Ouzhang/yale-project1

Physical Unit Quantity Video Generation

This repository contains the local research workspace for a proposal on physical unit quantity understanding in video generation. The central question is whether current video generation models can reliably express quantitative physical relations such as 1 L > 0.5 L, 2 m/s > 0.5 m/s, or 100 W > 20 W, while keeping the rest of the scene fixed.

What This Project Does

The project combines literature review, proposal writing, and pilot experiments for a two-stage research direction:

1. Relative physical quantity learning: train or adapt a video model to preserve ordering and ratio relations under fixed-scene counterfactual prompts, for example 0.5 L < 1.0 L or 5 N < 20 N.
2. Unit-anchored calibration: introduce a user- or data-defined unit anchor, then learn 0.5x, 1x, 2x, 4x scaling around that anchor instead of asking the model to directly know every real-world unit.

The current workspace includes:

• a Chinese research proposal and visual materials;
• paper notes and reference PDFs;
• fixed-seed pair experiments for open video models;
• image/video contact sheets used in the proposal;
• KIVI-based physics evaluation code and pilot outputs;
• scripts for running server jobs and generating proposal figures;
• model download instructions, with model weights intentionally excluded from the repository.

Main Findings So Far

Pilot experiments show that current video models can often generate plausible laboratory scenes, measuring tools, water cups, bulbs, tracks, carts, and other scientific objects. However, they are not reliably grounded in physical unit quantities.

The typical failure modes are:

• changing only the number in the prompt does not consistently change the intended physical quantity;
• fixed-seed generation still changes camera, object layout, labels, or background;
• a pair may show the correct direction, such as more water for 1.0 L than 0.5 L, but not the correct ratio;
• image/video editing or same-reference generation can preserve the scene better, but still does not solve quantity grounding by itself.

These observations motivate quantity-specific numerical tokens, pairwise/groupwise preference optimization, and unit-anchor calibration.

Directory Layout

Path	Contents
doc/	Proposal markdown and report assets. Main file: doc/research_proposal_quantity_unit_video_generation_zh.md.
out/	Canonical output directory for generated videos, contact sheets, figures, paper notes, and KIVI outputs.
scripts/	Local helper scripts, including proposal figure generation and Bernini server submission.
models/	Model download manifest and scripts. Actual model weights are not stored in this repo.
KIVI/	KIVI benchmark code plus added physics-evaluation scripts/configs used for this project.
reference/	Reference papers and literature-review materials.
LaTeX_Templates/	Report/paper template files.

See out/README.md for the detailed output layout.

Tested Model Families

The project contains outputs or scripts for these model families:

• HunyuanVideo 1.5 fixed-seed physical quantity pairs;
• Wan2.2 T2V fixed-seed pairs;
• Wan2.2 I2V same-reference counterfactual videos;
• LongCat-Video fixed-seed pairs and smoke tests;
• LongLive volume smoke test;
• Bernini-R-1.3B fixed-seed quantity pairs;
• older physics pilots with Zeroscope, ModelScope T2V, and Wan2.1-T2V-1.3B;
• static image prompt tests used as additional evidence in the proposal.

The model download instructions are in models/README.md, and the helper script is models/download_models.sh.

Reproduction Notes

Server jobs should be launched through scripts kept in this repository, rather than ad hoc shell commands. The current convention is:

cd /Volumes/PortableSSD/大三下/yale/project1
# create or edit a script under scripts/
# then submit/run that script on the remote server

Large model weights should live outside the repository on the server, for example under /mnt/scratch/l/liyj/kivi_models_cache/. Local model download targets are documented under models/.

Hugging Face Upload

This repository is intended to be uploaded as a Hugging Face dataset-style project repository. The .hfignore file excludes system metadata, caches, nested git directories, and model weight files while keeping proposal documents, scripts, outputs, and metadata.