CLASS 6June 10 · Creative AI Video + Coded Design

Class 6: Creative AI Video and Coded Design. "You can't generate a photo with code, but you can generate an image with a lot of depth."

Taught bySydney Seifert-GramArt director, photographer, designer
Educator, Creative AI Academy

WithTony JonesCo-instructor, Pratt AI Design Certificate

The class in brief

Sydney's third night moved from coded, structure-first visuals into full video craft. It covered keyframes as living stills, the six-part video prompt structure, and an audio-prompt companion. The night's biggest idea was character consistency, demonstrated start to finish on a single woodcut mouse named Baxter. The closing thesis was editing: AI output is a first draft, not a finish.

After this page you can write a video prompt in six deliberate slots. You can also build a character turnaround sheet before animating anything, and protect a flat illustration style from drifting toward realism mid-motion.

Structure over appearance

Why this matters · 6:25 PM

Describe the structure, not the finished look, and the result stays editable.

Coded generation flips the usual instinct. Instead of describing what a finished image looks like, you describe what is underneath it: the content, the structure, and the rules that generate it . The output stays in editable parts you can revise through conversation, unlike a flattened generated image .

The demo was a fully interactive perfume bottle prototype, built by chaining Claude for process logistics, Midjourney for ideation, and ChatGPT for tech specs, material, and lighting. The result was one working browser demo you could rotate and relight. Sydney opened this segment with the line in the headline above.

questions a coded visual has to answer: what's the content, what's the structure, what's it for later, does it need to be interactive.

tools chained into one prototype: Claude for logistics, Midjourney for ideation, ChatGPT for the material and lighting specs.

Keyframes, not reference images

The craft · 6:53 PM

A keyframe is pinned to a timestamp. A reference just guides the mood.

Keyframes are living stills: frozen moments that still carry motion potential, and they can start, continue, or resolve a sequence . A reference guides palette and mood and may never appear in the final output. A keyframe is pinned and locks composition, pose, and camera before the AI animates between two of them .

Prompting for keyframes runs on three principles: create with intention, consider what happens before and after the frame, and verbs matter . Load the still with elements whose physics imply movement, helium balloons, powdery snow, a walking gait, so the video has somewhere to go.

rules for prompting a keyframe: create with intention, think before and after, and let verbs carry the motion.

still can perfect a look before you spend a single generation on the full video. Efficiency is the whole point of anchoring first.

The framework · 7:05 PM

The video prompt structure

Formatwhat kind of visual

A TikTok video, an establishing shot, a timelapse, a Go-Pro clip, expired film. Sets the frame everything else lives in.

Subjectthe hero

A cowboy, a candle, a carousel, a carnival, a puppy, an apple tree. Whoever or whatever the video is about.

Detailswhat's happening

Floating in space, at the lake, in the kitchen. The scene around the subject.

Motionwhat moves, what the camera does

Flickering, dolly in, Dutch tilt, crash zoom, dancing. Say "static camera" explicitly if you want none.

Stylethe vibe, colors, textures, feelings

Dark and moody, high key, film noir, vintage camcorder. Same slot as the image framework's style row.

Parameterswhat else to establish

16:9, 4k, low motion, 24 fps, 1:1, loop. The technical spec the model needs to hit.

The full six-block video prompt structure

Three pitfalls got named right after the framework.

Overloaded prompts confuse the AI with too many ideas at once .
Static keyframes yield almost no movement to animate .
Unclear timing leaves the AI unable to stitch events in order.

Character consistency, live

The craft · 8:09 PM

Models hold WHO better than HOW.

Character consistency was the night's biggest idea, demonstrated live on a woodcut mouse named Baxter, chained through four tools.

Gemini analyzed the illustration style from an early Midjourney design to bank reusable prompt language .
ChatGPT built the character turnaround sheet, front, side, and back, as the master reference .
Five white reference animals were shot the same way up front, so the model could not invent features later . Then references were combined to test what carried over .
Claude was handed the mouse keyframe and asked to write a prompt formatted for Seedance, a video-generation model, using its own formula: subject, action, environment, camera, lighting, style .

models chained into one character pipeline: Midjourney to explore, Gemini to analyze style, ChatGPT to build the turnaround sheet, Claude to write the render prompt.

reference angles shot up front, before a single video frame, so the model had no room left to invent.

