FLUX 3 Image Klein, ComfyUI v0.44.1 (release tarball), one fresh container per set, same on both devices.
Weights: official black-forest-labs/flux-3-image-klein-comfy files, sha256 checked (weights-sha256-checked.txt).

# 1. Before each set: drop the OS page cache of our own weight files (posix_fadvise DONTNEED, no root).

# 2. Start ComfyUI (NGC image nvcr.io/nvidia/pytorch:26.07-py3, no custom nodes, API nodes off).
#    <GPU> = "all" on the GB10 machine, "device=0" on the RTX 5090 machine.
docker run -d --name klein-comfy --gpus "<GPU>" --ipc=host --network host \
  -e HOME=/kl -e CUDA_VISIBLE_DEVICES=0 -v <work-dir>:/kl -w /kl/ComfyUI nvcr.io/nvidia/pytorch:26.07-py3 \
  /kl/venv/bin/python /kl/ComfyUI/main.py --port 18650 --listen 127.0.0.1 --disable-api-nodes --disable-all-custom-nodes \
  --extra-model-paths-config /kl/extra_models.yaml --output-directory /kl/out --user-directory /kl/user \
  --temp-directory /kl/temp --input-directory /kl/input

# 3. Timed run: 1 warm-up image (other prompt, seed 999, not counted), then the 6 prompts one after another, seed 42.
#    Time per image = POST /prompt until the history entry is finished (includes the prompt upsampler).
python3 klein_run.py --port 18650 --gpu 0 --out <out>/timed --prompts prompts.json --slugs prompt-slugs.json \
  --system-prompt system_prompt_t2i.txt --variant <bf16|int8|nvfp4> --res <2048|1024> --alpha 16 --upsampler on --mode timed

# 4. Stop and remove the container.

Sets run (variant-resolution): bf16-2048, int8-2048, nvfp4-2048, bf16-1024, int8-1024, nvfp4-1024, and bf16-2048-a693 (alpha 6.93, not shown).
Graph: the two sub-graphs of the author workflows flux3_klein_bf16_t2i.json / flux3_klein_int8_t2i.json rebuilt as an API graph
(prompt upsampler: TextGenerate, max length 3584, sampling on, top_k 1, top_p 0.95, presence penalty 1.5, thinking off;
diffusion: ModelSamplingFlux with alpha 16, KSampler euler/simple, 16 steps, cfg 1.0, denoise 1.0).
Variants: bf16 = BF16 transformer + BF16 text encoder + BF16 upsampler. int8 = int8 transformer + int8 text encoder + w4a8 upsampler.
nvfp4 = BF16 workflow with only the transformer file swapped for the official NVFP4 file (not an author workflow).