Complete ComfyUI Workflow for Flux 2 Klein

ComfyUI has become the standard interface for running Flux models, and Black Forest Labs has embraced this by releasing official workflow templates for Flux 2 Klein. This guide walks you through everything from initial setup to advanced configurations that will have you generating images in minutes.

ComfyUI's node-based approach might seem intimidating at first, but it offers flexibility that simpler interfaces can't match. Once you understand the basic flow, you can customize Klein's behavior precisely to your needs.

Prerequisites

Before we start, ensure you have:

• ComfyUI installed (latest version recommended)
• Python 3.10+ with CUDA support
• 12GB+ VRAM GPU for Klein 4B (20GB+ for 9B)
• Sufficient storage (~15GB for model files)

If you don't have ComfyUI installed, clone the repository and follow the installation instructions on their GitHub page.

Downloading Model Files

Flux 2 Klein requires three main components:

1. Main Model Weights

Download from Hugging Face:

For 4B model:

black-forest-labs/FLUX.2-klein-4B

For 9B model:

black-forest-labs/FLUX.2-klein-9B

Place the .safetensors file in:

ComfyUI/models/diffusion_models/

2. Text Encoder (T5)

Klein uses the T5 text encoder. Download:

google/t5-v1_1-xxl

Place in:

ComfyUI/models/text_encoders/

3. VAE

The VAE handles image encoding/decoding:

black-forest-labs/FLUX.1-dev (ae.safetensors)

Place in:

ComfyUI/models/vae/

Basic Workflow Setup

Here's the minimal node setup for text-to-image generation:

Basic ComfyUI workflow for Flux 2 Klein text-to-image generation

Essential Nodes

1. Load Diffusion Model

• Select your Klein 4B or 9B model
• This is the core generation model

2. CLIP Loader

• Load the T5 text encoder
• Handles prompt interpretation

3. VAE Loader

• Load the Flux VAE
• Handles latent-to-image conversion

4. CLIP Text Encode

• Connect your prompt here
• Outputs conditioning for the sampler

5. Empty Latent Image

• Set your desired resolution (1024x1024 recommended)
• Batch size for multiple images

6. KSampler

• Steps: 4 (Klein is distilled for few steps)
• Sampler: euler
• Scheduler: normal
• CFG: 1.0-2.0 (Klein works with low CFG)
• Denoise: 1.0

7. VAE Decode

• Converts latent to viewable image

8. Save Image

• Outputs your final result

Connecting the Nodes

The basic flow is:

Prompt → CLIP Text Encode → KSampler
Model Loader → KSampler
Empty Latent → KSampler
KSampler → VAE Decode → Save Image
VAE Loader → VAE Decode

Optimal Settings for Klein

Klein's distilled nature means different settings than typical Stable Diffusion workflows.

Sampler Settings

Resolution Settings

Klein handles various resolutions well:

• 1024x1024 - Default, fastest
• 1152x896 - Landscape
• 896x1152 - Portrait
• 1536x1024 - Wide landscape
• 1024x1536 - Tall portrait

Higher resolutions increase VRAM usage and generation time.

Image-to-Image Workflow

Klein supports image-to-image generation for editing and variations.

Additional Nodes Needed

1. Load Image

• Brings in your source image

2. VAE Encode

• Converts image to latent space

Modified Settings

• Denoise: 0.4-0.8 - Lower values keep more of original
• Connect VAE Encode output to KSampler's latent input instead of Empty Latent

Multi-Image Reference Workflow

One of Klein's advanced features is accepting multiple reference images for generation.

Setup

1. Load multiple images
2. VAE Encode each
3. Use LatentBatch node to combine
4. Connect batched latents to KSampler

This allows Klein to blend concepts from multiple sources into coherent outputs.

Advanced workflow supporting multiple reference images

Using Official Workflows

Black Forest Labs provides official ComfyUI workflows optimized for Klein.

Importing Official Workflows

1. Download workflow JSON from BFL's GitHub
2. In ComfyUI, click "Load" or drag the file
3. All nodes configure automatically
4. Ensure model paths match your installation

Available Official Workflows

• Text-to-Image Basic - Simple prompt to image
• Image-to-Image - Edit existing images
• Multi-Reference - Combine multiple source images
• Batch Generation - Generate multiple variations

Troubleshooting Common Issues

Out of Memory (OOM) Errors

Solutions:

• Reduce resolution to 1024x1024
• Enable attention slicing in ComfyUI settings
• Use FP8 quantized model
• Close other VRAM-consuming applications

Slow Generation

Check:

• CUDA is properly installed
• GPU is being utilized (not CPU fallback)
• No background processes consuming GPU
• Latest ComfyUI version installed

Poor Quality Output

Verify:

• Using correct step count (4 for Klein)
• CFG scale is low (1.0-2.0)
• Proper model files loaded (not corrupted)
• Prompt is descriptive and clear

Model Not Loading

Ensure:

• File placed in correct directory
• File extension is .safetensors
• Sufficient VRAM for model size
• ComfyUI restarted after adding models

Performance Optimization

Maximize Klein's speed in ComfyUI:

Enable GPU Acceleration

In ComfyUI settings, ensure:

• CUDA execution provider selected
• FP16 precision enabled
• Memory efficient attention enabled (for lower VRAM)

Workflow Optimization

• Avoid unnecessary nodes
• Don't chain multiple KSamplers unless needed
• Use batch generation instead of repeated single runs
• Cache frequently used latents

Saving and Sharing Workflows

Export Workflow

Click "Save" to export your workflow as JSON. This preserves all node connections and settings.

Share with Others

Workflows can be shared as:

• JSON files
• Workflow images (metadata embedded)
• Node configurations in documentation

Frequently Asked Questions

What's the minimum ComfyUI version for Klein?

The latest version is recommended. Older versions may lack required nodes or optimizations for Flux models.

Can I use Klein with ControlNet in ComfyUI?

ControlNet support for Klein is developing. Some community implementations exist but aren't as mature as SDXL ControlNet.

Why is my generation taking longer than expected?

Klein should generate in 1-3 seconds on capable hardware. If slower, check that CUDA is being used and no memory offloading is occurring.

How do I add LoRAs to the Klein workflow?

Add a "Load LoRA" node between the model loader and KSampler. Connect model and CLIP outputs through the LoRA node.

Can I queue multiple generations in ComfyUI?

Yes, use the queue feature to batch prompts. Each will generate sequentially using your workflow.

What's the difference between the official workflow and custom setups?

Official workflows are optimized and tested by Black Forest Labs. Custom setups offer flexibility but may need tuning.

Does ComfyUI support Klein's image editing features?

Yes, through image-to-image workflows with appropriate denoise settings.

How do I update Klein models in ComfyUI?

Replace the model file in the appropriate directory and restart ComfyUI.

Can I run Klein and SDXL in the same ComfyUI instance?

Yes, simply load the appropriate model for each workflow. Both can coexist.

Where can I find community workflows for Klein?

Check OpenArt, Civitai, and the ComfyUI subreddit for community-created workflows.

ComfyUI provides the flexibility and power to get the most out of Flux 2 Klein. Start with the basic workflow, understand how the pieces connect, and gradually explore advanced features as your needs grow.

For users who prefer simpler interfaces or lack local hardware, Apatero offers browser-based generation with multiple models including video generation and LoRA training on Pro plans.