The exercise · 8:24 PM

Keeping the same character across two videos

The activity

25 minutes: lock a character, then animate it into two different scenes.

Build your reference set in ChatGPT, Copilot, or Gemini first, then generate the two clips .

Deliver a gif or a link, plus notes on what held and what drifted between the two scenes.

The exercise result, side by side

Scene one: the party.

Scene two: the wet street.

The same glowing hoodie figure carries across both scenes. Watch what the reference set holds onto, the identity, and what drifts anyway, the exact rendering style, which is the exercise's real lesson.

The audio prompt structure

The audio prompt uses four fields in the same order as the visual framework .

The first field is environment.
The second field is Action.
The third field is texture.
The fourth field is mix.

Three habits sit on top of the fields.

Visuals come first.
Label your audio, using tags such as "Music:" or "Sound Effects:".
Put spoken dialogue in quotes.

Editing is the real craft

"AI generations are infinitely better when we edit them, take them apart, and put them back together in ways only we as humans can." This was demonstrated two ways: a full retexture turning a photo into a stylized doll , and a logo composited onto a shirt, then animated .

Two ways to upscale

Traditional upscaling preserves and predicts pixels; generative upscaling reimagines detail that was never there . Sydney's own pick is Topaz, which now lives inside Photoshop too.

Methods and prompts

Five methods to take with you

METHOD 01 · TAUGHT 7:05 part of: structure your ask

Write it as Format, then Subject, Details, Motion, Style and Parameters

The video companion to the image and text frameworks. Motion is the new slot: name what moves and what the camera does, or say "static camera" if you want neither.

Where it came fromSydney presented the video prompt structure as the motion companion to the image and text frameworks, adding motion as its own slot so you name what moves and what the camera does, or explicitly say static camera if you want neither.Use it whenUse this as the starting structure any time you write a prompt for a video generation model.

Same idea elsewhereGive it two keyframes instead of one description (class 2) · Speak the language of the internet (class 3) · Write it as a GOLD block (class 4)

Working prompt

Format: [TikTok video, establishing shot, timelapse]. Subject: [the hero]. Details: [what's happening around it]. Motion: [what moves, what the camera does, or "static camera"]. Style: [vibe, colors, textures]. Parameters: [aspect ratio, fps, loop].

You will know it worked whenthe output moves the way your Motion field described, or holds a static camera if that's what you specified, with nothing else drifting.

METHOD 02 · TAUGHT 8:11 part of: context beats prompts

Build the turnaround sheet before you animate anything

Lock a character's front, side, and back as one master reference sheet before generating a single scene. Everything downstream points back to this one file.

Where it came fromSydney's character-consistency pipeline was demonstrated with her woodcut mouse Baxter. It has you lock a character's front, side, and back view as one master turnaround sheet before generating a single scene. Every later scene points back to that one reference file.Use it whenUse this before animating any character across multiple scenes, so its appearance does not drift from shot to shot.

Same idea elsewhereControl the input and build gradually (class 2) · Show it your voice and not just the output (class 3) · Evolve the reference instead of restarting (class 5)

Working prompt

Here's my character design [attach]. Analyze its style first: line weight, color, texture, era. Then build a full turnaround sheet, front, side, and back view, same style and proportions throughout. This becomes my master reference for every scene after this.

You will know it worked whenthe front, side, and back views hold the same line weight, color, and proportions as each other, not three slightly different characters.

METHOD 03 · TAUGHT 8:17 part of: build tools not chats

Let one model write the prompt for another

Hand your keyframe and your goal to a smart model like Claude or ChatGPT, and let it write the render-model-specific prompt. Different video models want different formulas.

Where it came fromSydney showed that different video render models want their own specific prompt formula, so she hands her keyframe and her goal to a smart model like Claude or ChatGPT and has it write the prompt formatted for the specific render model she is using.Use it whenUse this when you have a keyframe and a goal for a video model but are not sure how that particular model wants its prompt formatted.

