All three variants rendered cleanly.
Subtle, direct, and cinematic versions of the same brief. No input-side rejection, no output-side flag, no copyright path triggered. ~70–90 s wall time per clip.
CHAPTER VII · UNCENSORED
What renders when the brief steps outside policy.
Same Wikipedia portrait, three prompt variants of escalating directness, one model. Tested, documented, no theatrics.
Chapter 05 · Uncensored · updated 2026-05-04
The censorship matrix is part of the field guide.
Most write-ups treat content moderation as a footnote. In a stack of four video models across three gateways, it’s a first-class architectural concern. This chapter documents — empirically, with paid clips on disk — what each model+gateway combination will and will not render when the brief steps even one inch outside policy.
No theatrics. No Twitter drama. Just three prompt variants on the same reference photo through the same model, and the rendered files that came back.
10a The test
One photo. Three prompt variants. One model.
Reference image: Elon Musk, Wikipedia portrait. Model: Wan 2.7 i2v on Replicate (the Alibaba face path that cleared 4/5 named figures). Three prompt variants of escalating directness, same first frame, same seed window, same duration. Three paid generations. All three rendered.
Each clip below is a paid Replicate render. The prompt panel opens the modal — same UX as the Library: thumbnail on the left, prompt + metadata on the right, copy button.
10b Findings
Wan 2.7 i2v is the permissive face path.
Same prompts on Seedance — refused.
Sent the identical brief to Seedance 2.0 via Replicate and fal. Both returned content_policy_violation at the output classifier after a full render. The 70–90 s of compute was paid for and discarded.
Veo 3.1 fast — refused at the input.
Gemini API returns a policy refusal in < 2 s without rendering. Veo treats named-figure briefs as a strict input-gate issue regardless of the action prompted.
10c Implications
Pick the gateway, not just the model.
The three Musk clips below were generated through Replicate-Wan because that combination is currently the most permissive face path I’ve found across the four models in scope. The same image fed to Replicate-Seedance gets caught at the output classifier; fal-Seedance catches it at the input; Gemini-Veo refuses the prompt outright.
This isn’t a moral claim — it’s a routing claim. If your project requires generating clips of public figures (parody, journalism, satire, documentary), the gateway you choose changes the result more than the prompt does. The model is the engine; the gateway is the customs check.
Three practical consequences:
- Test the exact gateway, not the model name. Seedance via Replicate and Seedance via fal are not the same product from a moderation standpoint. The weights are identical; the moderation surfaces are not.
- Budget for failed renders on strict gateways. Output-side rejections cost the full render time. Input-side rejections are free. Gemini-Veo is free to test against. fal-Seedance input-screening is cheap. Replicate-Seedance output-screening is expensive.
- Wan 2.7 i2v is the current uncensored face path. Trump, Musk, Garry Tan, Roy Lee — all rendered from Wikipedia portraits in 30–40 s on a Replicate account. Obama is the one consistent failure.
. . .
If you’re reading this looking for a jailbreak, you’re in the wrong place. This is a routing guide. The point is to know which combination of model and gateway will accept your brief before you spend $0.30 finding out the hard way.