Flux 2 ComfyUI: 4 Workflows (10 Images, GGUF, Low VRAM)
Takeaways
- 😀 Flux 2 allows you to use up to 10 reference images inside ComfyUI to create a single new image.
- 🛠️ Four simple workflows have been created: text-to-image, single image edit, two image mix, and multi-image mix.
- ⚙️ The required model files include Flux 2's main model, the FP8 or GGUF quantized models, a text encoder, and a VA file.
- 💾 FP8 is ideal for low VRAM GPUs, while GGUF Q4 provides a similar quality at reduced VRAM consumption.
- 📸 The text-to-image workflow in Flux 2 generates high-resolution images with just one click, taking around 1 minute per 2K image.
- 🔄 Autoprompt automatically generates prompts based on uploaded images, eliminating the need to type your own text prompts.
- 👗 The single image edit workflow allows you to make specific edits like changing clothing colors or replacing objects without altering the rest of the image.
- 👚 The two-image workflow can combine two references, such as swapping clothing designs between two people.
- 🌍 The multi-image workflow supports up to 10 reference images and allows you to create complex scenes, combining various elements from each image.
- 💡 The process is highly customizable, allowing you to set resolutions manually and control the number of images used in the final output.
Q & A
What is the main purpose of the Flux 2 ComfyUI workflows described in the script?
-The workflows help users generate images, edit images, and combine multiple references inside ComfyUI using Flux 2 models, including support for up to 10 reference images.
Why did the creator split the Flux 2 processes into four separate workflows?
-To keep things simple and avoid confusion, since having all features inside one large graph becomes difficult to manage.
What key model files are required before running any of the workflows?
-You need the Flux 2 model file (FP8, BF16, or GGUF), the text encoder file (BF16 or FP8), and the Flux 2 VAE file.
Which model option is recommended for users with low VRAM?
-The FP8 version of Flux 2 or the GGUF Q4 model, which offers quality close to FP8 while using less VRAM.
Why is updating ComfyUI to version 0.3.75 or higher important?
-Because the native nodes ‘Empty Flux 2 Latent’ and ‘Flux 2 Scheduler’ only load correctly in the updated version.
What does the autoprompt feature do in the text-to-image workflow?
-It analyzes an uploaded image and automatically generates a Flux 2–compatible JSON prompt based on its content and style.
How does the singleFlux 2 ComfyUI workflows-image edit workflow work?
-You upload one image, describe the change you want in the prompt, and Flux 2 modifies that specific detail while keeping the rest of the image unchanged.
What is the purpose of the two-image workflow?
-It combines elements from two reference images based on the prompt—for example, keeping the person from one image while transferring clothing or design elements from the other.
How does the multi-image workflow utilize up to 10 images?
-Each reference image is plugged into the workflow as a separate branch that can be enabled or disabled. Flux 2 then merges selected elements from all active references according to the prompt.
What type of results can be expected when using many reference images together?
-Flux 2 can create complex, coherent scenes that blend characters, clothing, environments, and objects from all the chosen reference images while maintaining consistent style and realism.
Outlines
- 00:00
🚀 Flux Store Launch and Workflow Overview
Flux Store is launched with a significant new feature: the ability to use up to 10 reference images in Comfy UI for generating new images. The video introduces Flux 2 workflows, including autoprompt for automating prompt creation, and discusses how to run Flux 2 efficiently even with low VRAM on your GPU. Various model settings are explored, including FP8, Q4 GGUF, and how to set up these models for optimal performance based on GPU capabilities. The setup process for required model files and their installation in Comfy UI is also covered.
- 05:00
📸 Testing Flux 2 with Text-to-Image Workflow
This section demonstrates using Flux 2's text-to-image workflow with autoprompt and manual text prompts. A test image of a fashion designer is generated in 2K resolution with FP8, showcasing the workflow’s ability to produce high-quality results quickly. The user also explores how to use autoprompt to generate a JSON-style prompt based on a reference image, testing this with multiple images, such as a photo of an old couple and a muscular hand holding a soda can.
- 10:02
🔍 Comparing FP8 and GGUF Q4 Models for Image Quality
Here, the quality of the Q4 GGUF model is compared to FP8, demonstrating that the results are very similar in terms of detail, color, and composition. This section is particularly useful for users with limited GPU VRAMFlux Store workflows overview, as the Q4 GGUF model provides nearly identical results to FP8, making it a viable option for those with lower VRAM without compromising on image quality.
- 15:02
👚 Editing Images with Flux 2 Workflow
The focus shifts to the image-editing workflow, where users can upload a single image and specify prompt instructions to modify it. The example demonstrates changing the color of a woman's t-shirt from black to green while keeping the rest of the image intact. This shows the power of Flux 2's image editing capabilities, allowing for subtle yet powerful edits like clothing color changes, small object replacements, or lighting adjustments.
👕 Creating New Images by Combining Two References
This section introduces a two-image workflow where Flux 2 combines elements from two reference images based on a text prompt. The test involves using one image of a woman and another image with a t-shirt design, resulting in the woman wearing the t-shirt from the second image. This workflow is particularly useful for transferring designs, logos, or clothing styles from one image to another without distorting the original features.
