By Mohammed Al Marzooqi · Kimi Brain Journal
A striking AI image is a useful starting point for a film. The next question is whether it can belong to a sequence: does the character remain recognizable, does the action connect, and does each shot move the story forward?
My interest in photography and filmmaking predates my AI projects. As I explore ComfyUI and local generation workflows, I find it useful to separate visual experimentation from the decisions needed to build a coherent scene. The following is a planning framework, not a report of a completed production.
Start with a short scene brief
Describe who is present, where the scene takes place, what changes, and what the audience should understand by the end. Keep the action manageable. “A detective enters a room and notices a disturbed document” gives you clearer review criteria than a long list of cinematic adjectives.
Translate the brief into a small shot list. An establishing shot can explain the space; a closer shot can reveal the discovery; a reaction shot can show its significance. Each image should have a job within the scene.
Build a reference set before making variants
Keep the approved character reference, wardrobe, location, lighting direction, and important props together. Record which files you are using and where they came from. If a portrait is generated, label it accurately. Visual resemblance alone cannot establish which model or LoRA created it.
Separate identity from styling during review. The right coat or hat can make a character concept recognizable while the face still changes. Check facial structure and proportions alongside costume and lighting.
Save the workflow with the result
ComfyUI’s official getting-started guide covers loading workflow templates and images with workflow metadata, selecting the required model, and generating an image. Those features help explain why a saved workflow can be more useful than an isolated output. See the official ComfyUI guide for the current interface and setup steps.
For your own records, keep the relevant model names, prompts, settings, input references, and output filenames together. Treat a seed as one recorded setting, not a promise that results will remain identical after changes to models or software.
Review continuity before polishing
Place candidate shots in sequence and inspect what changes between them. Does the character hold the same object in the same hand? Does the room layout remain understandable? Does the next shot continue the movement or reset it? A beautiful frame may still be the wrong frame for the edit.
Keep a short review record: accepted, revise, or reject, with one specific reason. “Face changed” or “action restarts” is more useful than “something feels off.” Change one important variable at a time when diagnosing the problem.
Build a rough edit early
A rough sequence can reveal missing coverage, repetitive pacing, or confusing action before you spend time refining every shot. Add temporary sound only when you have permission to use it, and clearly separate previews from finished work.
The goal is a scene the viewer can follow. Generating more material helps only when you know what the scene still needs. This is the process I am continuing to learn and document through Kimi Brain Journal.
About the Journal: This is an independent founder’s publication affiliated with Kimi Brain. Discussion of ComfyUI is not an endorsement by, or partnership with, its developers.