# Image grid, RTX 5090 32 GB, 29 Oct 2026. One fresh server per variant and pass.
# Engine: stock ComfyUI v0.43.0 (commit 426f5eea52a7ece1dd599fff4cb6cd1f352ddd24), no custom nodes.
# Server start (only the 5090 visible, no extra GPU flags of ours):
CUDA_VISIBLE_DEVICES=0 python main.py --listen 127.0.0.1 --port 8201 --disable-auto-launch --disable-api-nodes \
  --extra-model-paths-config extra_models.yaml --output-directory out/<variant>-d<pass> --temp-directory temp --user-directory user

# Per variant (6 variants) and per pass (3 passes, fresh server each):
#  1. one warm-up image (neutral prompt, seed 1, see prompts.json), not counted
#  2. for each prompt I1..I4 (text in prompts.json), seed 42, 1024x1024:
#       POST /prompt with the workflow of that variant (workflows.json, <PROMPT> replaced by the prompt text)
#       time per image = POST /prompt until the history entry holds the finished image (text encoder, sampler, VAE, save)
# Variants: zimage_int8, zimage_nvfp4, krea2_int8, krea2_nvfp4, fibo2_int8, fibo2_nvfp4_fremd (unofficial third-party quantization)
# Sampler share (derived afterwards from the server logs, no new run):
python sampler_split.py server-logs