Wan 2.7 i2v on Replicate is the new face path.
Trump, Musk, Garry Tan, Roy Lee — all rendered cleanly from Wikipedia portraits in 30–40 s. ~3× faster than fal-Kling, on a Replicate account.
CHAPTER 0 · INDEX
AI Video, stress-tested.
Which models cleared the gate. Which didn't. The clips that prove it. A field notebook on the current AI-video stack — Seedance 2.0, Wan 2.7, Kling 3.0, Veo 3.1.
Six clips. Four models. Real receipts.
Click any tile to see the prompt, the model, the gateway and the wall time. Every file ships in the repo.
Have you ever wondered why the same Trump portrait will animate cleanly on one AI video model and get rejected in four seconds on another — even though both are wrappers over the same kind of classifier?
Or why a synthetic likeness generated by Flux 1.1 Pro, with no real photo touched, still gets face-recognized at the input gate of Replicate-Seedance? Or why a 6% morph plus Perlin noise renders all the way through — only to be rejected at the output, two minutes and thirty cents later?
Or what the actual difference is between sending a prompt to Seedance through Replicate versus through fal — given that they’re calling the same underlying ByteDance weights, but each gateway exposes its own moderation surface and you only find out which one fired by reading the error message carefully?
Or how to take a prompt that produces a watery, generic clip and turn it into something that looks deliberately shot — not by writing more, but by re-ordering five slots that every prompting guide independently converged on within twelve weeks of each model’s release?
If any of that sounds interesting, this is for you.
. . .
This is a notebook, not a tutorial. It records what happened when I sent prompts through the current video stack — what worked, what didn’t, and what I haven’t yet been able to verify.
It won’t teach you to use AI video from scratch. It assumes you’ve already opened a Replicate or fal account and burnt your first ten dollars wondering why your prompts keep failing. What it offers instead is a complete, dated record of which model+gateway combinations accepted which inputs — and the prompt structures, reference setups, and bypass attempts that produced usable clips.
You don’t need to be technical to read it. The findings are written for founders, marketers and prompt engineers shipping with these tools today. There’s a builder in here that compiles the request body for any of the four supported models, and a library you can copy verbatim if you just want a working starting point.
Every claim on the site is one of three things: tested, tested-and-failed, or documented-but-not-yet-tested-by-me. The clips are the evidence. The manifest JSON files are checked in alongside.
§ Current findings
Where the stack stands today.
A practical map of which model+gateway combinations accepted which inputs, where classifiers fired, and what prompt structures produced usable clips.
Seedance face screening is photo-specific, not fame-tier.
Sam Altman’s Wikipedia portrait passed fal-Seedance i2v in 235 s. Trump, Musk, Zuckerberg, Garry Tan, Roy Lee on theirs — all flagged. Test the exact image before assuming a model accepts the figure.
Synthetic-likeness bypass is dead at the input.
Flux 1.1 Pro likenesses, 6% morph + Perlin noise — Replicate-Seedance face-recognizes all of them in 3–5 s. The classifier is recognition-based, not pixel-hash based.
01 Models in scope
Four video models, each given a role.
No model wins on every axis. Speed, audio, references, classifiers, scene composition — each one differs. Below is what each accepted when I ran it, with the gateway treated as part of the model.