Same idea elsewhereStand up a reusable Project or Gem (class 4) · Weight your references on purpose (class 5) · Let the agent build the bones (class 8)

Working prompt

Here's my keyframe [attach] and what I want to happen next: [describe the action]. Write me a prompt formatted for [Seedance/Kling/Runway], following its formula: subject, action, environment, camera, lighting, style. Flag anything that risks breaking character consistency.

You will know it worked whenit names a specific risk to character consistency, not just a generic caution, before handing over the formatted prompt.

METHOD 04 · TAUGHT 8:06 part of: structure your ask

Structure the audio prompt separately

Environment, action, texture, mix. Visuals get designed first; the audio prompt describes what should support them, not lead them.

Where it came fromSydney taught the audio prompt structure as its own separate framework, environment, action, texture, and mix, and stressed that visuals get designed first, with the audio prompt written to support them rather than lead them.Use it whenUse this once your visual prompt is set and you are ready to add sound, music, or dialogue on top of it.

Same idea elsewhereGive it two keyframes instead of one description (class 2) · Speak the language of the internet (class 3) · Write it as a GOLD block (class 4)

Working prompt

Environment: [ambient baseline]. What's making the sound: [footsteps, a door, wind]. Texture: [joyous, muffled, sharp, gritty]. Mix: [foreground or background, loud or subtle]. Music: [label it]. Dialogue: "[quote exactly what's spoken]."

You will know it worked whenthe generated audio matches what you specified in each field, especially the exact words in Dialogue rather than a paraphrase.

METHOD 05 · TAUGHT 8:57 part of: reject the first draft part of: keep the judgment human

Diagnose the edit yourself first

Before asking AI to fix a rough output, name what's actually wrong with it. The output is a first draft; your read on what needs editing is the part only you can supply.

Where it came fromSydney's closing point on editing was that the AI output is only a first draft, and that before asking AI to fix a rough result you should name what is actually wrong with it yourself, since that read is the part only the human can supply.Use it whenUse this before sending a rough AI output back for a fix, so you go in with your own diagnosis instead of a vague request to improve it.

Same categoryJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Here's my raw output: [attach]. I answer first, then you check me. My read on what needs fixing: [color, composite, retexture, trim, pacing]. Tell me if I'm right, then help me execute it as an edit, not a full regeneration.

You will know it worked whenit confirms or corrects your read on what's broken, then executes a targeted edit instead of regenerating the whole piece from scratch.

Where it broke · all night

A flat illustration wants to become photographic the moment it moves

Claude's own Seedance prompt for the woodcut mouse came back with a caveat built into the output .

Style persistence may drift.
Keep camera movement minimal.
The linocut border is fragile.

Video models are trained mostly on photographic and 3D motion, so a flat 2D style drifts toward realism as soon as it starts animating.

The fix is holding the camera static and pre-flighting the prompt in a smart model before you generate.

Sydney named the honest state of the medium plainly: "we're still in the six-finger era of video." That is the same early, glitchy stage AI images went through before the errors washed out.

Try this prompt

Quiz me on the video prompt structure and the character-consistency pipeline. Then give me an illustrated character and make me plan the reference set and the camera constraints before I generate a single frame.

You will know it worked whenit quizzes you on the video prompt structure and the character-consistency pipeline first. Then it makes you plan the reference set and camera constraints before you generate anything.

The shelf

Tools and references

Tools that night

Video models named that night: Krea, Seedance and Seedance 2, Hailuo, Kling, Veo, Wan, and Runway (plus Aleph). Also named: Gemini (Omni) and Midjourney.
Claude Opus 4.8 and ChatGPT for prompt-writing and layered output
Topaz upscaler, inside Krea and now inside Photoshop

Named in the room

Ancestra, Primordial Soup x Google DeepMind: the 3D-previz to Veo to live-action composite pipeline
Baxter, Sydney's own demo character, a woodcut mouse built from a single reference set
The Miro board's cheat sheets: camera placement, focus and lens, cinematic language, the turnaround sheet card

Every frame from this class

59 captures, in order. Click any one to see it full size.

Structure over appearance

Slide titled A Different Way To Create Visuals, next to a 3D rendered bottle stopper with measurement callouts. — This slide introduces coded generation, a different creative approach where instead of describing a finished image, you describe the underlying structure, hierarchy, and rules that generate it.

