This is a very simple ComfyUI beginner friendly Kontext workflow that will work with very simple text editing instructions in natural language to build a consistent character dataset from your desired reference image so that you can build a character LORA later. This is the Flux Kontext adaptation of my previously released Character Dataset Builder Workflow for QWEN Image Edit. It does not have any lengthy process that you may have been seen for other consistent character dataset builder you may have seen using SDXL or Flux. Kontext is very good on it's own to conserve and reuse character details from reference image to provide outputs. You need a Hugging Face account to download your necessary Flux Kontext files (Details are mentioned below). Make sure you install GGUF addon for ComfyUI using ComfyUI manager and place the correct files in correct places. Also check out my other workflows for SD 1.5 + SDXL 1.0, HiDream, WAN and QWEN Image.
How to use this -
#1. Just select your desired Kontext GGUF models first
#2. now select your input character reference image and set the character's first and last name (Like "Blondie", "Nobody")
#3. on next step enter your image editing instructions - Batch character prompts like "subject seen from" followed by other details.
#4. after this select image sampling methods, CFG, steps etc. settings (this part is tricky, if you want you can just have the values as it is)
#5. finally press the run button to generate. That's it..
** I have given two reference images (1 SFW and 1 NSFW, both with their character builder prompts you can try out or modify, 60 for each) for you already in the zip file containing this workflow along with a text file with links to any downloads you may need to do for my other workflows
Enjoy!
## Flux Kontext Models
=========================================================================================================================
### Download Links for Kontext Checkpoint
https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF/resolve/main/flux1-kontext-dev-Q4_0.gguf
### Download Links for Flux Encoders
https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/clip_l.safetensors
https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/t5xxl_fp8_e4m3fn.safetensors
### Download Links for Flux VAE
https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/main/ae.safetensors
Description
FAQ
Comments (6)
Thanks for the great work. Using Qwen models locally is already impossible, so this is a good alternative at least. It took 16 minutes to generate 6 images without touching anything, using example prompts. I think it's a bit too heavy to use locally, at least for me. My graphics card is a 3080.
QWEN one does work, but only with certain tweaks.. I don't know how I did that but in one of my ComfyUI setup out of two ComfyUI inside Stability Matrix tool works with QWEN. But yeah, Kontext one works well across my Mac Machine and my Linux Machine. Use Stability Matrix to manage things (ComfyUI, WebUI Forge, Trainer etc.).. you can use shared resources across tools.. it may let you solve your QWEN Image issue.
@sarcastictofu Thank you for your reply. Actually, I can say I’ve resolved the issue—by selling my secondary monitor and graphics card, I was able to purchase a new one.
At first, the 65 GB models seemed all kinds of impossible to run localy, but after seeing the workflows shared today, I downloaded the qwen_image_edit_fp8_e4m3fn model and got it up and running. Although the official version is still quite slow, it runs impressively fast with the Lightning 4 Steps LORA. And results are really good.
You can't mix gguf with .safetensors and the workflow has both. I can't find gguf versions of the flux text encoders (clip_I and t5xxl) so I had to add DualClipLoader non gguf version. But then my system constantly goes OOM. I have a 3060 12vram and 32 ram. I reduced steps to 6. I've looked for the gguf clip_I and t5xxl and can't find them. Is there anything I can do? Am I missing something?
Well you can definitely use gguf and safetensors in the same workflow, who said you can't? I have done so in both my Mac and a system with AMD GPU.. if you are looking for the correct model to download look at the model downloads .txt file I included with the workflow on the .zip file.. and the trick to use one of the clip as gguf and another one regular safetensor is to use GGUF dual clip loader, not the default one.. make sure that you have both GGUF comfyUI plugins installed (I use ComfyUI with Stability Matrix and ComfyUI manager so it makes things bit more easier.. you can manage multiple ComfyUI, WebUI Forge instances and "port" models, LORAs etc across multiple platforms - like for me MacOS and Linux Mint, also had it on a Windows machine )