Magnific shows Wan 3.0 animating a character sheet

• A single character sheet anchored magnific's 10-second rusted-gardener test, and magnific's published prompt names face, build, and materials as the traits to retain.
• Shot direction can be written as blocking with timing and sound: magnific's restaurant-kitchen demo calls for one moving take, while magnific's two-shot car scene assigns dialogue to two timed shots.
• magnific's short-film breakdown treats continuity as pre-production, with magnific's short-film thread locking characters before generation and magnific's location reference reusing an empty kitchen across three timelines.
• Referenced illustration style still shifts under motion, according to GlennHasABeard's first test and GlennHasABeard's reply about a pull toward 3D rendering.

The gardener recipe specifies a crane down to sprout level, servo whirr, water trickle, and a warm string cue. The kitchen script includes a plate catch, a walk-in door releasing cold fog, and one Andalusian Spanish line at the 14-second mark. Alibaba's API reference frames Wan3.0 as reference-based video generation, while its Model Studio release page calls the input system Omni Reference.

The reference supplies identity traits, while the text supplies scene direction. In magnific's published prompt, the robot's face, build, and materials are the only requested carryovers from [imag1].

The rest of the prompt divides the take into four production controls:

• Action: the robot carries an oversized watering can, waters one sprout, then sits beside it.
• Look: stylised 3D animation, rounded shapes, subsurface scattering, teal and orange.
• Camera: a high starting angle that cranes down to the sprout.
• Sound: servo movement, trickling water, and a string cue.

The restaurant prompt specifies a 30-second unbroken handheld move from a ticket printer, down the pass, past the dish pit and walk-in, to a silent head chef. Its only dialogue is an off-camera Spanish call at roughly 14 seconds; extractor fans, frying oil, metal, printer, and plates form the audio bed.

The car scene applies the same structure in 15 seconds: a back-seat two-shot runs from 0 to 8 seconds, then reverses to the man from 8 to 15. The dialogue, delivery, silence, light, condensation, and ambient engine and gull sounds are each assigned in the prompt.

For a mother-and-son short spanning 15 years, magnific said it locked three characters before generating a shot. David and Susana appeared at three ages through character sheets, according to magnific's character-sheet post.

The team also generated its kitchen without people, then fed that image back into every scene set in the room across three timelines. A separate magnific demo claimed one face persisted across 22 years, three ages, and 13 scenes in magnific's opening-night post.

GlennHasABeard said the first test with his art style and references had already earned a place in his workflow. He also wrote that Wan 3.0 shifted the result toward a more 3D model, though GlennHasABeard's comparison put it closer to Seedance 2.5 than Flux 3 or H3 for that task.

He said a more precise prompt could lock the source style further in another reply. In a separate 30-second corgi test, he identified dogs morphing together near the water and named a starting keyframe as a possible control in his corgi test.

magnific advertises 30-second generations with audio, up to 20 references, 1080p output, five ratios or adaptive framing. Alibaba's launch announcement describes the model as a public beta built around multimodal references and videos up to 30 seconds.

Availability is already split across creative surfaces. magnific said Wan 3.0 is live in its availability post; Pika's launch post offers a 20-reference, 30-second generation through Pika API Club, while Runway's announcement says Runway supports multiple image, video, and audio reference inputs.