Slide titled Accuracy and Editability with two paragraphs of body text on a plain background. — This slide explains why coded visuals hold up under scrutiny. Building from underlying logic lets AI tools produce accurate, flexible assets that can be revised through conversation instead of starting over.

Slide titled From Images to Prototypes with an italic paragraph of body text. — This slide sets up the shift from static image generation to interactive prototypes, pushing coded design to account for interactivity, logic, and form, not just visual direction.

Five bottle renders in a row, each labeled with a workflow step and its AI tool, Process Logistics Claude, Bottle Ideation Midjourney, Tech Specs ChatGPT, Material Ideation ChatGPT, Lighting Scenarios ChatGPT. — This slide maps the five-step perfume bottle prototype workflow to the specific AI tool used at each stage: Claude for process logistics, Midjourney for bottle ideation, and ChatGPT for tech specs, material ideation, and lighting scenarios.

Interactive prototype interface showing a glass bottle render with material and lighting option buttons. — This is the interactive perfume bottle prototype itself, showing selectable material options like sterling silver and fluted gold alongside lighting presets like golden hour and moonlight, an example of a coded visual you can inspect and revise.

Slide titled Objectives with a numbered 01 to 04 grid listing four class topics and short descriptions. — This slide lays out the class roadmap in four objectives: where AI video stands right now, keyframe and video prompting language, building a multi-tool workflow for character consistency, and editing and upscaling AI video outputs.

Text quote from Primordial Soup on a plain background about inventing stories and tools with AI. — This quote from Primordial Soup, the studio behind ANCESTRA, describes their approach of merging narrative and experimental AI workflows, citing ANCESTRA, made with Google DeepMind, as a case study referenced in class for where AI video stands right now.

Keyframes, not reference images

Text slide defining keyframes as living stills that carry potential for animation and motion. — This slide defines a keyframe as a living still, a frozen moment that still carries the potential for motion, and explains that a keyframe can start, continue, or resolve a video sequence.

Slide titled Keyframes does not equal Reference Images with two explanatory paragraphs. — This slide draws a key distinction: reference images only guide palette, mood, and style, while keyframes are pinned to a specific timestamp and lock composition, pose, lighting, and camera position for the AI to animate between.

Slide titled Story Anchors with a single paragraph of explanatory text. — This slide introduces story anchors, the idea that keyframes can pin specific moments in a video to a scene, giving more granular control over how events unfold.

Slide titled Prompting for Keyframes with an explanatory paragraph of body text. — This slide frames the mindset shift for keyframe prompting: where image prompting aims for one high-quality still, keyframe prompting is about creating images specifically designed to move.

Prompt tag boxes beside a Midjourney image of dolphins diving down a snowy slope. — This is a Midjourney example showing how short prompt tags, dolphins, diving, deep, freshly-fallen, expand into a full descriptive prompt that produces a scene of dolphins diving down a snowy ski slope.

Prompt tag boxes beside a Midjourney image of a woman in a meadow holding white balloons. — This is a second Midjourney example, showing prompt tags like walking, gently floating, and cinematic expanded into a full prompt that produces a woman walking through a meadow holding white balloons.

Slide titled Prompting for Video with two paragraphs of body text. — This slide explains that prompting for video requires a different mindset than prompting for a still: you're shaping a moving sequence, so it matters not just what appears in frame but when it appears and how it transitions.

Grid of motion verb boxes with several highlighted in bold, including breaks, pauses, freezes, walks, strolls, runs, and sprints. — This is a cinematic language reference grid of motion verbs, breaks, pauses, freezes, walks, strolls, runs, sprints, that class discussed as vocabulary for describing how a subject moves in a video prompt.

Grid of camera terminology boxes covering shot types and camera movements. — This is a reference grid of camera placement and movement terms, covering shot types like close-up and over-the-shoulder alongside movement terms like pan, dolly, crane, and whip pan, vocabulary for directing the camera in a video prompt.

The video prompt structure

Table with six labeled rows, Format, Subject, Details, Motion, Style, Parameters, each filled with example prompt terms. — This is the completed Video Prompt Structure framework, Format, Subject, Details, Motion, Style, and Parameters, filled with concrete examples for each category, the six-part structure used to build a full video prompt.

