Directing the Reference, Part 3

The first moving result covers Nozomi waking at her desk after an all-nighter. Three manual drawings establish the wide, overhead close view, and reverse. They also establish what the finished video must not change: the desk, curtain, paper, drawing tablet, phone, wine, her sleeping posture, and the emotional delay before she looks up.

The video model contributes continuous movement, light, texture, and the small physical transitions between those directed states. The output lasts 15.07 seconds at 24 frames per second. It is not proof that the process has been solved. It is proof that the manual layer can remain legible after the production layer arrives.

That is the hybrid I am trying to build. Traditional practice supplies the judgment needed to design, diagnose, and revise. The models expand what a very small production can attempt. Neither becomes meaningful by being placed above the other. They become useful when they share one authored film.

The first session was a long voice conversation with a language model. I did not arrive with a plot. I had a summer setting, a three-to-five-minute festival constraint, and a desire for an emotional turn that would hurt without turning the film into a thriller. The conversation stayed with one question at a time until the subject became personal.

I spoke about the fear of being replaced as a filmmaker while using these tools to make the film. I wanted to redirect that fear, almost like judo, instead of disguising it as confidence or writing a lecture. The protagonist became an overworked animator. The deepest injury became loss of identity. The decisive emotional choice was self-betrayal rather than a machine or a friend becoming the villain.

That distinction shaped everything. The short does not condemn artificial intelligence or place traditional artists on a moral pedestal. It stays close to an artist who has spent years learning anatomy, perspective, acting, cinematography, animation, drawing, and storytelling while the economic meaning of those years begins to move beneath her. The technology can be exciting. The social pressure can be cruel. Both can be true in the same film.

The development conversation became a film bible, beat sheet, and screenplay. Nozomi is a skilled animator and animation director at a small studio. She can correct the weight of a shoulder, the timing of a glance, or the shape of a hand because thousands of invisible hours sit behind those small decisions. Her friend Mia works in machine learning and sees the same transition from the other side.

The story is designed so that enthusiasm and resistance both make sense. Mia is not careless because she is technical. Nozomi is not pure because she draws. The studio, the artist community, and the new production culture each place a different demand on her. She becomes surrounded by people, tools, messages, and output while feeling increasingly nowhere inside them.

I am keeping the later turns out of this journal. The public synopsis is simpler: a struggling animator tries to protect the craft on which she built her identity as her industry changes. Her own community rejects her for experimenting, survival pulls her further into the tools, and she starts to lose the boundary between her hand and the machine.

Short production does not remove the need for continuity. It concentrates it. The apartment had to support sleeping, drawing, a visit from Mia, livestreaming, and later changes in light without shifting its desk, entrance, window, or walking lane. The studio needed enough geometry to stage work from several directions while remaining the same cramped place.

We built character definitions, prop definitions, location notes, floor plans, a reusable apartment blockout, and a color-coded character block library before asking the image model for finished frames. The studio sweep below became a spatial authority: ten readable angles harvested from one environment instead of ten unrelated rooms that merely share a palette.

This is where foundational knowledge becomes practical. A reference can only control perspective if somebody recognizes a broken perspective. A character sheet can only protect anatomy if somebody sees when a shoulder, hand, or weight shift is wrong. Tools can produce options very quickly. They cannot replace the directing decision that says which option belongs to the same film.

The hand-drawn boards do the work that matters most. They choose where the camera stands, what the audience learns, when a character crosses the room, how long a pause lasts, and what must remain outside the frame. They are quick enough to change and specific enough to argue with.

Each board is paired with manual notes. The notes name the character, prop, and location authority, then state the performance, screen direction, background geometry, camera, and forbidden changes. The image model receives a directed problem rather than an invitation to invent the scene.

The more these tools enter the workflow, the more I need actual drawing ability. Moving quickly and accurately from an idea to a visual strengthens every later hybrid stage, and the gain compounds almost exponentially. When the first camera, perspective, pose, or acting idea is vague, everything downstream stalls. This production has pushed me to review fundamentals and spend more time warming up before I draw.

Without that preparation, using a model feels like playing gacha: pull again, hope the random result feels exciting, and accept that the next shot may belong to another film. The hybrid setup becomes useful when the drawings and boards carry authorship, while the models help test finish, lighting, materials, and motion against decisions that already exist.

The productionized boards are not a straight line from rough to approval. Early passes changed the room, added furniture, wrote dialogue into the image, weakened the camera, lost fingers, or followed the finish reference while ignoring the staging reference. Those results are not hidden. They explain what the next instruction had to become.

We tested direct reference transfer, style-only reference, rough-locked composition, local anatomy corrections, and failure-specific retries. A rejected output remained useful when it isolated one mistake. An approved output had to pass more than attractiveness: character identity, room topology, prop placement, perspective, anatomy, acting, and the exact job of the shot.

The image model is very good at making a plausible finished picture. Production requires a stricter question: is this the right picture, connected to the one before and the one after? That question still comes from drawing, staging, editing, and knowing what to look for.

Since the first snapshot, production packages have extended through Shot 32 of the bedroom demo. Shots 12, 25, and 32 were rerun from updated manual authority. Indoor footwear is also a hard continuity check now, so a character sheet cannot silently put shoes on a character who is barefoot in the room.

The repository grew quickly: ninety-four commits in seven days, ninety-four shot records, and many more reference, location, prop, prompt, and review files. A folder tree is precise, but it becomes difficult to direct from when the current board, note, and production state sit several levels apart.

I built a local production website that reads the project as scenes, sequences, and shots. The shot detail panel places the board beside editable manual notes. Character, prop, and location tags remain visible as authorities rather than being buried inside a long prompt. The site writes those decisions back to Markdown so the browser does not become a second, disconnected source of truth. The latest pass also tightened reference search and workspace sync.

This part is engineering in service of directing. It does not decide the shot. It shortens the distance between seeing a problem, changing the instruction, and checking the next result. That distance matters when a short film still contains nearly one hundred planned shots.