{
 "title": "MOSS-Image-SFT on rented GPUs: official 50-step sampler vs community 4- and 3-step LoRAs",
 "measured": "2026-10-24",
 "platform": "RunPod Secure Cloud, one node per GPU",
 "weights": "OpenMOSS-Team/MOSS-Image-SFT@f972bd4b4b9e3852133cc5da954d87bc32cd8fb9; LoRAs akhaliq/MOSS-Image-SFT-4-steps@2ded95c37a029360245926fe72e475e30039bbfb (moss_image_sft_4step_v2.0, moss_image_sft_3step_v1.0)",
 "software": "torch 2.11.0+cu128, diffusers 0.41.0, transformers 5.20.2, flash-linear-attention 0.6.0, causal-conv1d 1.7.0",
 "settings": "1024x1024, batch 1, official: 50 steps CFG 3 (negative branch with block 22 skipped); LoRAs: 4 or 3 steps, CFG 1. 6 prompts x 2 seeds after one warm-up image.",
 "prices_usd_per_hour": {
  "rtx4090": 0.89,
  "l40s": 1.09,
  "rtxpro6000": 2.49,
  "h100-sxm": 3.99
 },
 "runs": {
  "rtx4090-te-nf4": "RTX 4090 24 GB, Qwen3.5-9B text encoder in bitsandbytes NF4 (the official CPU-offload mode was OOM-killed in the pod's 31 GB RAM limit)",
  "rtx4090-te-int8": "RTX 4090 24 GB, text encoder in bitsandbytes LLM.int8",
  "l40s": "L40S 48 GB, official bf16, whole model on GPU",
  "rtxpro6000": "RTX PRO 6000 Blackwell Server Edition 96 GB, official bf16",
  "h100-sxm": "H100 SXM 80 GB, official bf16"
 },
 "files": "<run>/results-<variant>.json: per-image timings (encode/denoise/decode), peak memory, power; <run>/ocr-scores.json: OCR text match per expected string (rapidocr + rapidfuzz partial ratio) and layout IoU; prompts.json: prompts and expected strings",
 "harness": "Python harness on diffusers with the official pipeline; method described in the blog post"
}