🧩 Flux-Kontext Character Expansion Workflow
Overview
This custom Flux-Kontext workflow is a major extension of the base model, built to handle multi-image composition, LoRA stacking, and multi-step generation for high-quality, consistent character datasets.
It’s designed for character creation, LoRA training, and regeneration of low-quality inputs — giving you complete control over appearance, context, and detail.
The NSFW version is here: https://civarchive.com/models/2088506
It uses line art refernces with inpainting of the subject - so you find any position you like and your reference image will be painted into it.
Do you want ALL the NSFW LoRAs, checkpoints, and Workflows? Here is my patreon - i will uploading some CRAZY stuff I have been working on.
https://www.loboforge.com/@LoboForge
🖼️ Dual-Image Context System
The workflow allows you to combine two different input images into a single coherent generation:
If both inputs are active, the system blends their visual and semantic context — e.g., “the man and the woman are dancing.”
If only one image is active, the model evolves that image using the new prompt — e.g., “same face and clothes, the man is riding a horse.”
This setup enables dynamic, instruction-based image merging without losing character identity.
🧠 LoRA Power Loader
A custom LoRA Power Loader lets you stack multiple LoRAs simultaneously, each with independent strength control.
You can fine-tune how each LoRA influences the output — controlling aspects like:
Facial identity
Body type and proportions
Clothing style
Lighting or artistic detail
This provides granular creative direction for dataset generation or character refinement.
🔍 Latent Upscale & Restoration
A built-in Latent Upscale node enhances weak or low-resolution inputs before re-generation.
It restores missing structure and detail at the latent level, producing clean, high-resolution results without introducing blur or artifact noise.
Use this feature to:
Improve dataset quality
Restore bad or compressed source images
Regenerate old renders into consistent HD-quality material
🧩 Multi-Prompt / Multi-Step Generation
The workflow supports multiple prompts and generation passes, automatically rendering your character in:
Different poses
Varied lighting and environments
Multiple angles and camera distances
This produces a rich, consistent dataset ideal for training LoRAs or any character model requiring 360° diversity and visual coherence.
🎯 Use Cases
Creating training data for LoRA or DreamBooth
Generating pose/lighting variation sets for the same character
Cleaning and enhancing low-quality reference material
Building consistent visual identity across scenes
⚙️ Requirements
Flux-Kontext nodes (ComfyUI Flux / Florence integration)
LoRA Power Loader node
Latent Upscale node
Any Flux-compatible base model
Optional: additional LoRAs placed in
/models/Lora/
Description
🧠 Flux-Kontext Character Creator v2.1 — Advanced LoRA Trainer
This version expands on the original character training workflow with a complete expression and motion suite for ultra-consistent LoRA training.
It now captures every aspect of personality, anatomy, and lighting for professional-grade dataset generation.
🌟 What’s New
🎭 Full Expression Range
Added five emotional variants to teach facial consistency and versatility:
Neutral
Happy (natural smile)
Angry / focused
Sad / contemplative
Surprised / shocked
Each expression maintains identical clothing, proportions, and lighting to preserve identity across moods.
🏃 Dynamic Poses
Introduced movement and gesture awareness to improve body and limb accuracy:
Walking / mid-step
Reaching or waving
Action / tension poses (arms extended or turning)
These help LoRAs generalize character movement and physical realism.
💡 Lighting Control
Three distinct lighting setups expand realism and visual depth:
Soft studio lighting (neutral tone)
Backlit silhouette (outline training)
Harsh midday sunlight (contrast and detail recovery)
This ensures your LoRA adapts to multiple illumination conditions without losing likeness.
📝 Improved Caption System
The auto-caption sidecar .txt files have been upgraded to include full prompt context:
<lora_name>, smiling, close-up portrait, same outfit, white background
<lora_name>, inside dark castle, candlelight, scared expression, dramatic moonlight
<lora_name>, walking, outdoor trail, daylight, relaxed expression
...
Every generated image now includes a precise training tag with emotion, pose, and lighting data — ready for LoRA training pipelines.
🧩 Ideal Use Case
Designed for character LoRA creation, pose reference generation, and multi-scene identity training, this workflow automates complex dataset creation in a single run — all while maintaining perfect visual and contextual consistency.