 2.4x faster than NVIDIA DLSS, works on all GPU vendors — bdsqlsz · 2026-11-12
- Data center skeptic takes mocked: none hold up, says viral post — HealthcareAIGuy · 2026-11-12
- Developers predict AI model plateau: the real gains will come from inference optimization — bendee983 · 2026-11-12
- Cost-Aware GRPO: weighting RL samples by informativeness and cost to cut training compute — burkov · 2026-11-12