# Image comparison on one RTX 5090 32 GB, 5 Nov 2026 (night run). Engine: ComfyUI 0.44.0 (master 7ad7adb3), seed 42.
# 1) Start the server (paths shortened):
python main.py --listen 127.0.0.1 --port 8291 --disable-auto-launch --output-directory <out> --temp-directory <tmp> --user-directory <user>
# 2) For each setup a job list (6 prompts after one warm-up image, seed 1, not counted) was sent to the ComfyUI API as a full graph
#    (the graph of each image is stored in the jobs-*.json files of this folder, key "graph"). Wall time = submit to finished file.
python run_h3.py --port 8291 --out <run-folder> --jobs <jobs-file>.json
# Setups (graph source):
#  P-H3-Image w4a8 and int8: workflow p-h3-image_text_to_image.json of PrunaAI/P-H3-Image-Comfy (prompt upsampler bf16 in the graph,
#    8 steps, euler, ManualSigmas, MiniMaxH3SigmaShift 1/1, 8 MP = 2912x2912)
#  Qwen Image 2.1 "A": Comfy template image_qwen_image_2_1_t2i (int8_convrot, 1024x1024, 25 steps, cfg 1, euler/simple), short prompt as is
#  Qwen Image 2.1 "B": same graph, prompt = the upsampled text that P-H3 w4a8 produced for the same request
#  MOSS-Image-SFT "C": Comfy template image_moss_image_t2i with its own prompt upsampler (system prompt in moss-upsampler-system-prompt.txt),
#    1024x1024, 25 steps, cfg 2.5, er_sde/simple
#  MOSS-Image-SFT "A" / "B": same graph without upsampler; A = short prompt, B = P-H3 upsampled text