陳志信：讓圖片動起來之前，你所需要了解的前置步驟

1. In one sentence

AI animation is not "make a picture move and call it animation"; what actually decides whether motion looks good and has any soul is the compositional design done before you generate anything — and he distilled that method into a handful of general ideas you can apply in any piece of software.

2. The session

Why start with what happens beforehand

He stressed that an idea has to work across all software, because the operating techniques of any one tool may be obsolete tomorrow, while ideas about composition and narrative carry across every animation tool. He cited Meta's figure that the average post holds a viewer's attention for only 1.7 seconds: short video has to seize attention in a very short window, and the key is not the generation tool but whether you designed the frame beforehand as "a storyboard frame that will move", rather than as a beautiful finished image, poster or design comp.

He also noted that AI video generation has advanced to the point where AI actors produce their own performance and lines, not following the instruction exactly, and even improvising actions the creator never asked for — which shifts the creator's role from simply writing prompts to directing: communicating with and guiding AI actors who have ideas of their own.

Technique one: the Dutch angle, creating motion through a tilted frame

His first technique is deliberately breaking the frame's horizon with a tilted composition, a technique with its own name in photography and film, originating in a German Expressionist horror film from the 1920s which he noted won a number of awards and even led later to machines built specifically to shoot at such angles (on verification, the film is *The Cabinet of Dr. Caligari*, a German Expressionist horror released in 1920 — see section seven). He noted Hitchcock and Tim Burton as directors typically fond of using tilted framing to convey unease, and that in action films it is often used to signal that the audience is about to enter a tense sequence.

He demonstrated live that with the same composition, deliberately drawing the frame at a tilt makes the generated motion swing naturally — especially noticeable for flying and running animation. Same prompt, same frame, only the centre of gravity tilted, and the naturalness of the motion differs enormously. He also warned that useful as the tilt is, overuse draws criticism, and the best moment for it is the pivot where a frame turns from calm to tense; conversely he mentioned a well-known film making heavy use of vertical framing to present the spatial contrast between rich and poor (on verification, that film is *Parasite*, directed by Bong Joon-ho and released in 2019 — see section seven).

Different generation models also respond to the same prompt with different force: he warned that the same instruction to go "fast" can produce very different results in a different piece of software, and you have to try things yourself to get a feel for each tool's temperament.

Technique two: reference points, solving AI's inability to tell front from back or left from right

He identified a core blind spot: AI does not really understand relative relations like front, back, left and right — much as on a real set, a director saying "raise your right hand" often gets stuck on "your right or the camera's right?" He demonstrated several failed attempts at the instruction "eat the noodles on the right", with the AI picking up the wrong side, and then switched to a reference point solution: mark the frame with a coloured dot (green, say), change the prompt to "pick up the noodles by the green dot", and the AI executes accurately, with the colour marker removed in post afterwards. The same logic applies to teaching an AI dog actor to identify direction and objects (painting the head of the cat it should touch orange, swapping the thing it should look at for a pineapple) — a process he describes as suddenly becoming a dog trainer, needing repeated trial and error to work out what the AI is actually paying attention to.

Technique three: plate-based compositing

He shared a case shot at a barbecue restaurant: he first visited the shop and filmed the service flow and setting, then replaced the people in frame with cat and dog AI actors prepared in advance, aligning lighting and colour and compositing them straight into the plate — so the whole restaurant's cast becomes cats and dogs without looking wrong. He also mentioned effects can be done the same way: a frame with no smoke in it can be given a local effect and generate a whole cloud of it.

Details to check before generating

He cautioned that before instructing anything to generate animation, beyond the compositional angle you should also check that depth of field reads correctly (the depth relationship in front of and behind a wing, say), because rendering is not cheap, and if the source material's depth map reads wrongly the generated result will be inaccurate — money spent on rendering for nothing.

3. Figures and cases

• Per the Meta figure he cited, the average post holds a viewer's attention for only 1.7 seconds (speaker's own account, source not further specified).
• The animal animation work he shared reached over a hundred million views two months running (speaker's own account; no public link was found to corroborate it, so it is marked as the speaker's account).
• An animation he previously worked on appeared on a large advertising display inside Taipei Main Station (he mentioned a dog-themed advertisement; on verification, the closest match is the glasses-free 4D themed presentation for the "果然是你" app on the Taipei metro — see section seven — though it could not be confirmed with certainty that it is the same piece).

4. Lines worth keeping

• "Learn a prompt today and it's useless tomorrow; but an idea carries across every piece of animation software."
• "Once you've drawn your image, take a look at the depth of field… because if even the depth map reads wrong, those few dollars really are money in the water."
• "Its face was very reluctant, extremely reluctant — but it picked up the right one."

5. Tools and terms mentioned

• Dutch angle: deliberately tilting the camera or the frame to create unease and tension, common in German Expressionist cinema and in Hitchcock's and Tim Burton's work.
• The reference point technique: using colour markers instead of directional words (left, right, front, back) to steer AI-generated action.
• Plate-based compositing: combining a live-action plate with AI-generated characters, replacing the people in frame while keeping the lighting and environmental detail.
• AI video models improvising: he noted that current AI video tools let virtual actors improvise lines and actions from the situation rather than being bound entirely by the instruction given (he was referring to Google's latest generation video model, though which version could not be confirmed — marked unverified).

6. Wider observations

The most striking thing is how he pulls "AI animation" back to the old-school craft of traditional image-making: storyboards, composition, lighting, colour — a technical language that has existed since the film era and applies directly to generative tools, and usually decides whether the result works. The passage on using colour markers instead of directional words to communicate with AI echoes exactly the principles from the earlier sessions on how non-engineers communicate precisely with AI: the smarter AI gets, the more people need to learn to translate an abstract intention into a concrete signal it can understand.

7. Sources