Datasets:

rockeycoss
/

tmp

Step through the same prompt's images across the runs in test/simpleeval_sampled_inference/ (one column per run, e.g. ep10·cfg9, ep20·cfg7), one prompt at a time.

Run it

cd /scratch3/len091/nanogen_repos/nanogen_inf
python test/simpleeval_sampled_inference/viewer/gen_manifest.py --serve

It prints the URL, e.g.:

http://virga-login:8000/test/simpleeval_sampled_inference/viewer/index.html

Remote machine? Forward the port first, then open the localhost URL:

ssh -L 8000:localhost:8000 virga-login

Use --port N for a different port.

Controls

← / →, Prev / Next, or the slider — step through prompts
Random (or press r) — jump to a random prompt
prompt # box — jump to a specific index
size slider — scale the images; click an image to open it full-res

Scores

Each run's vqascore_qwen3-vl-8b.npy (one score per image, 0–1) is shown under the image, colored red→green by value, with the run's mean score in the column header. Runs without that .npy show —.

Notes

Under --serve, the manifest is rebuilt live on each page load, so adding more run folders (other epochs/cfgs) into simpleeval_sampled_inference/ just shows up as new columns on refresh. A not generated placeholder appears for any run missing a given prompt index.
Without --serve you can instead serve the repo root with python -m http.server and open the same path; it then reads the static manifest.json snapshot that gen_manifest.py also writes (re-run the script to refresh it).