ai-animation-pipeline
The video argues that AI video generation fits into a traditional animation pipeline when the artists' designs set the inputs, long structured prompts replace guesswork, and Blender previs supplies motion and camera that words describe slowly. It also states that subtle character acting is still unsolved.
The video presents AI as an addition to a conventional animation pipeline, with storyboards, department-level ownership and a small team of animators who were new to AI. It states that the team's artistic voice, not the tools, steers every decision.
It argues the film's look had to be negotiated with what video models can hold. Pure watercolor tested as sloppy and soulless, midday light as flat, and heavily textured props as cheap in motion. The style was found through repeated mistakes, not decided in advance.
It argues that a video model decides every detail you leave unwritten, so control comes from specifying far more than seems necessary, even for a 4-second shot. Where animating is faster than describing, the team shows the motion instead of writing it.
It claims problems are cheapest to catch before production, so the team tests assets in video early and treats what it learns as binding rules for the project.
Using three problem scenes, it argues that most failures traced back to the references the model was shown rather than to prompt wording. The references were a gray preview framed too close, a character design that had since changed, and a screenshot missing a detail. Debugging therefore means auditing the references.
It claims the hardest problem in this sequence was subtle character acting, not spectacle. The ceiling-handle gag looked simple to animate traditionally but resisted both text prompts and Blender references, which the team calls a 'performance gap'.
The video closes by presenting the sequence as partial progress: 18 scenes made in 3 days, some shots still unsolved, and setbacks used to pick the next tests. Production numbers, total generations, budget and total duration are deferred to a full breakdown of the finished film.
Asset-first approach — The pipeline rule that the quality of the image inputs (character sheets, prop sheets, backgrounds) dictates the quality of the whole film. The people who prepare them are 'asset managers', the art-director equivalent, who make sure everything belongs to one world. Apply: Finish clear, consistent character, prop and location designs before generating video, review them as a set, and attach them as the image inputs for every shot.
Colour-coded storyboard shot triage — Storyboard panels are marked orange when the shot gets a Blender previs stage, and left blank when it goes straight to video generation. Apply: Send shots with specific movement, a specific camera angle, or anything slow to explain in words to Blender, especially when animating is quicker than describing. Send simpler shots straight from the board to text-to-video.
Loose boards — Treating storyboards as a guide, not a fixed plan, because video generation is unpredictable. Apply: Plan story beats and main scenes rather than every shot, expect change, and leave some shots to generative chance.
LLM-assisted prompt drafting with human review — A short human request is expanded by Claude (any chatbot works) into a long, complete prompt for character sheets, locations, props or video. Apply: Give the chatbot a brief description, let it write the full prompt (appearance, clothes, pose, framing, lighting for a character sheet), then read it before generating and fix details that don't match the designs.
Batch-of-four critique and trait merge — Generating four options per run in Soul 2.0 via Cinema Studio on Higgsfield, critiquing each against the character goals, and combining the best traits into one direction. Apply: Judge each result for goals such as funny but cool, stylized rather than too realistic, standing out against the backgrounds, and suited to a flight. Then merge the winners, as with the hoodie and sneakers from the first option plus a skirt for silhouette.
Hand-finishing AI output — Manual correction and painting over generated character art to fix the palette and unify the hybrid 3D/watercolor look. Apply: Raise brightness and saturation in Photoshop for comedic energy. Add hand-drawn strokes around the silhouette, and on the driver's face, cap and clothes, so the characters keep a cel-style 3D look and read as the same film.
Visual formula: watercolor location plus golden-hour lighting — The film's resolved look: generate the location in watercolor style, then add late-afternoon lighting in the prompt during production. It is also carried into production as a lighting lock of low warm sun and long shadows. Apply: Replace midday lighting, which looked flat and dull, with warm evening light and harsh shadows in the prompt to add volume and depth.
