Krea2 Character --- Visual DNA to Human Character
A ComfyUI workflow that transforms a visual reference into an original human character design using Krea2.
The workflow first analyzes the supplied reference image with Qwen3-VL 4B, converts the reference's visual characteristics into a cohesive human-character prompt, and then uses Krea2 to generate the character while preserving the visual relationships of the source.
The core idea is:
Reference Image → Visual Analysis → Human Character Design → Krea2 Generation → Character Result
The workflow is designed to translate the visual DNA of a subject rather than simply turning the subject into a costume or anthropomorphic character.
What it does
Give the workflow an image of virtually any visual subject:
- animal
- plant
- object
- architecture
- artifact
- mineral
- natural phenomenon
- landscape
- abstract form
- other visual reference
Qwen3-VL analyzes the reference and produces a natural-language character-design prompt.
The prompt instructs the model to create an original human character, translating characteristics such as:
- silhouette and proportions
- distinctive shapes
- colors and color relationships
- patterns and markings
- textures and materials
- structural features
- hairstyle
- clothing construction
- accessories and ornaments
- pose and body language
- personality and visual presence
The translation is intentionally metaphorical rather than literal. For example, feathers can become layered fabric or ornamental shapes, leaves can influence garment construction, and the geometry of an object can influence the character's silhouette.
Workflow highlights
1. Reference image
Load the source image with the LoadImage node.
The image is resized to approximately 1 megapixel before being sent to the vision-language model.
2. Qwen3-VL character interpretation
The workflow uses:
qwen3vl_4b_bf16.safetensors
through the ComfyUI TextGenerate node.
The built-in instruction set is specifically designed to produce a style-agnostic human character design description.
The generated prompt:
- chooses male or female based on the visual character of the reference
- keeps the resulting character fully human
- translates visual characteristics instead of copying the source literally
- preserves color relationships and visual hierarchy
- separates character design from rendering style
- avoids generic rendering instructions
This makes the generated description suitable for subsequent interpretation by Krea2.
3. Optional LoRA
The workflow includes a LoRA Manager section.
The supplied workflow contains:
banjiesock_Krea2
at strength 1.0.
The LoRA trigger-word system is connected through
TriggerWord Toggle (LoraManager), allowing trigger words to be managed
without manually rebuilding the prompt.
If you do not have this LoRA, replace it with your own Krea2-compatible LoRA or disable/remove the LoRA section.
4. Krea2 prompt weighting
The generated character prompt is passed through:
Krea2PromptWeight
This allows the prompt to be processed using Krea2-specific prompt weighting before generation.
The workflow then applies:
Krea2T-Enhancer-Advanced
with the current workflow defaults:
- Enabled:
true - Strength:
1.5 - Text Scale:
1.5 - Debug:
false
These values can be adjusted depending on the desired balance between reference preservation and prompt influence.
5. Smart Seed Variance
The workflow uses:
RBG_Smart_Seed_Variance
to introduce controlled variation into the conditioning.
Current settings include:
- Variance preset:
Creative - Fine tune variance:
75 - Model type:
Krea2 (SingleStream) - Fade curve:
Linear - Noise injection:
All Steps - Protect mode:
First Half - Direction shift:
Bone Anatomical Coherence - Shift strength:
97 - Variance schedule:
decreasing - Cutoff step:
7 - Total steps:
10 - Cutoff strength:
0.1 - Vibe blend:
0.5
These settings are not universal requirements. They are the configuration included in this workflow and can be experimented with for different character references.
6. Two-stage Krea2 sampling
The workflow uses two ClownsharKSampler_Beta stages.
First sampler
Current configuration:
- Sampler:
linear/euler - Scheduler:
beta57 - Steps:
10 - Steps to run:
9 - Denoise:
0.9 - CFG:
1 - ETA:
0.5 - Sampler mode:
standard - Bongmath: enabled
Second sampler
Current configuration:
- Sampler:
linear/dormand-prince_6s - Scheduler:
kl_optimal - Steps:
10 - Steps to run:
1 - Denoise:
0.27 - CFG:
1 - ETA:
1 - Sampler mode:
standard - Bongmath: enabled
The second stage operates on the latent produced by the first sampler and provides the final denoising stage used by this workflow.
Model requirements
The workflow references the following model files:
Krea2
krea2_turbo_fp8.safetensors
Loaded with UNETLoader.
Qwen3-VL
qwen3vl_4b_bf16.safetensors
Loaded through the Krea2 CLIP configuration and used by TextGenerate
for visual analysis.
VAE
Wan2.1_VAE_upscale2x_imageonly_real_v1.safetensors
Used for encoding and decoding.
Custom nodes
The workflow uses several non-core ComfyUI nodes.
Important dependencies include:
comfyui-kjnodescomfyui-lora-managercomfyui-fearnworksnodesComfyUI-RBG-SmartSeedVarianceRES4LYFComfyUI-VAE-Utilscomfyui_layerstylergthree-comfy
The exact node versions recorded in the workflow may differ from the versions currently available in ComfyUI.
If a node is missing, install the corresponding custom node package and restart ComfyUI.
Main workflow structure
┌─────────────────────┐
│ Reference Image │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ Image Scale │
│ ~1 Megapixel │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ Qwen3-VL 4B │
│ Visual Analysis │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ Human Character │
│ Design Prompt │
└──────────┬──────────┘
│
┌───────────────┴───────────────┐
│ │
▼ ▼
Trigger Words LoRA
│ │
└───────────────┬───────────────┘
▼
┌─────────────────────┐
│ Krea2 Prompt Weight │
└──────────┬──────────┘
▼
┌─────────────────────┐
│ Krea2 T-Enhancer │
└──────────┬──────────┘
▼
┌─────────────────────┐
│ Smart Seed Variance │
└──────────┬──────────┘
▼
┌─────────────────────┐
│ Clownshark Sampler │
│ Stage 1 │
└──────────┬──────────┘
▼
┌─────────────────────┐
│ Clownshark Sampler │
│ Stage 2 │
└──────────┬──────────┘
▼
┌─────────────────────┐
│ VAE Decode │
└──────────┬──────────┘
▼
┌─────────────────────┐
│ Krea2 Character │
└─────────────────────┘
Comparison output
The workflow also contains an image comparison section.
It keeps the original reference and generated character available for side-by-side evaluation and creates an image reel labeled:
ReferenceKrea2 Character
This is useful for checking how successfully the generated character preserves the visual identity of the original reference.
Output
The main generated image is saved using the filename prefix:
Krea2-Character/Krea2-Character
A comparison image is also generated using:
Krea2-Character/Krea2-Character-Compare
Important design principle
This workflow does not attempt to create an anthropomorphic version of the reference.
Instead, it follows:
Visual DNA → Human Character Design
The source subject influences the human character at multiple levels simultaneously:
- Overall silhouette
- Proportions and shapes
- Color palette
- Hair
- Facial characteristics
- Clothing construction
- Patterns and ornamentation
- Materials and textures
- Accessories
- Pose and body language
- Personality and mood
The objective is a unified character design in which the visual relationships of the source remain recognizable after being translated into human design language.
Style neutrality
An important part of the workflow is separating character design from art style.
The Qwen3-VL prompt deliberately avoids prescribing:
- a specific artist
- a specific medium
- photorealism
- illustration style
- rendering technique
- line quality
- shading method
- lighting setup
- level of finish
This means the resulting character concept can later be interpreted through different visual styles without rebuilding the underlying character design.
Recommended use
For the strongest results, choose a reference with:
- a recognizable silhouette
- distinctive shapes
- clear visual relationships
- identifiable materials or textures
- meaningful color organization
- interesting patterns or structural details
Highly generic references may produce less distinctive character designs.
Notes
This workflow is intended as a creative character-design system rather than a conventional image-to-image workflow.
The Qwen3-VL stage acts as a visual-to-concept translation layer, while Krea2 performs the final visual synthesis.
The workflow was created for experimentation with Krea2, Qwen3-VL, prompt weighting, controlled variance, and RES4LYF sampling.
License
Apache License 2.0.