Welcome back, fellow digital artists and AI enthusiasts! Remember those mind-blowing Flux models we talked about? The ones that let you create stunning images even on a potato powered device?
Well, buckle up, because we're about to take things to the next level with an amazing image-to-image workflow!
Forget starting with just text prompts. Today, we're diving into the fascinating world of image-to-image generation using ComfyUI and Flux – a powerful combination that will unleash your inner AI Picasso.
Why Image-to-Image Matters (Even When Text-to-Image is Great)
You might be thinking, "Why bother with image-to-image when text-to-image is already so good?". Hold on! This workflow opens up a treasure chest of possibilities, giving you the power to:
Generate endless variations of an existing image, keeping the core elements intact. Want your pet dressed as an astronaut? Done. How about transforming your vacation photos into surreal landscapes? Easy peasy!
Blend different art styles seamlessly. Ever wanted to fuse a renaissance painting with a modern anime aesthetic? Now you can.
Maintain visual consistency across a series of images. Perfect for creating a cohesive look for your graphic novel, comic series, or design project.
The Secret Weapon: Image-to-Image Workflow
Image-to-image with Flux gives you a unique edge, allowing you to guide AI creativity using your own images as a starting point. Imagine using a simple sketch, a captivating photo, or even a random screenshot to generate an array of variations. You can tweak the style, inject a dose of fantastical elements, or even completely reinterpret your source image.
The real magic lies in the control you have. You can adjust how much the original image influences the variations, add artistic filters and effects, or sprinkle in your favorite Loras for that extra stylistic flair. This technique is a game changer for anyone looking to add a personal touch to their AI creations.
If you rather watch the video instead, watch it here
Ready to Dive In? Let's Set Up Our Playground!
Before we unleash our creativity, we need to make sure our tools are ready to rumble. We'll be using ComfyUI, a fantastic open-source AI image generator that offers a highly customizable and powerful workflow. To make things even easier, we'll be using the ComfyUI Manager Node, your one-stop shop for seamless installation and updates.
Here's the step-by-step breakdown:
Download the Workflow: Grab the "Flux Image-to-Image Workflow" from the Google Drive link in the description below.
Install ComfyUI Manager Node:
If you don't already have it, clone the ComfyUI Manager repo directly into your ComfyUI custom nodes folder (instructions in the description).
Restart ComfyUI. For detailed instructions, check out my previous Flux video (link in the description).
Load the Workflow: Head to the ComfyUI web page and load the downloaded workflow.
Install Missing Nodes:
If you see the dreaded "Missing Nodes" message, don't fret! The ComfyUI Manager has your back.
Go to the "Manager" tab.
Click "Install Missing Custom Nodes."
Install each node on the list.
Click "Restart" to finalize the installation and reload ComfyUI.
Download Models: It's time to equip ourselves with the models needed for this artistic adventure. Go back to the ComfyUI Manager and click "Model Manager."
Flux-Guntf Models: Search for "flux guntf" and choose either the "dev" or "schnell quantized guntf" model, based on your preference. There are several quantized versions available. For this tutorial, I'll be using the "Q4 dev" model.
Clip Models: Search for "fp8" and install the "Google TT 5fp8" clip model.
Clip L Model: Search for "clip l" and install the "clip l" model by Comfy Anonymous.
Flux VAE Model: Search for "flux vae" and install the "flux.1 vae" model.
Upscaler Model: Search for "fool" and install the "foolhardy" model for upscaling.
Lora Model: For Loras, you can easily download them from Hugging Face or CivitAI. Download the Lora of your choice and drag and drop it into the "lora" folder within your "models" folder. In this video, I'll be using the XLabs Realism Lora.
Refresh ComfyUI: Give ComfyUI a couple of refreshes to ensure all the models are loaded.
Select Models for Each Node: Click the arrows on each node to select the corresponding model and make sure everything is connected correctly.
Load Base Image & Write Prompt:
Upload your base image to the "Load Image" node.
Type in your desired prompt in the "Positive Prompt" field of the "Clip Text Encode" node.
Pro tip: Use a multimodal vision model like Chat GPT, BLIP, or even a local model, to analyze your base image and generate a detailed description for you. Then, simply copy and paste relevant parts from the model's description into your prompt field.
Adjust Settings:
Steps: In the "Basic Scheduler" node, adjust the number of steps. I usually find that 20 steps strike a good balance between quality and speed, but you can experiment and see what works best for you.
Denoise Strength: The "Denoise Strength" setting in the "K Sampler" node controls how much the original image will be altered during generation. A value of 1 means maximum noise, which can dramatically change your image. As you get closer to zero, you're adding less noise, and the generated image will stay closer to the original. Values between 0.50 and 0.91 work well for subtle variations.
Batch Size: To control how many images are generated per run, adjust the batch size in the "RepeatImageBatch" node. Instead of generating one image at a time, you can bump it up to four or more. Just keep in mind that larger batch sizes will increase generation times.
Upscale (Optional): I've added an upscaler node in the workflow to boost the resolution of the generated images to a crisp 4x. You can skip this step by directly connecting the VAE decode image node to the save image node.
Hit "Queue Prompt" & Generate!
Unlocking Your AI Artistry
Don't be afraid to play around with different prompts and settings – a little experimentation goes a long way. Try different Loras, adjust the denoise strength, tweak the batch size, and see what kind of amazing variations you can come up with.
Bonus tip: Once you find a style you absolutely love, note down the seed value displayed in the console. You can save this seed and switch the "Control After Generation" setting in the "Random Noise" node to "fixed" to replicate that exact style whenever you want!
Now go forth and create! I can't wait to see what incredible masterpieces you conjure up using this powerful image-to-image workflow. If you have any questions or need further assistance, leave a comment below, and I'll do my best to help you out. Until next time, happy creating!
🔗 Links
ComfyUI Manager Github Repo - https://github.com/ltdrdata/ComfyUI-Manager
Flux ComfyUI GGUF Workflow:
https://drive.google.com/file/d/1UL1MFsSj05wXd9jSwN-qDUcFuVmL97En/view?usp=sharing
How to Run Flux Image Models In ComfyUI with Low VRAM - https://youtu.be/VvjwTneI2xk
How To Run Flux Dev & Schnell GGUF Image Models With LoRAs Using ComfyUI - Workflow Included - https://youtu.be/nncY3dJLV78
0 Comments