Early asset test stage — A stage between pre-production and production where finished assets go straight into video generation to see whether they hold up in the final shot. Apply: Test characters, props and backgrounds early, while edits are still easy, and turn the lessons into project visual rules. The three here were how much texture to keep on the car, putting characters in the references, and the lighting lock.
Texture-load rule for props — Heavy watercolor texture on a prop in motion, over a watercolor background, doubles the information the model must hold. In video the textures looked like a PNG stretched over 3D geometry. Apply: Make the prop clean 3D and keep the painterly texture in the world around it. Briefly seen props such as the highway trucks can stay watercolor.
Image-to-image asset editing — Producing variants of a locked asset by editing the existing image with an edit model instead of generating from scratch. Apply: Attach the main design and edit it into the needed versions, as with the roofless taxi and the backseat interior view.
Prop sheets only for recurring props — Assets are not built for everything, only for props that reappear across scenes and must stay consistent. Apply: Give recurring items their own prop sheet, such as the 3 recurring trucks. Leave one-off items to text-to-video.
Literal-wording check for effects sheets — Effect descriptions can be rendered literally by the image model, as 'mushroom-shaped explosion' produced an actual mushroom with smoke. Apply: Reword the prompt (for example 'nuclear blast') and add hand-drawn textures to the base so the effect matches the playful art direction.
Blender previs as video input — A plain gray Blender animation, with no colour or materials, only geometry, motion and camera, attached to Higgsfield as the video input. It locks camera angle and timing, and squash and stretch on the wheels removes stiffness. Apply: Animate the shot and adjust the camera alongside the animation in Blender, export the gray preview, and attach it with image references for the taxi, trucks, characters, road and light. The video guides movement and camera and the images guide the look.
Five-block video prompt structure — A fixed prompt order, long even for a 4-second shot. It has scene context and style; active references; shot structure; locks and constraints; and animation principles. Apply: Write plain-sentence context and style rules first. List each attached asset like a call sheet, with personality and quality but no visual description, plus the line 'characters, environments and props, 100% match the reference'. Describe each hard-cut shot beat by beat with timing. Add strict locks, always 'no music, SFX only', plus behavior rules such as keeping characters in their seats and landing the taxi only after it clears the trucks. End with the 12 principles of animation.
Front-loading the prompt — The claim that the first 20% of a prompt matters most, because the model reads top to bottom and follows less strictly the further it reads. Apply: Put the most important content first in the prompt.
Shared reusable Claude skill — The prompt structure is written down once for Claude as a 'skill' (system prompt), which Claude applies to every prompt. Apply: Share the same system prompts across the team so everyone follows a similar structure, with each person writing prompts for their own department.
Generate-and-compare review loop — Checking each generated shot against its Blender preview for fidelity. Apply: Watch the generation back against the preview and check timing, camera, and whether the car and characters still match the designs. Simple shots without previs follow the same review and adjustment.
Always include characters in vehicle references — A lesson from the empty-car tests: random people kept appearing inside the car although they were in neither the references nor the prompts. Apply: Add the characters to the references so the model stops inventing occupants.
Zoom out (Rule one) — If the model misreads a reference, especially a textureless gray previs, the camera may be too close for it to see enough of the scene. Apply: Pull the Blender camera back so more of the surroundings are visible and the action is easier to follow, then regenerate.
Screenshot-as-camera-reference — Using screenshots from earlier takes that have the wanted camera angle as references for later attempts. Apply: Capture frames with the right framing and attach them to the next generations, which made the framing consistent in the seat-belt scene.
Reference consistency audit with Photoshop paint-over — Checking that every reference agrees with the character sheet, because a screenshot chosen for camera angle can carry wrong details, such as missing headphones. Apply: Open the screenshot in Photoshop, hand-draw the missing details back in, and use that corrected frame as the reference.
Prompting small props for consistency — Small interactive details need explicit description so they stay consistent, as with the ceiling handles. Apply: Describe where the props are attached, how they come off, and how they look once the character is holding them.
