Add LoRA extraction, verification, and comparison scripts by ShreejithSG · Pull Request #865 · hao-ai-lab/FastVideo

ShreejithSG · 2025-11-06T07:24:17Z

This PR adds a set of scripts under scripts/lora_extraction for working with LoRA adapters in FastVideo models based on Wan 2.2 TI2V:

Description:

This PR adds a set of scripts under scripts/lora_extraction for working with LoRA adapters in FastVideo models based on Wan 2.2 TI2V:

extract_lora.py – extracts LoRA weights from a fine-tuned model relative to a base model using low-rank decomposition. Supports checkpointing.
verify_lora.py – reapplies extracted LoRA to the base model and compares per-layer transformer weights to verify extraction correctness.
compare_lora_outputs.py – batch comparison of generated outputs between base, fine-tuned, and base+LoRA models. Outputs image/video metrics (MSE, MAE, LPIPS, SSIM).
README.md – explains usage and workflow.

Notes:

Layer-level verification shows near-zero MSE (≈3.6e-08) between original and re-applied weights.

Preliminary perceptual checks show image-level differences (MSE ≈0.19, LPIPS ≈0.62, SSIM ≈0.128). The cause is under investigation.

Scripts are designed to be callable individually rather than a single pipeline, aligning with typical workflows (fused/finetuned checkpoint -> extract -> verify).

gemini-code-assist · 2025-11-06T07:24:37Z

Summary of Changes

Hello @ShreejithSG, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a comprehensive set of utility scripts under "scripts/lora_extraction" to facilitate the management and validation of LoRA adapters within FastVideo models, specifically those built on the Wan 2.2 TI2V architecture. The scripts enable users to extract LoRA weights, verify their numerical accuracy, and perform perceptual comparisons of model outputs, providing essential tools for researchers and developers working with fine-tuned diffusion models.

Highlights

LoRA Extraction Script: Introduces "extract_lora.py" for extracting LoRA weights from fine-tuned models relative to a base model using low-rank decomposition, with checkpointing support for large model runs.
LoRA Verification Script: Adds "verify_lora.py" to reapply extracted LoRA to the base model and compare per-layer transformer weights to numerically verify extraction correctness, showing near-zero MSE.
Output Comparison Script: Includes "compare_lora_outputs.py" for batch comparison of generated outputs (images/videos) between base, fine-tuned, and base+LoRA models, providing metrics like MSE, MAE, LPIPS, and SSIM.
Documentation: A new "README.md" is added to provide a clear overview and usage instructions for the new LoRA utility scripts, explaining the workflow and default configurations.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces a valuable set of scripts for LoRA extraction, verification, and comparison. The implementation is generally sound and serves its purpose well. My review focuses on enhancing the scripts' robustness, configurability, and adherence to Python best practices. The main suggestions include wrapping script logic in main functions guarded by if __name__ == "__main__":, using argparse for configuration instead of hardcoded constants, improving error handling with more specific exception catching, and addressing potential runtime errors. I have also noted a discrepancy between the documentation and script behavior for resuming extraction.

gemini-code-assist · 2025-11-06T07:26:22Z

+# Configuration
+BASE_MODEL = "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+FINETUNED_MODEL = "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers"
+OUTPUT_PATH = "fastwan2.2_transformer_lora.pt"
+CHECKPOINT_PATH = "lora_checkpoint.pt"
+RANK = 16


The script's logic is executed at the module's top level. This is not a good practice as it prevents the script from being imported without running its code, and makes it hard to reuse. All script logic should be encapsulated in a main function, called from an if __name__ == '__main__': block. This would also be a good place to introduce argparse to handle configuration parameters like model names, paths, and rank, instead of using global constants. This would also resolve the discrepancy with the README.md regarding the --resume flag.

ShreejithSG · 2025-11-06T07:26:41Z

@SolitaryThinker, I’ve raised a PR adding the LoRA extraction, verification, and comparison scripts under scripts/lora_extraction.

- Added test trigger in .buildkite/pipeline.yml for scripts/lora_extraction/ - Created fastvideo/tests/lora_extraction/test_lora_extraction.py - Test validates: extract → merge → verify → inference (SSIM/LPIPS) - Configured for L40S GPU with 45-min timeout - Removed deprecated compare_lora_outputs.py

SolitaryThinker

Thanks for the PR, looks good overall

SolitaryThinker · 2025-12-08T06:31:20Z

+    if ssim is None or lpips is None:
+        raise AssertionError("SSIM or LPIPS metrics missing")
+
+    print(f"\nMetrics: SSIM={ssim:.4f}, LPIPS={lpips:.4f}")


add newline char

please check all the other files as well

…#865)

Add LoRA extraction and validation scripts

96def4e

gemini-code-assist Bot reviewed Nov 6, 2025

View reviewed changes

SolitaryThinker mentioned this pull request Nov 24, 2025

[Feature] Development Roadmap 2025 Q4/2026 Q1 #899

Open

18 tasks

Update merge script to match model configs and inference

5bdc7aa

ShreejithSG force-pushed the lora-extraction branch from 0bcaded to 5bdc7aa Compare November 26, 2025 20:49

ShreejithSG added 3 commits November 26, 2025 12:50

outdated inference script

622b2b5

Merge branch 'hao-ai-lab:main' into lora-extraction

256b958

SolitaryThinker added the go label Dec 8, 2025

SolitaryThinker approved these changes Dec 8, 2025

View reviewed changes

ShreejithSG added 4 commits December 8, 2025 08:09

Add missing newline at end of test_lora_extraction.py

91a98b8

Fix CI timeout

f2ee2f4

CI fix

799bcc4

Refactor LoRA scripts for importability

f017acd

SolitaryThinker merged commit 92fb660 into hao-ai-lab:main Dec 9, 2025
1 check failed

ShreejithSG mentioned this pull request Jan 23, 2026

[Feature] Are LoRA weights available for Wan? #1031

Closed

shijiew555 pushed a commit to Gary-ChenJL/FastVideo that referenced this pull request Apr 8, 2026

Add LoRA extraction, verification, and comparison scripts (hao-ai-lab…

a50de84

…#865)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add LoRA extraction, verification, and comparison scripts#865

Add LoRA extraction, verification, and comparison scripts#865
SolitaryThinker merged 9 commits intohao-ai-lab:mainfrom
ShreejithSG:lora-extraction

ShreejithSG commented Nov 6, 2025

Uh oh!

gemini-code-assist Bot commented Nov 6, 2025

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

gemini-code-assist Bot Nov 6, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ShreejithSG commented Nov 6, 2025

Uh oh!

SolitaryThinker left a comment

Uh oh!

SolitaryThinker Dec 8, 2025

Uh oh!

SolitaryThinker Dec 8, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

ShreejithSG commented Nov 6, 2025

Description:

Notes:

Uh oh!

gemini-code-assist Bot commented Nov 6, 2025

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

gemini-code-assist Bot Nov 6, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ShreejithSG commented Nov 6, 2025

Uh oh!

SolitaryThinker left a comment

Choose a reason for hiding this comment

Uh oh!

SolitaryThinker Dec 8, 2025

Choose a reason for hiding this comment

Uh oh!

SolitaryThinker Dec 8, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants