How to Use LoRA Training for Character Consistency in Vertical Drama

Reference images condition a generation. A trained LoRA encodes identity into the model weights. That sentence contains the complete technical distinction that determines which approach is correct for vertical drama character consistency at series scale.

A reference image tells the model what the character looks like in the images provided. The model interprets those images and produces output that is influenced by them. In a single session with a skilled operator, this produces good results. Across multiple sessions, across multiple operators, across a 70-episode series where generation batches happen over weeks, the model's interpretation of the same reference images produces variation that compounds into visible drift.

A trained LoRA encodes the character's appearance into the model's weights directly. Once the training is complete, the character's identity is not a conditioning input that the model interprets session by session. It is part of the model itself. The same face, the same features, across every output, because the training has made that face part of what the model knows rather than something it is being shown each time.

For projects requiring more than twenty clips of the same character, the training investment pays back immediately. A 70-episode vertical drama series requires between 210 and 420 clips per series regular. The LoRA training investment for that character is justified after approximately ten generation sessions.

This is the complete guide: what LoRA training is at the technical level, how to execute the training workflow for vertical drama characters, which generation tools support LoRA inputs, and how LoRA-trained models compare to reference image approaches across the specific requirements of vertical drama production.

What LoRA Training Actually Is

LoRA stands for Low-Rank Adaptation. It is a parameter-efficient fine-tuning technique that adapts a pre-trained model to specific new knowledge without modifying the model's core weights across its full parameter set.

A conventional diffusion model used for AI video or image generation contains billions of parameters that encode its general knowledge of how the visual world looks. When you prompt the model to generate a woman with dark hair in a luxury office, the model draws on all of that general knowledge to produce output. The output looks like a woman with dark hair in a luxury office because the model's training included enormous amounts of imagery that matches those descriptions. The output does not look like your specific character because the model has no specific knowledge of your character's exact facial geometry, skin tone, bone structure, or expression patterns.

LoRA training adds specific knowledge to the model without retraining it from scratch. A LoRA file is not a new model. It is a small adapter layer that modifies the existing model's behavior at specific points in its processing pipeline. Think of it as training an AI artist to recognize and consistently recreate your character's DNA: face proportions, expressions, accessories, and style cues.

The practical implication: a LoRA trained on 15 to 30 images of your vertical drama character teaches the model the specific individual whose face you want it to reproduce. After training, when you use the LoRA during generation, the model does not interpret your reference images to produce someone who looks like your character. It produces your character, because your character's specific visual identity is now encoded in the adapter weights the model applies at generation time.

The training process takes 30 to 60 minutes on a modern GPU. This is not a weeks-long research undertaking. It is a production pre-production step that fits within the schedule of any vertical drama series.

LoRA vs Reference Image: The Practical Comparison

Understanding when LoRA training is the correct approach and when reference image conditioning is sufficient requires comparing their specific performance characteristics across the requirements that vertical drama production actually creates.

Within a Single Generation Session

Reference image conditioning and LoRA training produce comparable results within a single generation session. A skilled operator using a high-quality reference image set in Seedance 2.0 or Kling 3.0's Elements system can produce character-consistent output at production quality within a continuous session.

The session boundary is where the approaches diverge.

Across Multiple Generation Sessions

Reference images condition individual generations but do not persist production state, enforce identity across regenerations, or transfer cleanly between team members over a 60-plus episode season. When a new generation session begins, the operator uploads reference images that the model interprets freshly. Small differences in which images are selected, how they are weighted, and how the model's stochastic generation process distributes its interpretation of them produce variation between sessions that compounds across a 70-episode production timeline.

A trained LoRA produces the same character in a new session without requiring image upload or re-interpretation. The LoRA is loaded at inference time, and the character's trained identity is applied across every generation in that session and in every subsequent session. The operator who works on episode thirty of a series generates the same character as the operator who worked on episode one, without any session-to-session transfer of reference images.

Across Multiple Operators

The multi-operator problem is the most commercially significant consistency challenge in vertical drama franchise production. A sequel produced by a different generation operator than the original series must produce the same characters the franchise audience recognizes.

