ComfyUI for Vertical Drama Production: The Node Workflow That Combines Character, Style, and Motion
The reason ComfyUI is the correct production environment for vertical drama generation is not that it produces better individual outputs than Seedance or Kling in isolation. For dialogue close-up generation, Kling 3.0 with Soul ID produces character consistency that a base ComfyUI workflow cannot match at equivalent generation time. The reason ComfyUI is the correct production environment is the workflow graph itself.
Workflow.json is diffable. You can git diff two versions of a workflow and see exactly which sampler changed, which LoRA was added, which conditioning node was rewired. No other AI image stack gives you this. That reproducibility is the production requirement that vertical drama's 70-episode consistency challenge demands and that consumer-facing generation tools cannot provide.
A generation operator using Kling 3.0 through a consumer interface who needs to reproduce the exact generation configuration from session three in session eight has no systematic mechanism for doing so. The session three configuration exists in the operator's memory and in the generation log they manually maintained. The session eight reproduction is an approximation.
A generation operator using ComfyUI who needs to reproduce session three's configuration in session eight opens the session three workflow.json and runs it. The reproduction is exact. No approximation is required.
What ComfyUI Is and How It Works
The node-based architecture means every decision in the generation process is explicit and modifiable at the node level. A character LoRA's strength value is a node parameter. The ControlNet conditioning strength is a node parameter. The IP-Adapter weight is a node parameter. These parameters are stored in the workflow.json and reproduce identically when the workflow is run again, on the same machine or on a different one.
The ComfyUI Desktop app shipped in 2026 for one-click install on macOS and Windows, managing updates, models, and custom nodes through a GUI. For production operators who are not comfortable with command-line installation and configuration, the Desktop app provides the same workflow capabilities in a managed environment without requiring Python environment configuration.
The Core Vertical Drama Generation Graph
The core generation graph for vertical drama close-up production combines four control layers stacked in a specific order. The order matters: each layer's output conditions the next layer's input, and the stacking order determines which layer's specifications take precedence when layers partially conflict.
The most reliable character consistency in ComfyUI comes from combining IP-Adapter FaceID for face consistency, a character LoRA for style and body consistency, and ControlNet for pose control. For vertical drama production, a fourth layer, the style LoRA, enforces the production's visual register on top of the character identity and pose control.
Layer 1: The Base Checkpoint.
The foundation model for the generation graph. FLUX.1 Dev is the current production standard for character-driven close-up generation in ComfyUI. FLUX's 12-billion-parameter transformer has stronger prompt-following than most earlier diffusion checkpoints. Load the base checkpoint with a Load Checkpoint node.
Layer 2: The Character LoRA.
A LoRA trained on specific characters bakes identity into the weights themselves. It survives dramatic pose, lighting, and scene changes in ways that seed locking cannot. Load the character LoRA with a LoRA Loader node connected to the base checkpoint. Strength: 0.6 to 0.75 for primary character LoRAs. This range enforces character identity without overcrowding the generation's response to other conditioning inputs.
Layer 3: IP-Adapter FaceID.
IP-Adapter FaceID specifically targets face consistency, making it ideal for character work. The IP-Adapter node takes the character reference image as input and applies facial feature conditioning to the generation. Weight: 0.80 to 0.85. The higher weight here relative to the character LoRA reflects IP-Adapter's role as the primary face identity anchor while the character LoRA handles body type and clothing.
Layer 4: ControlNet.
The camera geometry enforcement layer. For vertical drama close-up production, the ControlNet node takes an OpenPose skeleton or depth map as its conditioning input and enforces the camera geometry specification from the production's style guide. Strength: 0.70 to 0.80 for standard close-up coverage. This range enforces the geometric specification while allowing natural variation in the character's micro-expression within the specified frame.
Layer 5: Style LoRA (optional per production).
For productions that use a production-specific style LoRA rather than relying entirely on prompt-layer style specification, the style LoRA node connects alongside the character LoRA. Strength: 0.30 to 0.45. The lower strength relative to the character LoRA ensures the style conditioning supplements rather than overrides the character identity conditioning. Stacking LoRAs is multiplicative, not additive. A style LoRA at 0.4 combined with a character LoRA at 0.7 produces a combined conditioning effect that is stronger than either individually.
The Video Generation Extension: LTX and IC-LoRA
The core graph described above produces still images: character reference frames, style verification outputs, and ControlNet geometry confirmation outputs that feed into video generation tools like Seedance and Kling. For productions that want to generate video output directly within ComfyUI, the video generation extension uses LTX-Video with IC-LoRA motion adapters.
The IC-LoRA motion adapter extends the character consistency from the still image graph into the video generation: the character's visual identity established by the LoRA and IP-Adapter layers carries into the motion sequence, maintaining character consistency across the clip rather than only in a single frame.
For vertical drama production, the LTX IC-LoRA workflow is the appropriate routing for:
Close-up emotional performance clips where the character's micro-expression motion across 3 to 8 seconds is the primary commercial asset and consistency from frame one to frame last is required.
Atmospheric and transition clips where the style LoRA's visual register needs to be maintained across motion without the character identity control that Seedance and Kling provide through their native reference mechanisms.
The Workflow.json as a Production Asset
The vertical drama production graph's workflow.json file is the production asset that ComfyUI's architecture makes possible and that no consumer-facing generation tool provides.
A workflow PNG is the entire pipeline. Drag it in, get identical results. For production management, this means:
Session reproducibility. The session eight operator who opens the session three workflow.json and runs it produces output with the same character LoRA strength, the same IP-Adapter weight, the same ControlNet conditioning strength, and the same style LoRA configuration as session three. The generation log's prompt documentation combined with the workflow.json produces exact session reproduction.
