Dataset Generator: Flux 2 Klein 9B - Quick Summary
What It Does
Automated batch image generation workflow for AI influencers using Flux 2 Klein 9B. Creates 200+ character-consistent images with automatic captioning for LoRA training or social media content.
Key Features
Built-in prompt list - Store multiple prompts directly in workflow nodes
Auto character name replacement - One trigger word replaces "Character name >" in all prompts
Batch processing - Queue hundreds of images, walk away
Auto-captioning - Saves
.txtfiles alongside images (LoRA training ready)Custom save paths - Organized output by character/project name
How to Use
Create a headshot image
Edit prompts in the orange text node (one per line, use "Character name >" placeholder)
Set character name (e.g., "P0rtia") - auto-replaces in all prompts
Configure save path for outputs
Optional: Load reference image for consistency
Queue and generate - outputs save as
image_0001.png+image_0001.txt
Requirements
Flux 2 Klein 9B model
16GB+ VRAM (24GB recommended)
Custom nodes: Prompt List, Text Replacement, Save with Captions
Best Settings
Quality: Euler, 4 steps, CFG 1.0
Resolution: 1024x1280 or 1080x1920
Output
textoutput/character_name/ ├── char_0001.png ├── char_0001.txt (caption) ├── char_0002.png ├── char_0002.txt
Perfect for: AI influencer datasets, LoRA training prep, automated social media content
Description
FAQ
Comments (14)
the version of qwen referrenced on the markdown note is incorrect and fails with a mat* error. It needs to be updated to reflect Qwen3-8B-Q8_0.gguf as the text_encoder. running now on a 5060ti 16gb. I'll report the total runtime.
I was having a problem thinking the images were not saving, but it will only save a image after ALL IMAGES are done, thats really, really stupid
any ideas what we could do in order to make it save images as it progresses?
Thank you, your process always makes my job easier.
You are a genius - is there anyway to do this but for a text to image as well?
holy shit these prompts are fantastic, i was googling around for some prompts i could use to generate a dataset just from a single image i really liked from sdxl, and this created a 1:1 perfect close up just from that blurry sub-1080p sdxl gen. Crazy times we live in.
Thanks for your positive feedback :)
Sorry, total newbie here. Workflow wouldnt work because it couldnt find the unet file. flux-2-klein-base-9b-fp8.safetensors was in the unet folder. Am i missing something here?
Hi, the workflow wouldnt work because the creator provided inaccurate information. If you would still like to attempt to use this workflow here are links to what you will actually need.:
https://huggingface.co/unsloth/FLUX.2-klein-9B-GGUF/tree/main
https://huggingface.co/Qwen/Qwen3-8B-GGUF/tree/main
https://huggingface.co/unsloth/Qwen3-8B-GGUF/tree/main
Note - You only need 1 of the Qwen text encoders. It must match the Flux 2 variant you choose to download (Q8, Q6, Q4, etc.) or the Qwen text encoder and Flux model will not be compatible. The Flux model in unet folder and Qwen in text encoder folder.
You can also change the node to make use of the safesensor (load model) instead of the gguf.
@0loser The OP could've also created and uploaded a fully functional and correct workflow from the beginning. Hope this helps!
this workflow uses gguf but specifically calls to download fp8. How stupid.