Skip to content
Prompt Mortem

The Graveyard Wan 3.0 Prime Case #003

Published 3 Nov 2026

[Case study] Video

Wan 3.0 Prime: I cry if I want to!

It's my party, and I'll cry if I want to Cry if I want to, cry if I want to You would cry too if it happened to you!

Foundation model

Wan

Alibaba · 3.0 Prime

Provider

ArtCraft

Recorded cost

USD 1.88

Generation clips

1 clip

10.031s

Resolution

720×1280

30 fps

Outcome

Mixed

The assignment

Objective

Test whether WAN 3.0 Prime can follow a precise four-shot sequence using separate references for two characters’ facial identities, body proportions and clothing, alongside environment references.

The central test is control over the character introductions: hear Dorian first, clearly see Ethan first, then reveal Dorian before ending on a full-body shot of both men walking.

Expected outcome

A ten-second scene following this exact order:

  • 0–2 seconds: Dorian speaks while only his defocused shoulder and indistinct profile are visible. His face remains unrecognisable.
  • 2–4 seconds: Ethan receives the first sharply focused facial close-up, silently listening.
  • 4–7 seconds: Dorian’s face becomes clearly visible for the first time as he continues speaking.
  • 7–10 seconds: A wide, head-to-toe shot shows both men continuing to walk, with Dorian noticeably taller and broader.

The street references should establish atmosphere and architectural style while the model constructs its own compositions. Character identity, wardrobe and proportions should remain consistent throughout.

// Generations

Attempt 01

The Walk

  • Video generation

Reference media

// Dissection

The autopsy

Overall observations

WAN understands the broad ingredients. There are two recognisable characters in tailored suits, a wet London street, warm lights and glossy reflections. The close-ups are attractive, and the sequence broadly moves from Dorian to Ethan, back to Dorian and finally to both men.

But the opening instruction was more specific than “make Dorian blurry.”

I wanted his shoulder and an indistinct profile, with no recognisable facial features. Instead, the opening puts a substantial portion of his frontal face in the frame. It is soft, but still recognisably Dorian. The planned reveal has already happened before Ethan gets his introduction.

That distinction matters. Hearing someone before seeing them clearly is a deliberate storytelling choice. A slightly blurred portrait does not achieve the same thing.

The environment references also appear to have been treated as compositions to reproduce. The final wide shot closely follows the supplied street view, despite the explicit instruction to use those images for environment and atmosphere only.

Then there is the blocking. These men are supposed to be walking together throughout the conversation. In the wide shot, they are standing facing one another, turn and then walk away. It adds a little piece of staging I never requested.

The result contains much of the requested subject matter. It does not reliably preserve the direction governing how that subject matter should appear.

What worked

  • Ethan and Dorian remain visually distinguishable, with recognisable facial and wardrobe cues from their references.
  • The close-ups retain useful facial detail and broadly photorealistic rendering.
  • The wet pavement, warm streetlights and elegant buildings establish the requested atmosphere.
  • The sequence includes separate character close-ups and a final wider view.
  • Dorian appears taller than Ethan in the wide shot.

What failed

  • Dorian is recognisable in the opening, undermining the explicitly delayed facial reveal.
  • The opening frames his face rather than restricting the composition to a shoulder and indistinct profile.
  • The final wide shot closely reproduces an environment reference’s composition instead of constructing an independent view.
  • The continuous walking conversation becomes a stationary exchange followed by a turn and departure.
  • The requested timing and blocking are only approximately followed.
    "You love me too much" wasn't part of the script.

// Before the next run

Production implications

For this scene, the order of the reveals is part of the storytelling. An attractive close-up cannot compensate for introducing the wrong face first.

The extra stop-and-turn also creates a continuity problem with the intended preceding shot of their shoes walking. I would need to work around the generated staging or replace footage to preserve the planned sequence.

Individual shots might have value as isolated material. As a complete, directed scene, this take is not dependable enough for the intended edit.

This attempt demonstrates a gap between generating the requested people and setting, and following the precise camera and performance instructions that make them a scene.

// Final assessment

The verdict.

I gave WAN a shot list. It gave me its interpretation.

“Do not reveal Dorian’s face” became “show his face, but soften it a bit.” “Atmosphere only” became a composition reference. “Keep walking” became “stop, have a moment, turn around and leave.”

To be fair, the men look good. The suits look good. London looks expensive and thoroughly damp. Ethan and Dorian are also considerably more recognisable than the unfortunate replacement actor in my Grok experiment.

But looking good is only part of the job.

I need to decide whose face the audience sees, when they see it and what the characters are doing. Those instructions were the actual test.

Failed for the intended sequence. Lovely wardrobe department. Director apparently unavailable.

Conditional for my production workflow

Search the archive

Search cases, notes, rants, and models.