Reference image approaches require the new operator to learn how to use the reference images correctly: which images to prioritize, how to structure the reference input, how to compensate for the specific ways the reference images interact with the generation tool's interpretation. Different operators using the same reference images produce different output because their reference image selection and prompt engineering decisions differ.

A trained LoRA eliminates this operator dependency. Any authorized operator who loads the trained LoRA file produces the same character, because the character's identity is in the LoRA weights, not in the operator's reference image selection skills.

Under Different Lighting and Angles

Reference images condition the model on the specific lighting and angle conditions represented in the reference set. When a generation session requires the character in a lighting condition or angle not well-represented in the reference images, the model has less information to work from and produces more variation.

LoRA training encodes the character across the angle and lighting diversity present in the training dataset. A training set that includes the character in multiple lighting conditions produces a trained model that generates the character consistently across those conditions, because the training has explicitly taught the model how the character's specific facial geometry behaves under each of them.

The Vertical Drama LoRA Training Workflow

Step 1: Dataset Curation

The training dataset is the most consequential decision in the LoRA training process. The dataset quality determines the trained model's output quality at a higher rate than any parameter setting in the training configuration.

Dataset size: 12 to 30 images is the production-grade range for vertical drama character LoRA training. Below 12 images, the trained model does not have sufficient angle and lighting coverage to generalize well across generation conditions. Above 30 images, the incremental quality improvement diminishes relative to the curation time required. A curated dataset beats a huge messy folder.

Angle coverage requirements:
Front-facing portrait at close to medium distance. This is the primary reference for the model's learning of facial geometry and should constitute 3 to 5 images in the training set.
Three-quarter left. 2 to 3 images at slightly varying distances.
Three-quarter right. 2 to 3 images at slightly varying distances.
Slight upward angle from below eye-line. The angle vertical drama uses to communicate urgency and dominance. 2 images.
Slight downward angle from above eye-line. The angle used for vulnerability or defeat. 2 images.
Profile left and right. 1 to 2 images each for full angle coverage.

Lighting coverage requirements:
Warm key light from camera left. The CEO romance genre's standard interior lighting setup.
Neutral even lighting. For scenes where environment lighting has not been established.
Cool or high-contrast dramatic lighting. For confrontation scenes and the paywall episode's key moments.

Expression coverage requirements:
The neutral-to-controlled baseline the character occupies in most scenes.
The suppressed intensity expression, the most commercially important expression for vertical drama's close-up register.
The moment of involuntary vulnerability, if the character arc includes this transition.

Image quality requirements:
High resolution: 1024 x 1024 pixels minimum, 2048 x 2048 preferred.
Sharp focus on the face. Blurred images teach the model blurred faces.
No sunglasses, masks, heavy props, or anything that obscures the facial geometry the training needs to learn.
No heavy retouching or filters that alter the skin tone in ways that do not represent the character's actual appearance.
Consistent with the character's visual identity as it will appear in the series: correct wardrobe category, correct hair state, correct makeup register.

Naming and captioning: Each training image requires a caption that describes the image and includes the character's trigger word. The trigger word is the term you will use in generation prompts to activate the LoRA. Don't just name your character sara. Name them ohwx_sara. Using a rare token prevents the AI from bleeding in concept data. If you name a character rocky, the model might import boxing associations. A rare token like ohwx_character or vd_lead_female creates a unique identifier that the model associates exclusively with the trained character.

Step 2: Training Tool Selection

The training tool determines which generation models the trained LoRA can be deployed in. The selection depends on the production's primary generation tool.

For productions using Flux.1-based generation tools: The AI Toolkit by Ostris is the standard training tool. Flux.1 Dev is the base model for quality-priority training. Z-Image Turbo is the base model for speed-priority training. Flux.1 is the gold standard for prompt adherence. If the character needs to hold a specific object while wearing a specific shirt, Flux.1-based LoRAs are the correct choice.

For productions using LTX-based video generation: LTX's native LoRA trainer supports video-native training. LTX LoRA training runs on video data natively, not image frames, producing better temporal consistency, motion quality, and character continuity across frames. The resulting LoRA weights can be loaded into ComfyUI workflows or deployed via fal.ai and Replicate endpoints.