Version control integration. The workflow.json is diffable through standard git version control. The production company that version-controls its workflow.json files can identify exactly when any parameter changed between production sessions and what the change was. This is the audit capability that the continuity bible's generation log requires but that prompt-based generation systems cannot provide at the same precision.
Quality control standardization. The approved production workflow.json is the generation standard that all operators on the production use. An operator who modifies the workflow parameters to produce an output they prefer has made a documented change that the git diff will identify. The quality control system that requires workflow.json version control eliminates undocumented parameter drift between operators and sessions.
When to Use ComfyUI Versus Seedance and Kling Directly
ComfyUI is the correct generation environment for vertical drama production in three specific scenarios. For other scenarios, Seedance 2.5 and Kling 3.0's native interfaces produce equivalent or superior results with lower setup overhead.
Use ComfyUI when production-scale reproducibility is the primary requirement. A production company generating 70 episodes across six to eight weeks with multiple operators requires the workflow.json reproducibility that ComfyUI provides. The consumer-facing Seedance and Kling interfaces do not offer equivalent session reproducibility documentation.
Use ComfyUI when multiple conditioning layers must be combined simultaneously. A scene that requires character LoRA, IP-Adapter, ControlNet, and style LoRA conditioning simultaneously is a scene where ComfyUI's node graph produces the combined conditioning that consumer-facing tools apply only partially. The Seedance and Kling interfaces support reference inputs and some ControlNet conditioning, but not the full stacked conditioning that ComfyUI's graph enables.
Use ComfyUI for the character reference pack build and approval. Building the approved character reference images that feed into Seedance and Kling generation sessions is a still-image production task that ComfyUI handles optimally. The character LoRA, IP-Adapter, and ControlNet combination in ComfyUI's still-image graph produces the most precisely controlled character reference images for the production's reference pack.
Use Seedance 2.5 directly for 30-second performance clip generation. Seedance 2.5's native 30-second generation and 50-reference support produces video performance clips more efficiently than ComfyUI's video extension for most vertical drama dialogue scene types. The ComfyUI LTX workflow is appropriate for specific motion control scenarios where IC-LoRA motion transfer is the primary requirement.
Axis AI Studios Perspective
ComfyUI is not the first tool a generation operator learns when entering vertical drama production. It is the tool they learn when they want to operate at production scale rather than at session scale. The consumer-facing Seedance and Kling interfaces are faster to start with and produce excellent single-session results. ComfyUI produces the session reproducibility and multi-layer conditioning control that 70-episode production at professional quality requires.
At Axis AI Studios, ComfyUI is the reference pack build environment for every production. The character LoRA, IP-Adapter, and ControlNet combination in the ComfyUI still-image graph produces the approved reference images that feed into Seedance and Kling generation sessions. The workflow.json for each production's reference pack build is version-controlled and carried across the full episode run.
For businesses who want to commission AI-native vertical drama from a production partner whose generation infrastructure includes production-scale reproducibility through workflow version control, reach out at business@axisaistudios.com.
FAQ
What Hardware Does ComfyUI Require for Vertical Drama Production?
16GB or more VRAM minimum for the video generation extension. 24GB or more for meaningful quality output at vertical drama production standard. For the still-image reference pack build workflow using FLUX with character LoRA, IP-Adapter, and ControlNet, 16GB VRAM produces acceptable results at standard resolutions. For the LTX IC-LoRA video generation workflow, 24GB VRAM is the practical minimum for 9:16 video output at 1080p resolution. Cloud GPU access through RunComfy or equivalent platforms provides the required VRAM without local hardware investment for production companies that do not own compatible hardware.
Can ComfyUI Be Used for the Full 70-Episode Generation or Only for Reference Pack Building?
ComfyUI can be used for full 70-episode generation through the video extension workflows, but this is the correct routing only for productions where session reproducibility and multi-layer conditioning control are worth the higher setup overhead relative to Seedance and Kling's native interfaces. For most vertical drama productions, the optimal workflow uses ComfyUI for reference pack building and still-image quality verification, and routes video generation to Seedance 2.5 for dialogue close-ups and performance clips where the native 30-second generation and 50-reference support produce efficient high-quality output.
How Long Does It Take to Build the Initial ComfyUI Workflow for a New Production?
Building the initial production workflow.json for a new series, including character LoRA loading, IP-Adapter configuration, ControlNet setup, and style LoRA integration, takes two to four hours for a generation operator who has previous ComfyUI experience. The first run of the workflow for reference pack approval adds one to two hours. Subsequent productions that use the same workflow structure with updated character LoRA and style LoRA inputs take thirty to sixty minutes to configure rather than two to four hours, because the base graph structure carries forward.
Further Reading
For the character LoRA training that feeds the ComfyUI character identity layer described in this post, the guide to using LoRA training for character consistency in vertical drama covers the complete LoRA training process, training dataset requirements, and deployment configuration.
For the ControlNet camera geometry reference library that the ComfyUI ControlNet node uses, the guide to using ControlNet for consistent camera angles in AI vertical drama covers the OpenPose skeleton library, depth map configuration, and strength settings for vertical drama close-up coverage.
For the tool evaluation battery that confirms ComfyUI's production readiness for a specific production's requirements before the workflow is built, the guide to how to evaluate a new AI video tool for vertical drama covers the five tests and pass criteria that apply to every generation environment including ComfyUI.
Let's set
the new standard together.
If you're working on something, we'd like to hear about it.