The near-dead-end on watercolor shows how the style was found. After pure-watercolor tests looked 'sloppy, dirty and soulless', the team seriously discussed dropping watercolor and going more 3D. It kept testing both and found its own balance.
A gray previs can be readable to its makers and unreadable to the model. The close Blender camera looked fine to the team because they knew the scene. Without textures or faces the frame was mostly large gray shapes, so the model could not tell how characters and objects were positioned.
Previs detail has a two-sided cost for performance shots. A simple gray character lost expressions and hand movements in the generations, while a more detailed one made the model copy the Blender geometry instead of the character sheet.
Swapping a single reference can destabilize other things. When the final character design replaced the early one, the same prompt and other references produced camera angles that kept changing between takes, so the team spent more time fixing camera than appearance.
A reference chosen for one property can silently override another. The screenshot picked for its camera angle lacked the character's headphones, and every new take lost them even though the character sheet still had them.
Without people in the references, the model invents them. The empty-car tests, run before character sheets existed, kept producing random occupants until the characters were added.
One lighting decision came from an accident: a teammate added 'evening light' to a prompt just to see what happened. Shots got warmer and gained depth, which became the lighting direction.
Descriptive effect wording can be taken literally. 'Mushroom-shaped explosion' produced an actual mushroom with smoke, and 'nuclear blast' mostly fixed it.
Asset effort is allocated by recurrence. Only props that reappear and must stay consistent get sheets (3 trucks), while briefly seen trucks stayed watercolor and were left to text-to-video.
In reference blocks, assets get personality and quality descriptions but no visual description, because the attached images already carry the look.
There is no separate prompt engineer. Each person writes prompts for their own department, and shared system prompts keep the structure similar across the team.
The video says the large action scenes were not the hardest part. Ripping off ceiling handles seemed easy next to a car chase with flying logs and explosions, but it needed tracking of small details and a timed beat of reaction.
The video states the model does not follow all 12 animation principles consistently, but including them got movement closer to what the team wanted.
«our artistic voice is the most important part of our work and that continues to guide how we use these tools.»
— 01:23
«We call it asset-first approach. Essentially, the quality of the image inputs dictates the quality of the overall film.»
— 04:16
«Sometimes, you have to make a lot of mistakes before finding the right direction to go with.»
— 10:50
«This fix became the film's visual formula. Essentially, you generate location in watercolor style, then add in the golden hour lighting and prompt during production.»
— 11:26
«When we already know how we want something to move, it can be quicker to animate it than to explain every single detail with words in a prompt. We are visual artists.»
— 12:39
«Heavy watercolor texture is a lot of information for a model to hold once things start moving. Put a watercolor prop on top of a watercolor background, and that's double the load.»
— 13:37
«The first 20% of the prompt matters a lot because the model reads it from top to bottom, like we humans do. And the further it reads, the less strictly it follows the prompt.»
— 16:51
«The video guides the movement and camera, while the images show how everything should look.»
— 20:20
«The lessons and learnings we find during this stage later become our project's visual rules.»
— 21:47
«Rule one, if the model doesn't understand your reference, you probably can't see enough of it. Zoom out.»
— 23:36
«Her character sheet still had the headphones, but we were giving the model references that didn't quite match.»
— 24:41
«That little pause mattered. That small bit of reaction was crucial to deliver the gag.»
— 26:03
«We called this a performance gap.»
— 26:40
«turns out AI struggles with the subtle character acting instead.»
— 27:04
«For this sequence, we worked on 18 scenes in just 3 days. I know, crazy timeline.»
— 27:26
This is a specific, failure-inclusive account of an AI-plus-Blender animation workflow. Its main value is the concrete rules it draws from tests: prompt structure, reference handling, and the limits of previs for acting. It is the first episode of a series and defers costs, generation counts and total production time to a later full breakdown.

28:05