For productions using Wan 2.2 or open-weights models: Kohya_SS provides the training interface for open-weights character LoRAs. The trained LoRA files are in Hugging Face diffusers format and can be deployed locally or through cloud inference endpoints.

For productions using ComfyUI custom pipelines: Any of the above training approaches produces LoRA weights that can be loaded into ComfyUI node workflows, allowing the trained character to be combined with multiple generation tools, style LoRAs, and IC-LoRA motion adapters within a single generation graph.

Step 3: Training Parameter Configuration

The specific training parameters vary by tool, but the key parameters that affect vertical drama character LoRA quality are consistent:

Learning rate: 1e-4 is the standard starting point for character LoRA training. Too high a learning rate causes overfitting, where the trained model reproduces the training images too closely and cannot generalize to new angles, lighting conditions, or expressions. Too low a learning rate produces underfit output where the trained model's character is inconsistently applied.

Training steps: 1,000 to 2,000 steps is the production range for a 15 to 30-image dataset. Fewer than 1,000 steps produces an undertrained model. More than 2,000 steps risks overfitting on a small dataset.

LoRA rank: The rank parameter determines how much of the model's parameter space the LoRA adapter modifies. Higher rank produces a more powerful adapter that can capture more character-specific detail but also increases the risk of overfitting and produces a larger LoRA file. Rank 16 to 32 is the production standard for character LoRAs. Rank 64 and above is appropriate for complex character designs requiring high-fidelity reproduction of specific details.

Network alpha: Set to half the rank value. Rank 16, alpha 8. Rank 32, alpha 16. This ratio prevents the LoRA from over-dominating the base model's generation.

Step 4: Validation and Approval

After training completes, the LoRA is tested against the production's approval criteria before being filed in the character asset library.

The validation test runs 10 to 15 generations using the trained trigger word, varying the scene description, angle specification, and lighting condition across the test set. Every validation generation is reviewed against the training dataset's primary front-facing reference image.

The approval criterion: does every validation generation produce output that a reviewer identifies as the same specific individual as the training set's reference without prompting? If yes, the LoRA is approved and filed in the character asset library with its trigger word, training tool, base model version, and training date.

If any validation generation produces output where the character's facial geometry, skin tone, or key physical features have drifted outside the approval threshold, the training requires adjustment. The most common training adjustment: review the training dataset for images where the character's appearance was affected by heavy makeup, unusual lighting, or props, and remove those images before retraining.

Generation Tools That Support LoRA Inputs

Not all generation tools accept LoRA files as inputs. The tool selection for a LoRA-based vertical drama production must account for which tools in the production's pipeline can accept and apply the trained LoRA weights.

ComfyUI

ComfyUI is the most flexible LoRA deployment environment available in 2026. It accepts LoRA files for any compatible base model through the Load LoRA node, supports multiple LoRAs simultaneously through stacked loading, and allows fine-grained control of each LoRA's weight at generation time.

For vertical drama productions using ComfyUI, the trained character LoRA is loaded alongside a style LoRA that enforces the series' visual register and potentially an IC-LoRA motion adapter that controls camera behavior. The combination, character identity plus visual style plus motion control, produces the most consistent output available from any non-proprietary generation workflow.

The limitation: ComfyUI requires technical operator skill. Node-based workflow setup, LoRA file management, and parameter configuration require expertise that a generation operator without ComfyUI experience will take days to develop.

LTX-2.3

LTX-2.3 supports native character LoRA training and deployment through its IC-LoRA adapter system. The strongest multi-scene workflow combines all layers: reference image at frame 0, character LoRA loaded at inference, a fixed 50 to 80-word character prompt block across all scenes, and IC-LoRA adapters for scenes requiring precise body movement.

LTX-2.3's LoRA system is specifically designed for video generation rather than adapted from image generation, which means the trained character LoRA's output maintains temporal consistency across frames rather than producing identity that holds in still frames but drifts across motion.

