Skip to content

HOWTO_REGENERATE_CATALOG.md

Latest commit

 

History

History
172 lines (125 loc) · 5.95 KB

HOWTO_REGENERATE_CATALOG.md

File metadata and controls

172 lines (125 loc) · 5.95 KB
Raw

How to regenerate the artwork catalog from an empty checkout

Start from a fresh git clone with no images on disk and end with a running installation showing all 169 artworks. This is the recovery path after the image directories have been deleted — they are excluded from git by design, so a clone never contains them.

Total wall-clock: ~3 hours on an Apple M-series, most of it unattended.

Prerequisites

  • Python 3.10–3.12 (MediaPipe has no 3.13 wheels yet)
  • A GPU: Apple M-series (MPS) or NVIDIA (CUDA). CPU works but takes ~10× longer.
  • ~15 GB free disk (362 MB sources + 11 GB pictures + 4 GB model cache)
  • Network access to collectionapi.metmuseum.org and huggingface.co
  • Optional: Ollama with the llava model, for prompt generation
cd uncanny_maker
pip install -r requirements.txt

Step 1: Restore the source artworks

python restore_catalog.py

Downloads the exact 169 Met Museum images listed in CATALOG_MANIFEST.md, each under its original filename.

Expected output:

Manifest: 169 artworks — 0 present, 169 missing
Target:   .../uncanny_maker/catalog

  [  1/169] 436284  A_Donor_Presented_by_a_Saint_436284
  ...
Restored 169/169 file(s).

Use this, not download_human_figures.py. That script discovers artworks by keyword search and selects them by index into the result set. Met search results shift over time, so re-running it yields a different catalog: different artworks, different filename stems, and therefore different seeds and different pictures. It is the right tool for building a new catalog, the wrong one for restoring this one.

Verify before continuing:

ls catalog/*.jpg | wc -l     # expect 169

If some artworks fail (the Met occasionally withdraws an image or revokes public-domain status), the script lists them and continues. A smaller catalog is fine — CatalogManager loads whatever is present.

Step 2: Generate the picture sequences

Optional but recommended, in a second terminal:

ollama serve
ollama pull llava

Then:

python iterate_degrade.py

This is the long step: 169 artworks × 10 pictures, about 1 min per artwork on M-series, ~30 s on NVIDIA. Stable Diffusion v1.5 (~4 GB) downloads on first run and is cached in ~/.cache/huggingface/.

Expected output:

Catalog: 169 image(s) — 0 already complete, 169 to process
Settings: 10 pictures · direct 1–5 0.1→0.3 · chain 6–10 0.22→0.42 · guidance=6.0 · steps=25

Fetching LLaVA prompts in parallel (8 workers)…
  [A_Woman_Reading_435991.jpg] hyperrealistic woman reading, glassy eyes, waxy skin…

Loading Stable Diffusion pipeline…
[1/169] A_Donor_Presented_by_a_Saint_436284.jpg
   10/10  5.8s/iter  ETA 0.0 min
  Done in 1.0 min → catalog_iterations_10/A_Donor_Presented_by_a_Saint_436284

Fully resumable. Kill it and re-run any time; artworks whose 0010.png exists are skipped, and chained pictures reload their predecessor from disk.

If torch.compile() errors on your PyTorch build:

python iterate_degrade.py --skip-compile

Step 3: Verify

ls -d catalog_iterations_10/*/ | wc -l              # expect 169
find catalog_iterations_10 -name '0010.png' | wc -l # expect 169 — all complete
du -sh catalog_iterations_10                        # expect ~11 GB

Every artwork directory should hold 11 files, 0000.png (the untouched source) through 0010.png.

Step 4: Run the installation

cd ../ars_aut_abeat
./start.sh

Open http://localhost:8000. Raise both hands for 1.5 s to trigger a run, or switch to SHOW mode and press Space.

How exact is the reproduction?

Pictures are deterministic given the source image and its filename: iterate_degrade.py seeds each artwork with zlib.crc32(stem) and each chained step with seed + i. Same input file, same stem, same DIRECT_*/CHAIN_*/GUIDANCE/STEPS → same pictures, bit for bit.

Two things break exactness, neither of which affects how the piece works:

  1. The LLaVA prompt. Ollama's output is not deterministic, so a re-run produces a differently-worded anchoring prompt and therefore visually different (not worse) pictures. Running without Ollama is fully deterministic — every artwork falls back to the fixed prompt "classical painting, human figure, museum artwork, detailed".
  2. Model or library versions. A different diffusers, torch, or stable-diffusion-v1-5 revision changes the numerics.

To reproduce the pictures exactly as first generated, run without Ollama. To reproduce the installation as exhibited, any run is fine — the measurement concept does not depend on specific pixels.

Troubleshooting

No images found in .../catalog from iterate_degrade.py Step 1 did not run or wrote elsewhere. Check ls catalog/*.jpg | wc -l.

The app starts but stays on IDLE and never triggers CatalogManager found zero artworks, so pick_next() returns None and the state machine cannot leave IDLE. It scans uncanny_maker/catalog/*.jpg and requires a matching catalog_iterations_10/{stem}/0010.png. A source JPEG with no completed sequence is silently skipped. Re-run Step 2.

Pictures 404 in the browser, artwork title shows /frames is mounted at import time only if catalog_iterations_10/ exists. If you generated pictures while the server was running, restart it.

RuntimeError: Cannot reach Ollama Harmless. Prompt generation falls back to the fixed prompt. Start ollama serve first if you want LLaVA-guided prompts.

Out of memory on MPS/CUDA Lower STEPS in iterate_degrade.py (25 → 15 roughly halves runtime and memory pressure with minor quality loss), or run with --skip-compile.

Related