What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The Awesome Seedance project connects video prompts to their original posts and records how selected prompts performed when tried on other models. Its headline figures describe different things: Carl Lee’s September 2026 article reports 463 source-traced cases and 264 cross-model retest runs, and says the 463 cases were not all rerun. These are project-reported totals, not an independent audit.
What the project tracks
Awesome Seedance is presented as a way to connect a prompt with both its provenance and practical reuse. A source link helps answer where a prompt came from; a template helps explain how to adapt its structure. A retest record adds a third kind of evidence: how a particular attempt performed under a specified model and platform.
As an Amazon Associate I earn from qualifying purchases.
Carl Lee’s September 23, 2026 article describes a collection of 463 Seedance 2.5 and 2.0 cases traced to original posts, alongside 25 reusable templates and 60 installable AI-video Skills. Those figures describe the article’s snapshot of the project, not an independently audited or industry-wide dataset. Read the article on DEV Community.
Why 463 cases does not mean 463 retests
The 463 figure counts source-traced cases in Lee’s article. The 264 figure counts retest runs. In the project documentation’s statistics dated September 19, 2026, 254 cases had been rerun in 264 runs. A case and a run are not interchangeable: a case can have more than one attempt, and many cases had not been rerun.
#1 Best Overall
The documentation reports 193 reproduced verdicts, 68 degraded verdicts, and 3 failed verdicts across those 264 runs. These are the project’s own classifications and results; they do not establish how prompts will perform for other people or under different generation conditions. See the project documentation.
Retest results depend on the conditions
The project documentation separates its reported results by model tier and platform. Its August 2026 batch used Flova’s MiniMax H3 768p; its September 2026 batch used fal.ai’s minimax/h3-max/text-to-video endpoint. The documentation cautions against combining them as though they were the same model condition.
| Documented model and condition | Runs | Reported reproduction |
|---|---|---|
| MiniMax H3 Max 768p, fal.ai endpoint; September 2026 batch | 253 | 73% |
| MiniMax H3 768p, Flova; August 2026 batch | 11 | 82% |
These figures come from the project documentation’s September 19, 2026 statistics snapshot. They are not a controlled comparison of the two tiers: the platform and batch differ, and the smaller H3 sample has only 11 runs. The rates should be read as project-reported outcomes under those recorded conditions, not as general success probabilities.
How a creator can use the collection
- Open a case and inspect its source. Follow the creator’s original post to see the context in which the prompt appeared, rather than treating a prompt as source-free text.
- Choose a template and adapt it. Templates are reusable structures, not guarantees that a new prompt will recreate the source video.
- Generate in Seedance. The article describes using the adapted prompt in Seedance after reviewing the source and structure.
- Read retest notes before changing the prompt. Check the recorded model, platform, and verdict so you can judge whether the result is relevant to your own setup.
One example in Lee’s article is a UGC creator-review template that separately locks the person and the product, then links spoken lines to visible actions. That structure is useful for making the intended relationship between narration and shots explicit; it is distinct from evidence that a specific generation will match a reference.
Templates and Skills serve different workflows
A template is material a creator can copy and edit directly. An installable Skill is intended to package a prompt workflow for an AI coding agent. Lee’s article reports 60 AI-video Skills, and describes support for Claude Code, Codex, and other coding agents. These are different ways to apply the project’s structures, not different levels of proof that a prompt is reproducible.
Provenance is useful, but it is not independent validation
The indexed project documentation says cases are checked against original posts and describes provenance fields including author, original-post link, and publication date. It also says prompts reverse-engineered from output alone, without a source or submission, are rejected under its collection standards. That explains the project’s stated collection method; it does not independently verify every source or retest.
A source post, a project retest, and an independent evaluation answer different questions. A source post supports where a prompt was found; a retest documents what the project observed under its stated conditions. Neither alone proves that another user can reproduce the same result.
Counts vary by snapshot
The project documentation dates its statistics to September 19, 2026. A separate AtomGit mirror reports a September 30, 2026 sync with 593 cases. That later number may use a different snapshot or inclusion basis, so it should not be combined with the 463 in Lee’s article as though the figures came from one count taken at one time. View the AtomGit mirror.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