Wan 2.2

Wan 2.2 with a trained LoRA is the production-standard open-weights workflow for production volumes that justify the training investment. The Wan 2.2 community has produced substantial documentation for character LoRA workflows, and the model's open-weights architecture means the trained LoRA files are fully portable: exportable, shareable, and self-hostable.

Cloud Deployment via fal.ai and Replicate

For production companies that do not want to manage local GPU infrastructure, trained LoRA weights in Hugging Face diffusers format can be deployed via fal.ai and Replicate endpoints. LTX exports LoRA weights in standard Hugging Face diffusers format for cloud deployment via fal.ai and Replicate. This deployment approach makes the trained character LoRA available to any authorized operator via API without requiring local GPU infrastructure.

Tools Where LoRA Is Not Currently Supported at Consumer Tier

Seedance 2.0, Kling 3.0, and Veo 3.1 do not accept external LoRA file inputs in their consumer-facing interfaces as of mid-2026. These tools implement their own character consistency mechanisms: Seedance 2.0 through Soul ID trained models and multimodal reference inputs, Kling 3.0 through the Elements system and Subject Consistency features, Veo 3.1 through character reference image conditioning at generation time.

For productions using these tools as primary generators, Soul ID is the functionally equivalent to LoRA training within the Higgsfield ecosystem, producing a trained character model that persists across sessions without reference image re-upload. For productions requiring LoRA-based character deployment in these tools specifically, ComfyUI integration with API access to these tools' underlying models is the technical pathway, requiring development expertise beyond the consumer interface workflow.

The Stacked LoRA Approach for Advanced Production Control

The most sophisticated character consistency workflow for vertical drama in 2026 uses multiple LoRAs simultaneously, each targeting a different dimension of the series' visual consistency requirements.

For the highest consistency, the technique is a specific stack: a low-strength character LoRA at strength 0.6 to get the general body shape, hair, and visual register; a PuLID adapter at strength 0.8 to lock in the precise facial features; and a ControlNet OpenPose to force the posture. This triple-threat approach allows studios to make AI productions where the character looks identical in every frame.

Applied to vertical drama specifically, the stacked approach looks like this:

Layer 1: Character identity LoRA. Trained on the character's 20 to 30 reference images using the workflow above. Applied at inference strength 0.6 to 0.8, producing the character's general appearance including body type, hair, skin tone, and overall visual register. The lower strength prevents the character LoRA from over-constraining the generation at the expense of scene-specific visual elements.

Layer 2: Visual style LoRA. A second LoRA that enforces the series' visual style, color register, and production aesthetic. Trained on a selection of approved generation outputs from the series that collectively represent the intended visual style rather than on character-specific images. Applied at inference strength 0.4 to 0.6, ensuring it influences but does not override the character identity layer.

Layer 3: IC-LoRA motion adapter. For scenes where the character's posture, movement, or camera angle needs to match a reference, an IC-LoRA motion adapter extracts the pose structure from a reference video or image and applies it to the generation while the character and style layers handle the visual identity. IC-LoRA unlocks precise motion control by separating motion from visual styling. The key to success: let the reference video handle motion, and use prompts purely for visual style.

The stacked approach requires ComfyUI for deployment and a generation operator who understands how to balance multiple LoRA strengths without producing conflicts between the layers. For productions with the technical expertise to implement it, the stacked approach produces the highest available character consistency from any open-weights generation workflow.

LoRA Training vs Soul ID: The Decision Framework

For vertical drama production companies choosing between custom LoRA training and Soul ID as their primary character consistency infrastructure, the decision depends on three operational factors.

Primary generation tool: Soul ID integrates directly with Higgsfield's generation tools and, through the trained model connection, with Seedance 2.0. For productions primarily using Seedance 2.0 and Higgsfield's ecosystem, Soul ID is the production-rational choice because it requires no external training infrastructure and produces output within the same workflow. For productions using open-weights tools, LTX-2.3, Wan 2.2, or ComfyUI-based pipelines, custom LoRA training is the correct approach because those tools accept LoRA inputs but not Soul ID models.