A cowboy in a white hat, semi-transparent, with prompt tag boxes like cowboy, standing still, and vintage camcorder overlaid on the frame. — This is the Video Prompt Structure framework applied to a real result, showing a cowboy video generated using the tagged elements: subject, details, motion, style, and format.

Slide titled Writing for Motion and Temporal Structure with three lines of body text. — This slide teaches how to write for motion and temporal structure: break events into ordered steps, like the subject enters, then the camera zooms in, because video narratives unfold over time, not just space.

Slide listing three crossed-out pitfall terms, Overloaded Prompts, Static Keyframes, and Unclear Timing, each with an explanatory sentence. — This slide lists three common video prompting pitfalls: overloaded prompts that confuse the AI with too many ideas, static keyframes that yield little motion, and unclear timing that leaves the AI unable to stitch events together.

Exercise brief slide with Goal, Tools slash Info, Deliver, and Remember columns of instructions. — This is the brief for the in-class exercise Make a Video in 5 Words or Less: add motion to a still image using a prompt of five words or fewer, then note how the AI interprets the added language.

Two nearly identical bird close-up images side by side, labeled Still and quote A bird unquote. — This side-by-side compares a still reference image against the video frame produced by the five-word prompt "A bird," showing how even a minimal prompt shifts the still into motion.

Slide titled Prompting for Sound with two paragraphs of body text. — This slide introduces audio prompting: directing what the viewer hears as a scene unfolds, since video models will infer ambience, effects, dialogue, or music on their own if you don't specify them.

Grid of sound-quality word boxes with several highlighted, including piercing, intimate, abrasive, melodic, raspy, gritty, metallic, and crisp. — This is an auditory language reference grid, vocabulary like piercing, intimate, abrasive, melodic, raspy, gritty, metallic, and crisp, for describing the quality of a sound in an audio prompt.

Table with four labeled rows, Environment, Action, Texture, Mix, filled with example audio prompt terms. — This is the Audio Prompt Structure framework, Environment, Action, Texture, and Mix, filled with example terms for what's making sound, how it feels, and where auditory attention should be focused.

Slide reading Visuals Come First and Label Your Audio, each with two lines of explanatory text beside it. — This slide adds a labeling technique to audio prompting: use tags like Music:, Sound Effects:, or Audio: to isolate different audio elements within a larger video prompt. It also reinforces thinking through the visual narrative first, before layering in audio notes.

Character consistency, live

Slide with two lines of italic text about using multiple references to prevent AI hallucination. — This slide explains why AI video workflows gather multiple reference images of a subject: seeing it from different angles and in different environments keeps the AI tool from hallucinating features or characteristics it was never shown.

Slide titled Input in Gemini 3.5 Flash slash Omni with a small woodcut mouse illustration and a request paragraph below it. — This is the first step of the Baxter the mouse character-consistency workflow: asking Gemini to analyze the linocut illustration style of a reference mouse image so its visual language can be reused across future frames.

Slide showing a small woodcut mouse image next to a bulleted style analysis covering art medium, line work, and composition. — This is Gemini's style analysis of the woodcut mouse reference, breaking down its linocut art medium, contour-following line work, and inverted border framing, the output used to keep the character's style consistent in later generations.

Slide titled Input in ChatGPT with a small woodcut mouse illustration and a request paragraph naming the mouse Baxter. — This is the next workflow step, asking ChatGPT to build a character turnaround sheet for the mouse, now named Baxter, using the same reference image so his design stays consistent from multiple angles.

Five photos of white animals, a dog, mouse, sheep, cat, and horse, each on a green background. — This is a set of five reference photos, a borzoi, a mouse, a sheep, a cat, and a horse, gathered to feed multiple animal likenesses into the AI tool without letting it hallucinate features.

Five close-up black and white fur or feather texture photos in a row. — This is a companion set of five close-up black and white fur or feather texture references, used alongside the animal photos to guide the AI's rendering of surface detail.

Two images of a white borzoi dog and a white rat, both with black spotted markings, in a muddy outdoor setting. — This is the combined result of feeding in multiple animal and texture references: a borzoi and a rat rendered with matching black-spotted white fur, showing how reference images shape a new generated look.

