Flux Kontext Pictures can be modified through natural language, which can not only make localized and precise modifications to pictures while maintaining character consistency and scene consistency, but can also make continuous modifications.
It can be said that Kontext is the strongest multimodal model available, even GPT4o is not far behind it.
Most importantly, Flux Kontext is now available in a free and open source dev version, with unbeatable features and free of charge, who can't love it?
Kontext Application Scenario Demo
Open comfyUI and open the official flux example from the templateWorkflowThe first one is the same as the second one. Two sample workflows are provided here that are essentially the same, so let's open the first one for demonstration.
Writing prompts using the official "prompt technique"
When we open the template, we are given an official tip on how to use the prompt word, so we can put it to good use.
Copy this cue word trick and come to the big language models like beanbag, deepseek, and passyi and tell it that these are flux kontext cue words, and I want you to learn how to use it, and if you get it reply to understand that I'll use it.
I gave the cue word to the Tongyi Big Model and it easily understood the use of the cue word.
Later, we can have Tongyi write the kontext prompts for me.
1. Remove the watermark:
Tell tomyi to write me a prompt based on the flux kontext prompt word rule: Remove the watermark from the screen.
Paste the prompt words given by Tongyi into the CLIP text encoder in the Kontext workflow, upload an image full of watermarks, click run, and it removes the watermarks from the image perfectly.
2. Extraction of elements:
Tell Tongyi to extract the robot character with the gun on the right, delete the rest, and set the background to white.
Tongyi not only helped me write the prompts, but also told me why the prompts were written the way they were.
Pasting the cue words into the text encoder and uploading a picture of the two robots, Kontext accurately extracted the red robot on the right.
3. Role reversal perspective
Have Tongyi write a prompt that says, "Rotate the camera 90 degrees clockwise to show a side view of the rabbit."
After pasting the prompt into the text encoder and also uploading a picture of the front of the rabbit and clicking run, we get a picture of the side of the rabbit.
4. Successive revisions:
Click "refresh" in the Load Image node to load the just-generated rabbit side image, so that we can continuously modify the just-generated image as long as we give a new cue word.
This time I asked Tongyi to help me write the prompt: the main body of the picture remains the same, the background becomes the prairie
In this way we have accomplished both the continuous modification of the image and the realization that the main body remains unchanged and only the background of the image is modified.
5. Style reference:
I told Tongyi that I will upload a style image first, please help me write a prompt word according to the flux kontext prompt word rule: refer to the style of the image I uploaded to generate an image of a beautiful warrior in armor with an exceptionally good figure, the female warrior is holding a sword with both hands and is making an attacking movement.
Upload a special styled image and paste the cue word just generated by Tongyi
Click Generate to see the result, which perfectly replicates the style of the reference image.
6、Double map merger
flux kontext has an image loading node disabled by default, turn it on and upload two images at the same time, a woman and a goose.
Write a cue word with a generic meaning: help me write a cue word based on the flux kontext cue word rule: let the woman carry this goose in her arms
Essentially kontext generates a picture that is also drawing cards, and the effect is that if you are not satisfied with the generated picture, you can draw a few more until you are satisfied.
Zoom in on the picture to observe that kontext extends the height of the original picture and then lets the woman hold the goose, and overall the consistency of the woman and the goose is still relatively high. However, kontext's support for dual-image is not particularly good, and the consistency is not very high when multiple images are generated, so it is recommended that we still focus on single-image.
7. Remove background
Writing Prompt Words with Generalization: Removing the Context
With the cue word, it was easy to remove the background of the picture other than the characters.
We could also go a step further and have kontext set the background to white after removing it.
8, style to draw
Have Tongyi write a prompt: help me write a prompt based on the flux kontext prompt rules: change the style to Ghibli style
Modifying the image style is a piece of cake for kontext, you can also try to make it modify to other types of image styles.
9. Local fine-tuning
If you want to modify only a part of the picture, really "where to hit", we need to modify the workflow a little, but also use the mask function, see the article for details:
Flux Kontext guide, two methods to achieve accurate local redrawing of AI images
Write a cue with a generic meaning: help me write a cue based on the flux kontext cue rule: change the red area to a pink wreath
Drawing a mask on the character part of the picture, Kontext is very precise in making the character wear a wreath on her head.
That's all for today.