Light-A-Video
[ICCV 2025] Light-A-Video: Training-free Video Relighting via Progressive Light Fusion
Could not verifysource: GitHubhomepagePythonApache-2.0commit 1b444823a725
Python, Apache-2.0 licensed. The project labels itself: video editing and video generation.
Argusic could not get Light-A-Video running. An Argusic agent installed it in 38 minutes and hit 3 errors and fixed every one without help on a clean machine with no GPU, and the whole session was recorded.
Install succeeds in a CPU venv (torch 2.3.0+cpu; diffusers upgraded 0.32.1 to 0.33.0), all 16 scripts compile, IC-Light/AnimateDiff/RealisticVision weights download and load, merge, and strict-load on CPU, configs parse, and sample input decodes; the GPU-only inference run and its video output cannot execute without an NVIDIA GPU (harness cgroup is 8 GiB, no CUDA).
What the agent ended up with on a clean machine, in its own words. How this is measured
Measured by Argusic on a fresh machine every time. Every number links to its evidence. Argusic Score 35 of 100 (the mean of this project's run scores).
At a glance
Subject data from GitHub, linked at the top of this page, refreshed . Test data by Argusic (CC BY 4.0); every number links to a run page with the full log, the recording, and their sha256 hashes.
What was measured
From the run that decided this verdict. Every line is a recorded value, not an estimate.

What went wrong, in the agent's own words
The agent's own words, worst first, never rewritten into advice.
- Runtime requires CUDA: lav_relight.py hardcodes torch.device('cuda'); container has a CPU-only torch and no NVIDIA GPU26 minutes
- README install (pip install -r requirements.txt) failed: diffusers==0.32.1 lacks AutoencoderKLWan/WanTransformer3DModel, breaking lav_wan_relight.py and src/wan_pipe.py; sd_origin keys also mismatched IC-Light's 8-channel conv_in (excluded by the repo's merge, strict load fails on 'missing feats' conv_in.weight)12 minutes
- First CLI launch was OOM-killed (exit 137): control-group memory.max is 8 GiB while loading the full fp32 Undefined pipeline needs >3.2 GiB just for the UNet (vae 319 MB, text 469 MB, peak exceeds cgroup)
- Install time
- 38 minutes
- Cold machine to finish
- 8 minutes
- Errors hit and fixed
- 3 hit, 3 fixed with no human help
- How the result was proved
- venv/pip install exited 0 and torch imported 2.3.0+cpu; compile-all of all 16 py files printed COMPILE_ALL_OK True after diffusers 0.33.0; CAR_SMOKE_OK printed after IC-Light merge+strict load (686 tensors, merged-sha256 d766f682c6a8ba701ade2991) + MotionAdapter (397 M params) + config parse + 16x512x512x3 video decode on CPU; CLI run (python lav_relight.py --config configs/relight/car.yaml) exited 137 (OOM, 8GiB cgroup) and before that printed Loading pipeline components 100% 6/6 then Killed
- Model tokens used
- 212,287
- Exact commit tested
- 1b444823a725
- Container image
- sha256:cdd920bce783
What was not measured
The other half, so you can see exactly how far this evidence goes. A gap is printed as a gap and never filled in.
- Peak GPU memory
- this ran on a CPU machine, so there was no GPU to sample
- Behaviour over a long run
- each attempt uses the smallest documented settings and a fixed time budget, so this measures whether it runs, not how it holds up under load
- Behaviour against real services
- this run was verified against stand-in services only
The end of the recording
The last lines of the terminal session that decided this verdict, exactly as they were recorded.
import os, glob ics = glob.glob(\"/work/repo/models/*\") print(\"models dir:\", ics) import json EOF" in /work/repo succeeded in 608ms: GPU check: False models dir: ['/work/repo/models/model_download_here', '/work/repo/models/iclight_sd15_fc.safetensor… Everything observed is now recorded. Final report: tokens used 212,287 Everything observed is now recorded. Final report:
Replay the whole session, every command from a clean machine to this point.
How it was tested
One agent, one container, no human help. It installs the project the way its own documentation says to, runs it, and fixes what breaks. Everything below is recorded as it happened: the terminal session, the log and the exact commit. The full method.
Strengths and limits
Measured facts, not opinions. How this is written.
What went well
- Installed in 38 minutes.
- Recovered from all 3 errors without a human stepping in, which says the failures are documented well enough to solve.
- Licensed Apache-2.0, as reported by its host.
- Measured 2 times, so the result is not a one-off.
What did not, or is not known
- An agent with no time limit did not get it running.
Also tested, in the same area
Every one of these was installed and run by Argusic on a clean machine. Nothing appears here that was not tested.
Run history
| tested by | attempt | outcome | score | cost USD | when |
|---|---|---|---|---|---|
| Argusic Runner | 2/3 | Did not run | 20.00 | 0.07 | |
| Argusic Runner | 1/3 | Did not run | 50.00 | 0.05 |
Topics (from GitHub)
Embed the badge
Markdown for the project README. It links back here; terms on the terms page.
[](https://argusic.com/subject/light-a-video)Questions
- Does Light-A-Video run?
- Argusic could not verify that Light-A-Video runs. Argusic installed and launched it on a clean machine in 38 minutes, hitting 3 errors on the way, and recorded the session.
- How did Argusic test Light-A-Video?
- On a fresh, disposable machine, with every command recorded and the repository pinned at commit 1b444823a725. 2 attempts are recorded, and the full method is on the methodology page.
- What did it cost to test Light-A-Video?
- The run that produced this verdict cost $0.05: real compute and model cost, not a list price per million tokens. It is recorded on that run's page.
- How long does Light-A-Video take to install?
- 38 minutes in the run that decided this verdict, from a clean machine to a running state. That is one measured attempt, not an average.
- Does Light-A-Video need a GPU?
- Not to start. Argusic reached a running state on a machine with no GPU.
- What goes wrong when installing Light-A-Video?
- 3 things broke in the recorded run, and 3 were fixed without human help. Each one, and the time it cost, is listed on this page.
- Where is the evidence for Light-A-Video?
- All 2 recorded runs are on this page, each linking to its full log and terminal recording, stored with a sha256 fingerprint so it cannot be quietly altered.