Slide titled Remember with a single italicized paragraph of body text. — This slide's takeaway: it's far more efficient to spend time perfecting a character's style, look, and composition through keyframes than to generate full videos from scratch each time.

Slide titled Input in ChatGPT with a small woodcut mouse illustration and a request paragraph about a ramen restaurant scene. — This continues the Baxter workflow, asking ChatGPT to place the mouse in a new scene, a ramen restaurant, at a smaller scale pulling a noodle from a bowl, while keeping the same style and reference frame.

Two side by side black and white woodcut illustrations of a white mouse, one among potted herbs in a greenhouse and one perched on books reading an open page. — This slide shows two additional woodcut style scenes generated for Baxter the mouse character, one tending herbs in a greenhouse and one reading an old book, extending his consistent linocut look into new settings.

A text slide reading that AI can help write prompts and that prompt structure should be tailored to the model being used. — This slide advises using AI itself as a prompt writing assistant even when you already know what you want, since each model takes in information differently and your prompt structure should match the tool you plan to use.

A slide titled Input in Claude Opus 4.8 showing a small woodcut mouse thumbnail beside a request to write a Seedance prompt keeping style and framing consistent. — This is the next workflow step, asking Claude Opus 4.8, an Anthropic AI model, to write a Seedance video prompt using the Baxter mouse keyframe image. It keeps the same style and framing and makes the mouse the main character exploring landscapes.

Where it broke

A two column slide showing the Claude prompt on the left and a written caveat about video model style drift and camera movement on the right. — This slide pairs the Claude Opus 4.8 prompt request with its output. The output is a caveat: video models are trained mostly on photographic and 3D motion, so they tend to drift toward realism. It recommends minimal camera movement and reinforcing the static border to keep style and framing intact.

A text slide explaining that ingredients differ from keyframes because they may not show up exactly as given in the final output. — This slide distinguishes ingredients from keyframes: some AI models let you feed in ingredient images to shape a result, but unlike keyframes those ingredients might not appear exactly as given, acting more as influences on creative choices or subject control.

A slide titled Input in Seedance 2.0 via Runway showing a small woodcut mouse thumbnail and a full paragraph prompt describing the mouse walking. — This is the workflow step moving Baxter into Seedance 2.0 via Runway, using the tagged mouse image plus a detailed prompt describing him taking slow deliberate steps across straw textured ground. It keeps the linocut woodcut style and locked camera frame.

A slide titled Input in Seedance via Krea showing a small keyframe thumbnail and a full prompt describing the mouse eating ramen, with no result image shown yet. — This is the next workflow step, moving Baxter into Seedance via Krea with a new keyframe, the mouse perched on the rim of a steaming ramen bowl lifting noodles to his mouth with chopsticks. It again pins the same woodcut linocut style and locked camera frame.

A slide titled Input in Seedance via Krea with prompt text on the left and a detailed woodcut illustration on the right of a mouse eating noodles from a large steaming bowl. — This shows the Seedance via Krea prompt alongside its output, the woodcut mouse perched on a steaming ramen bowl slurping noodles with chopsticks, keeping the same linocut style and framing established in the earlier keyframe.

A text slide stating that using both start and end keyframes gives tighter control over style, characters, and narrative. — This quote slide states that using both start and end keyframes together can give you tighter control over your style, characters, and narrative than a single keyframe alone.

A screenshot of the Krea tool interface for the Seedance 2.0 Fast model with Start frame and End frame buttons circled in red. — This is the actual Krea interface for the Seedance 2.0 Fast model, with the Start frame and End frame buttons circled to show where you attach the two keyframes discussed on the previous slide.

Keeping one character across two videos

A slide titled Input in Midjourney showing two small keyframe thumbnails, prompt text, a Note the steam annotation, and a large woodcut illustration of a mouse eating from a steaming ramen bowl. — This is the Baxter workflow continued in Midjourney, using two keyframe thumbnails and a prompt describing the mouse sniffing and slurping ramen broth, with a note calling out the steam effect visible curling up from the bowl in the result.

Two side by side woodcut style video frames, a mouse crouched among flowers on the left and a mouse eating noodles from a bowl on the right. — This side by side compares two of Baxter's Midjourney video results in the same linocut style, one of him foraging among garden flowers and one slurping noodles from the ramen bowl. It shows the character holding his look across two different scenes.

