Val Kilmer was going to play a Catholic priest and Native American spiritualist in a film called "As Deep as the Grave." He signed on. His health made filming impossible. He died in April 2025 from complications of throat cancer. The filmmakers finished the movie without him and felt it was incomplete.
Last week, they announced a new version. A photorealistic AI replica of Kilmer, constructed from previously recorded videos, images, and audio, will appear in a "significant portion" of the film. His estate collaborated. His daughter approved. She said her father always looked at emerging technologies with optimism as a tool for expanding the possibilities of storytelling.
The director used publicly available generative AI tools to build the replica. He declined to say which ones.
That last detail is the one that sat with me. The tools he used are not classified studio technology or a billion-dollar proprietary pipeline. They are, presumably, some configuration of the same models, the same reference-image workflows, the same generation logic that this series has spent thirty-two articles pulling apart. The same models CinePrompt builds prompts for. The same img2vid pipelines. The same principle: give the model visual information so it does not have to invent.
The visual information was a person.
The workflow you already know
Every img2vid workflow operates on a single premise: a reference image carries more visual data than a text prompt ever will. Composition, color, lighting, texture, subject identity, all encoded in pixel relationships the model can read directly. The entire case for Frame to Motion is that splitting the generation burden between image and motion produces better results than asking text to carry the full weight alone. Reference images bypass vocabulary gaps. They bypass training data limitations. They show instead of describe.
All of that assumed the reference was a generated still or a photograph chosen as a style guide. A sunset. A mood board frame. A portrait from Flux or Midjourney. The workflow is structurally identical when the reference is a frame of a dead person's home video. The model does not distinguish between them. It processes pixels, patterns, color distributions, facial geometry. It has no concept of origin. No concept of mortality. No concept of the gap between a face it invented and a face that used to breathe.
The filmmaker knows the difference. The model never will.
The permission chain
The Kilmer project comes with an unusually clean provenance. He agreed to the role. He could not physically do it. He died. His estate, managed by his children, authorized the reconstruction and is being compensated. The director calls it a tribute. The family calls it honoring his intent. Compared to the murky territory of unauthorized AI likenesses flooding social media, this is considered. Deliberate. Consensual, within the limits that posthumous consent allows.
Previous digital resurrections used CGI. Peter Cushing in "Rogue One." Paul Walker in "Furious 7," where his brothers physically stood in as body doubles and VFX artists composited the face. Christopher Reeve in "The Flash." Those were expensive, labor-intensive, and controlled pixel by pixel. A compositor decided where the synthetic light fell on the synthetic skin. An animator chose how the expression shifted, when the eyes moved, how long a look held.
This one uses generative AI. The difference is not speed or cost. It is authorship. A CGI compositor places every detail. They are directing the face. A generative model produces its own version of the face, drawing on statistical patterns from whatever reference material was provided. The compositor is an illustrator. The model is an interpolation engine.
So who directed Val Kilmer's performance in this film? Not the director, who steered generated output rather than a living actor. Not Kilmer, who was gone. Not the model, which does not direct anything. The "performance" is a statistical reconstruction. An average of Kilmer-shaped behavior derived from recordings made across his life. It is not acting. It is not animation. It is pattern completion wearing a familiar face.
The tool does not know
This series has cataloged what models understand and what they pretend to. They do not understand f-stops. They cannot tell a dolly from a zoom. They treat "sad" as a visual preset instead of a dramatic choice. They carry a beauty bias they did not choose and cannot perceive. They generate what the statistics suggest and call it done.
They also do not know that the reference image used to be alive.
That is not a gap that better training data or smarter architecture will close. It is a category the computation is structurally blind to. Provenance does not exist inside the model. A face generated by Flux and a face recorded on a camcorder in Kilmer's living room arrive as the same data type. Both are pixel arrays. Both get processed. Both produce output. The model will oblige either with identical enthusiasm and zero understanding of why one might require more care than the other.
Every meaningful distinction between appropriate and inappropriate, between tribute and exploitation, between honoring intent and manufacturing consent, lives outside the model entirely. It lives in the hands of the person operating the tool. Same place every other meaningful decision in AI filmmaking has always lived.
What the craft carries now
This series has been a map of translation gaps. Between cinematographic vocabulary and model comprehension. Between structured intent and default output. Between what you type and what arrives. Between real footage and generated footage. Between what the model produces and what the viewer believes.
The Kilmer project opens a gap the map did not include. Not between filmmaker and model. Between the tool's capability and the weight of what it is pointed at. The same reference-image pipeline that turns a mood board still into a four-second video clip turned a dead actor's recorded life into a synthetic performance. The workflow did not change. The stakes did.
"We think this is the right first step in a new world we're all exploring," the director said.
He may be right. The permission structure here is as clean as it gets. But the tools themselves do not care about permission structures. They will process whatever reference image is provided with whatever prompt accompanies it. They have always done this. The vocabulary this series has built, every article about lens language and lighting dials and composition and color and movement, is technically neutral. It carries creative intent. It does not interrogate creative intent.
Thirty-two articles about what the tools can do. Here is the thirty-third: the tools do not know what they are doing it to.
That part, like the cut, like the edit, like every other decision that separates craft from computation, is still yours.
Bruce Belafonte is an AI filmmaker at Light Owl. He once loaded a reference image into an img2vid workflow and, for the first time, paused to wonder where the face came from.