Chris Wesselman
© Chris Wesselman

Generative Filmmaking

Directed image and video development for cinematic production

I develop generative imagery and video as part of a larger filmmaking process, using the same production standards I bring to feature animation and visual effects. The work is grounded in composition, cinematic lighting, lens behavior, spatial continuity, and shot evaluation. I take shots from visual intent through keyframe development, motion tests, comparison, selection, and finish, treating generated material as source material rather than final output.

Old Growth: 1815





The studies below are drawn from my current independent project, a frontier folk-horror film set in Cascadia in 1815. I am developing it as a connected body of shots with defined rules for geography, period detail, character behavior, scale, weather, palette, lens language, camera movement, atmosphere, and sound.


The work combines directed keyframe development, generative motion, model comparison, painting, 3D, and compositing. I select each system and control method around the dramatic and visual needs of the shot.

1815, Oregon Country. British and American claims at the mouth of the Columbia remain unsettled, and the North West Company is sending trapping brigades south from Fort George. One brigade enters an abundant coastal valley and mistakes what it cannot read for vacancy. The country is neither empty nor unclaimed. The valley answers on its own terms.

geography / route logic

drowned stone threshold

SHOT STUDIES

An ongoing series. Each study documents one shot from dramatic objective through failure, comparison, and selection.

Shot Study 01 · THE WOODSMAN

Directed Shot Development


The objective was to move slowly toward a frightened woodsman while a much larger figure approached through the fog behind him. The woodsman needed to remain the emotional center of the shot. The giant had to stay partially concealed, visible without becoming a clean reveal.


Early versions failed in two directions. The camera either rotated far enough to lose the giant entirely or allowed the figure to resolve into a clear central reveal. Both produced usable motion. Neither held the intended story beat.


The three versions began with the same source keyframe and creative direction but differed in model and implementation. This is a production comparison rather than a controlled benchmark. I evaluated character performance, camera behavior, spatial continuity, concealment, and fidelity to the intended dramatic beat.


My workflows combine local inference with cloud API nodes inside ComfyUI, allowing shots to move between local and remote compute according to model access, hardware demand, speed, and production need. For this study, I ran MiniMax H3 locally while developing additional MiniMax H3 and Seedance passes through cloud infrastructure.

Seedance · Cloud

This version delivered the strongest forward movement and clearest escalation in scale. As the giant approaches, however, it becomes increasingly prominent. The shot shifts from sustained dread into direct pursuit, and the threat begins competing with the woodsman for control of the composition.

MiniMax H3 · Cloud

This version came closest to the intended dramatic structure. The camera continues toward the woodsman while the giant remains distant and secondary. The restrained movement preserves the sense that the threat has been discovered before it has been fully understood.

MiniMax H3 · Local

Run locally in ComfyUI, this pass sustains the woodsman’s defensive movement, shifting attention, axe handling, and performance through the controlled push toward his face.


The camera leaves the giant behind too early, changing the shot from a two-subject composition into an isolated character reaction. The performance improves, but the shot no longer holds the intended story beat.

DIRECTION AND SELECTION

The comparison separated two production problems, both sustaining the woodsman's performance and preserving the spatial relationship between the characters.


The MiniMax H3 cloud version is the selected result. It is the only one that holds both characters in the intended relationship for the length of the move. The local version delivers a better performance and loses the shot. The Seedance version moves well and resolves the threat too early, which is the exact failure the shot was designed to avoid.


NEXT ITERATION

The next pass begins with a simple Blender blocking render to lock the camera move, character positions, scale, timing, and occlusion before generation. That reference will drive the composition and spatial relationship while the generative pass focuses on the woodsman’s performance, concealment of the giant, and atmosphere.


WORLD DEVELOPMENT

Directed Keyframe and Environment Development


For this project, I use Ideogram to develop the first cinematic keyframes for each major location and sequence. These frames establish composition, lens language, weather, environment scale, architecture, props, and character placement before moving into video generation.


Selected frames are then refined through Krea 2 and other ComfyUI workflows. I use those tools for reference-driven iteration, inpainting, material correction, character continuity, and shot-to-shot cleanup. Once a frame is working, it becomes a visual anchor for blocking, motion tests, video prompts, and later shot evaluation.

The story is set in a remote 1815 Cascadian coastal valley. The expedition arrives by water, passes through a drowned stone threshold, unloads at a rotunda landing, cuts into the forest, and eventually fortifies the ruins. Over time, the men realize the valley is not abandoned. The structures, forest paths, altered objects, and signs of maintenance point to an occupied territory they do not understand.


The world is built through visual evidence: fitted basalt under shallow water, wet stone stairs, cedar palisades, worked wood, boundary trees, caches, oversized skulls, moved camp objects, returned tools, redirected canoes, and partial giant figures held in fog. These details support the central idea that the expedition mistakes an unfamiliar system of life for vacancy.


Palette, weather, architecture, props, scale, and geography are treated as continuity rules across the project. The visual language stays cold, wet, and photographic: marine overcast, black basalt, cedar, moss, mud, dull silver fog, practical 1815 materials, heavy 35mm grain, controlled firelight, telephoto compression, and incomplete views of the giants. The goal is to make the world feel specific, consistent, and production-ready rather than invented shot by shot.

HOW I WORK

Sequences, not single frames.


I work from an established world rather than generating shot to shot. Story, visual rules, recurring characters, locations, and props are defined before any shot begins, and every shot after that has to preserve them. Each one moves through the same path.


Story beat · Visual bible · Shot brief · Controlled keyframe · Motion versions · Selection and QC · VFX finishing · Versioned delivery


Every shot starts from a brief.

Before a model is selected I define camera behavior, character blocking, lens, lighting, atmosphere, continuity requirements, first and last frame intent, and the conditions that would break the story beat. Composition, scale, and perspective are creative decisions made in advance, not outcomes discovered afterward.


Every result is versioned.

Prompts, references, model settings, outputs, and review notes stay connected to the shot ID. A successful result can be reproduced, revised, and carried forward rather than existing as a lucky generation.


Generation is controlled, not sampled.

ComfyUI is the orchestration layer, running modular workflows across local and cloud inference. ControlNet, IP-Adapter and other reference-conditioning methods, LoRA training and integration, targeted inpainting, and clean plates protect composition, character identity, props, lighting, and atmosphere across iterations. Model choice follows the needs of the shot rather than a fixed stack, and local or cloud execution follows model access, hardware demand, and speed.


Development is nonlinear.

An approved frame can branch into several models and control strategies, with the strongest performance, camera behavior, composition, or continuity carried into the next iteration. A failed pass is not discarded without analysis. It identifies which part of the workflow needs to change.


I use an LLM-assisted development workflow to keep that branching legible: retrieving shot information, comparing versions, flagging continuity conflicts, and assembling the next prompt package against the project bible. It supports organization and analysis while creative direction, model selection, review, and approval stay with me.


Photographic realism is judged, not assumed.

I evaluate believable light, optics, materials, anatomy, motion, and spatial relationships rather than surface detail. A frame that resolves cleanly and reads wrong is still a failed frame.


VFX INTEGRATION AND FINISHING

Generated material is treated as source material, not as a finished shot. I develop and clean selected imagery in Photoshop, with Nuke and Blender providing the established production path for compositing, projection, spatial correction, atmosphere, color, and continuity as shots move toward final.


The objective is material that survives revision, integrates with conventional VFX elements, and moves downstream without remaining dependent on the model that created it.