Portability requirements: A trained LoRA file is a portable asset. It can be exported, transferred, deployed via cloud inference endpoints, and integrated into any ComfyUI workflow. Soul ID models are platform-bound: they exist within the Higgsfield workspace and are not portable to other generation tool ecosystems. Productions that require cross-tool character deployment need LoRA training for portability.

Technical infrastructure: Soul ID requires no GPU infrastructure, no training configuration, and no ComfyUI expertise. A production company with no technical development capability can train a Soul ID character in 5 minutes through the Higgsfield interface. Custom LoRA training requires either local GPU infrastructure or cloud training credits, training configuration expertise, and ComfyUI or API deployment capability. The technical barrier is real and should be assessed honestly before committing to a LoRA-based production pipeline.

Axis AI Studios Perspective

LoRA training is the character consistency infrastructure that scales where reference images cannot. For a 70-episode series with 3 to 4 series regulars, each requiring 200 to 400 generation clips across multiple sessions and potentially multiple operators, the training investment for each character's LoRA is paid back before the first episode batch is complete.

The production companies that invest in LoRA training infrastructure before production begins generate a different asset class from the production companies that rely on reference image conditioning. A trained LoRA file for a character who achieved 10% paywall conversion is not just a production tool. It is the evidence that the character's specific visual identity can be reproduced on demand, which is the technical foundation of franchise development, sequel production, and cross-platform licensing.

At Axis AI Studios, LoRA training is part of the pre-production infrastructure for every series where franchise development is a commercial objective. The trained files are managed within the character asset library system described in the franchise guide, versioned across productions, and deployed through ComfyUI-based pipelines that allow the character and style LoRAs to be combined with scene-specific motion control adapters.

For production companies building vertical drama content designed for franchise development and requiring LoRA-based character consistency infrastructure, reach out at business@axisaistudios.com.

FAQ

How Long Does LoRA Training Take and What Does It Cost?

Training time runs 30 to 60 minutes on a modern GPU for a 15 to 30-image dataset at 1,000 to 2,000 training steps. Cloud training via services like fal.ai or Replicate costs approximately $2 to $8 per training run depending on the base model and step count. Local GPU training on a consumer-grade GPU costs the electricity and the time. For a vertical drama series requiring 3 to 4 trained character LoRAs, the total training investment is $6 to $32 in cloud credits and approximately 2 to 4 hours of dataset curation and training time per character.

Can a Single LoRA Be Used Across Multiple Generation Tools?

Partially. A LoRA trained on a specific base model can only be deployed in generation tools that use the same or a compatible base model. A Flux.1-trained LoRA works in ComfyUI with Flux.1 models and in generation tools that use Flux.1 as their underlying architecture. It does not work in LTX-2.3, Wan 2.2, or tools built on different base models. For productions requiring the same character identity across multiple generation tools with different base models, separate LoRAs need to be trained for each base model ecosystem. The training dataset is reusable across multiple training runs on different base models, so the dataset curation investment is made once even when multiple LoRAs are produced.

What Is the Most Common LoRA Training Failure and How Is It Fixed?

Overfitting: the trained model reproduces the training images too closely and fails to generalize to new angles, expressions, or lighting conditions. The failure presents as a character who looks correct in generation scenes that are similar to the training images and who drifts in scenes that are significantly different. Overfitting is fixed by reducing the training steps, increasing the dataset size with more diverse images, or reducing the LoRA rank. The most reliable prevention: ensure the training dataset covers the full angle and lighting diversity the series requires before training begins rather than discovering coverage gaps in the validation test.

Further Reading

For the character asset library system that stores trained LoRA files alongside Soul ID models and approved generation outputs across franchise productions, the guide to building an AI character asset library covers the complete franchise-grade character infrastructure workflow.

For the prompt engineering language that deploys trained LoRA characters correctly in ComfyUI and LTX-2.3 generation sessions, the prompt engineering guide for vertical drama generation covers the specific trigger word syntax and stacking approaches that produce production-grade output.

For the generation tool routing framework that determines which scenes use LoRA-based open-weights tools versus Soul ID-connected Seedance 2.0, the Seedance vs Kling vs Veo comparison guide covers the complete scene-type routing logic across all three primary tools.