A slide with the activity title Create Two Videos with the Same Character and columns for Goal, Tools slash Info, Deliver, and Remember. — This is the brief for the in-class activity Create Two Videos with the Same Character: generate a character, build reference angles with a multimodal tool. Then use start and end keyframes or ingredients to produce two different videos of the same subject.

Front-facing glowing wireframe figure in a hoodie and pants against a black background, with green-glowing sneakers. — This is Tim's Volt character from the same-character video activity: a figure generated entirely as a glowing white-cyan LED light line, with an empty dark hood and green-glow sneakers, generated as the front-facing hero image.

This side-profile view of Volt is one of the reference angles Tim generated to test character consistency, part of the finding that the design's empty hood makes it unusually stable across camera angles.

Back view of the glowing wireframe figure, smooth rounded hood, green-glowing sneakers visible at the feet. — This back view of Volt shows the finding from Tim's exercise: because the hood has no face, the character's front and back look nearly identical, making him easy to keep consistent from any angle.

A text slide stating that AI generations are infinitely better when edited, evolved, and put back together by a person. — This quote slide states that one of the biggest mistakes with AI tools is taking outputs at face value, and that AI generations improve significantly once you edit, evolve, and recombine them in ways only a person can.

A text slide titled Editing Keyframes describing an iterative generate, evaluate, refine, build workflow. — This slide introduces editing keyframes, explaining that the strongest AI workflows are iterative, generate, evaluate, refine, and build on what works, adjusting light and color, compositing mockups, and using effects, without being afraid to step outside AI tools entirely.

Two side by side photos of a person holding a watermelon slice, labeled Original on the left and Fully Retextured on the right with a softer pastel style. — This side by side compares an original photo of hands holding a watermelon slice against a fully retextured version, showing how retexturing can restyle clothing, hands, and lighting while keeping the same pose and composition.

This is the input step for the Nano Banana editing exercise, asking the tool to place a logo in the center of a woman's t shirt, with placement marked by a red line on the example image.

This shows the Midjourney result of the logo placement exercise, a video of the woman wearing the t shirt with the logo now composited on, generated from a prompt describing a minimalist editorial shot against a clear blue sky.

Two side by side video frames of a man in a yellow raincoat on a fishing boat, his boots on the left and red stiletto heels in their place on the right. — This side by side compares the original fisherman video prompt against a mild remix, where a prompt edit swapped his rubber boots for bright crimson stiletto heels while keeping the rest of the fishing boat scene intact. It is an example of conversational editing revising a detail through language.

A text slide titled Upscaling defining traditional versus generative or creative upscaling approaches. — This slide explains upscaling, increasing image or video resolution with AI to add detail rather than simply enlarging pixels, contrasting a traditional approach that preserves the original exactly with a generative or creative approach that can reimagine and transform details.

A text slide listing five takeaways in a row, Thoughtful Prompting, Don't Buy the Hype, Don't Disregard Bad Outputs, Good Things Take Time, and Let Your Prompts Unfold. — The full class takeaways slide lists five instructions: be thoughtful and specific, stay curious without buying the hype, keep apparently bad outputs, allow time for iteration, and let motion prompts unfold.

Class 6: Creative AI Video and Coded Design. "You can't generate a photo with code, but you can generate an image with a lot of depth."

Structure over appearance

Keyframes, not reference images

The video prompt structure

Formatwhat kind of visual

Subjectthe hero

Detailswhat's happening

Motionwhat moves, what the camera does

Stylethe vibe, colors, textures, feelings

Parameterswhat else to establish

Character consistency, live

Keeping the same character across two videos

25 minutes: lock a character, then animate it into two different scenes.

The exercise result, side by side

The audio prompt structure

Editing is the real craft

Two ways to upscale

Five methods to take with you

Write it as Format, then Subject, Details, Motion, Style and Parameters

Build the turnaround sheet before you animate anything

Let one model write the prompt for another

Structure the audio prompt separately

Diagnose the edit yourself first

A flat illustration wants to become photographic the moment it moves

Where this class connects

Tools and references

Tools that night

Named in the room

Every frame from this class