FLUX.2 Klein 9B - Reference-Based Face Swap Workflow

⚡ Production-tested workflow.

This workflow originated from an internal automation pipeline capable of processing hundreds of images automatically through the ComfyUI API on a single NVIDIA T4 GPU.

A highly consistent face swap workflow for ComfyUI built around FLUX.2 Klein 9B, latent references, and a dedicated face swap LoRA.

Unlike traditional face swap methods, this workflow uses both a Target Reference and a Source Face Reference to preserve composition while transferring identity.

Requirements

Before loading the workflow, make sure you have:

• FLUX.2 Klein 9B

• FLUX2 VAE

• Qwen 3 8B FP8 Mixed Text Encoder

• Face Swap LoRA

• LanPaint Custom Nodes

Required Models

FLUX.2 Klein 9B

Place inside:

ComfyUI/models/unet/

Model:

flux-2-klein-9b.safetensors

FLUX2 VAE

Place inside:

ComfyUI/models/vae/

Model:

flux2-vae.safetensors

Qwen 3 8B FP8 Mixed

Place inside:

ComfyUI/models/text_encoders/

Model:

qwen_3_8b_fp8mixed.safetensors

Face Swap LoRA

Place inside:

ComfyUI/models/loras/

Model:

bfs_head_v1_flux-klein_9b_step3500_rank128.safetensors

Required Custom Nodes

LanPaint

This workflow uses the LanPaint sampler.

Install it through:

ComfyUI Manager → Install Missing Custom Nodes

Workflow Overview

This workflow uses two different references:

Target Reference

Preserves:

• Composition

• Pose

• Framing

• Perspective

• Scene structure

• Lighting context

Source Face Reference

Transfers:

• Face identity

• Facial proportions

• Hair characteristics

• Eyes

• Skin details

By combining both references, the workflow achieves significantly better consistency than traditional face swap approaches.

How To Use

Step 1 - Load The Source Face

Node:

Source Face Image

This image provides the identity that will be transferred.

Recommended:

✅ Front-facing portraits

✅ High-resolution images

✅ Clear facial visibility

✅ Good lighting

Avoid:

❌ Motion blur

❌ Heavy occlusions

❌ Sunglasses

❌ Extreme angles

Step 2 - Load The Target Image

Node:

Target Image

This image provides:

• Background

• Body

• Clothing

• Pose

• Composition

• Framing

• Perspective

The workflow attempts to preserve these elements while replacing the identity.

Recommended:

✅ Portrait photos

✅ Medium shots

✅ Visible face

✅ Clear lighting

Avoid:

❌ Tiny faces

❌ Hidden faces

❌ Extreme side profiles

Step 3 - Queue Prompt

Press:

Queue Prompt

The workflow automatically:

1. Reads target dimensions

2. Creates target latent references

3. Creates source face latent references

4. Applies dual-reference conditioning

5. Generates the face-swapped result

6. Preserves original scene composition

No additional setup is required.

What This Workflow Preserves

• Original background

• Original composition

• Original lighting

• Original perspective

• Original pose

• Original framing

• Original camera distance

• Original scene layout

Ideal Use Cases

• UGC Creator Replacement

• Influencer Face Swaps

• Character Consistency

• Marketing Creatives

• AI Avatars

• Personal Branding

• Dataset Generation

• Batch Face Swap Pipelines

Performance

Tested on:

• NVIDIA T4

• RTX 3060

• RTX 4070

• RTX 4090

VRAM requirements will vary depending on image resolution.

Batch Automation

This public release contains the single-image workflow.

Internally, I use a custom Python automation system connected directly to the ComfyUI API that can process large batches automatically.

Example setup:

• 1 source face

• 400+ target images

• Automatic queue management

• Automatic retries

• Organized output folders

• GPU-friendly processing

• Designed for T4 cloud instances

This automation system is not included in the public release.

Commercial Applications

This workflow can be used for:

• UGC generation

• Marketing campaigns

• Creator replacement

• Character consistency projects

• AI influencer pipelines

• Content automation

Need Batch Processing?

I also develop custom automation systems built around ComfyUI.

Examples:

• Face swap automation

• Folder-based processing

• ComfyUI API integrations

• Large-scale image generation pipelines

• Dataset creation workflows

• Custom workflow development

If you're interested in automation or custom solutions, feel free to send me a direct message on Civitai.

Credits

Built and tested using FLUX.2 Klein 9B, latent reference conditioning, and a dedicated face swap LoRA for high-consistency identity transfer.

If you find this workflow useful, consider leaving a rating and sharing your results.