About GenVidKit
One maintainer, one owned RTX 3060 12GB test bench, and a standing rule: a number only enters the measured tables after it has been produced here, under pinned conditions, and can be re-run.
Why this site exists
MiniMax H3 shipped as an open-weight video model with bold community numbers attached: minutes-long renders, small-VRAM success stories, dramatic speed-ups. Some of those numbers are real, some depend on settings the posts never mention, and a beginner cannot tell which is which. This site exists to close that gap for one concrete case: running MiniMax H3 locally in ComfyUI on ordinary hardware.
Instead of aggregating claims, the site runs the model on its own hardware and publishes what actually happened — wall times, memory ceilings, temperatures, failures included — next to installation and workflow guidance a first-time user can follow.
Who runs it
The site is written and maintained by TF6cool, whose public profiles are DEV Community and GitHub. The maintainer owns the test hardware, runs every benchmark personally and writes every page. There is no editorial team, no sponsor input into results, and no affiliation with any vendor covered here.
How the numbers are produced
- Owned hardware. Published runs come from the site's own RTX 3060 12GB bench, not from screenshots of other people's terminals.
- Pinned conditions. Workflow files, model files and template revisions are recorded by hash before a run; seeds, canvas, frame counts and step counts are fixed and stated with the result.
- Repeated runs, raw values. Long runs are executed at least twice and reported as raw per-run values with a range — not blended into a single average.
- Failures stay on record. Crashes, freezes and aborted runs are documented rather than deleted; several troubleshooting pages began life as one of them.
- Untested means untested. Cells in the benchmark matrix that have not been measured say so. Support is never inferred from a neighbouring result.
What this site will not do
- Community-reported numbers stay labeled as community reports. They are never converted into site measurements, and the site does not claim to have reproduced or validated them.
- No model weights are distributed here. Download links point to the upstream repositories, and the license map records which terms each file arrives under.
- Machine-generated media is labeled where it appears.
- When a published figure turns out to be wrong, it is corrected on the page rather than quietly removed — reports are welcome.
How a page is reviewed before publication
The tools are checked against recorded inputs, expected results and source references before changes are accepted:
- Trace the numbers. The estimation engine and the three-light hardware checks use the site's own measurements alongside attributed hardware reports. Their reference data records the source and the conditions behind each figure.
- Test expected results. Engine tests compare calculations and hardware verdicts with specified values, including the boundaries where a light changes colour.
- Check the checks. Review includes deliberately changing a value or removing a source to confirm that the relevant assertion fails. The review record documents what was changed and which check caught it.
- Review across models. Backend pull requests require read-only review by Codex, a different model from the backend author. Findings are recorded in the repository and must be addressed before the change is accepted.
The process verifies evidence and presentation; it is not independent peer review. If a reader finds a contradiction or a source changes, the correction channel reopens the affected claim.
How the site is funded
Everything here is free to read. To cover GPU hardware, electricity and hosting, the site may show advertising served by Google AdSense. Advertisers do not see drafts, do not choose topics, and have no influence on any measured result. Details of what advertising means for your data are in the privacy policy.