🎨 Multi-Reference Workflow for Complex Image Creation
In this final workflow, Flux 2 supports up to 10 reference images to create a single composite image. The user demonstrates this by combining multiple reference images, such as characters, clothing, and environments, into one cohesive scene. The test images involve four characters, and the resolution is manually set to match the original images. Further testing with six reference images shows how Flux 2 can create highly detailed and realistic composite images by blending elements from different sources seamlessly.
Mindmap
Keywords
💡FluxJSON code correction 2
Flux 2 is a powerful generative AI model used for creating and editing high-resolution images based on given prompts and reference images. It allows users to generate clean 2K images in one click without additional upscaling or Loras. In the video, Flux 2 is demonstrated through various workflows like text-to-image generation and multi-image workflows, showing its flexibility and efficiency in image creation.
💡ComfyUI
ComfyUI is the user interface (UI) used to interact with Flux 2, facilitating the creation, management, and modification of image workflows. In the video, the speaker discusses setting up and updating ComfyUI to the latest version (0.3.75) to ensure compatibility with the Flux 2 model. The UI enables users to easily apply different reference images, choose models, and tweak settings like resolution and VRAM usage.
💡GGUF
GGUF (Generic Graph Universal Format) refers to a quantized model format designed to be lighter on GPU memory usage while still maintaining good image quality. In the video, GGUF is used as an alternative to the full Flux 2 model for users with low VRAM. Specifically, the Q4 GGUF model was found to offer quality comparable to the FP8 model, allowing users to generate high-quality images even with lower VRAM GPUs.
💡FP8
FP8 refers to a floating-point format with 8 bits per value, used to optimize memory usage for AI models. In the video, Flux 2's FP8 model is recommended for users with low VRAM as it reduces the model size and memory usage, making it more practical for devices with limited GPU resources. The video demonstrates generating clean, high-resolution images using FP8 with a resolution of 2048x1536.
💡Autoprompt
null
💡Text Encoder
A text encoder is a neural network model responsible for converting input text prompts into a format that can be understood by the generative AI model. In Flux 2, the Mistral text encoder is used, with two versions available: BF16 for higher VRAM users and FP8 for low VRAM users. The video explains the importance of the text encoder in ensuring that Flux 2 correctly interprets and generates the desired images based on prompts.
💡VA File
The VA file (possibly referring to a 'Variational Autoencoder' file) is required to run Flux 2. It contains model data that helps in the generation of images based on the input provided. The speaker in the video highlights the need to place this VA file in the appropriate directory within ComfyUI's model folder for proper workflow execution.
💡Single Image Edit Workflow
The Single Image Edit Workflow in ComfyUI allows users to modify a single image by keeping most of its features intact while changing specific elements, like clothing color or objects. In the video, an example is shown where the color of a woman's t-shirt is changed from black to green, demonstrating the capability of Flux 2 to perform precise edits without disturbing the overall integrity of the image.
💡Two Image Workflow
The Two Image Workflow combines two reference images into a single output. Flux 2 uses these references to generate a new image based on a prompt that mixes elements from both sources. In the video, the example involves a woman from one image and a t-shirt design from another, with the model merging the face and pose from one image with the t-shirt design from the other, creating a coherent and realistic final result.
💡Multi-Image Workflow
The Multi-Image Workflow is a powerful feature in Flux 2 that allows the use of up to 10 reference images in generating a single final output. The video demonstrates how users can specify different reference images for various elements of a scene, such as clothing, backgrounds, and objects. With this setup, complex scenes can be created by combining diverse visual elements, such as characters, clothing, and accessories, all matching the style and composition of the reference images.
Highlights
Flux Store now allows the use of up to 10 reference images in ComfyUI for generating a new image.
The tutorial demonstrates how to use all 10 images in Flux 2, along with autoprompt to generate images without typing a prompt.
A guide on how to run Flux 2 even with GPUs that have low VRAM by using FP8 and Q4 GGUF models.
The video introduces four different Flux 2 workflowsJSON code correction: text-to-image, single image edit, two-image merge, and multi-image reference workflows.
Flux 2 models, including the FP8 and GGUF versions, are shown to be optimized for low VRAM usage, with Q4 GGUF performing similarly to FP8.
ComfyUI must be updated to version 0.3.75 or later for Flux 2 workflows to work correctly, including new Flux 2 nodes like 'flux to latent' and 'flux to scheduler'.
The text-to-image workflow allows generation of high-quality 2K images in just one click, with no upscaling or extra steps.
Autoprompt functionality automatically generates a JSON-style prompt from an uploaded reference image for accurate and efficient image creation.
Using autoprompt, users can upload images to generate similar outputs without typing prompts themselves.
The quality comparison between FP8 and Q4 GGUF models shows that Q4 provides near-identical results to FP8 for users with limited VRAM.
Single image editing in Flux 2 allows for detailed changes like altering clothing color while maintaining the original image structure.
The two-image workflow merges elements from two reference images, such as combining a person's face with different clothing from a second image.
In the two-image workflow, Flux 2 can accurately replicate text from a reference t-shirt, preserving details like font style and readability.
The multi-reference setup can handle up to 10 reference images, allowing for complex scenes combining elements like people, backgrounds, clothing, and objects.
The multi-image workflow is demonstrated with up to 6 references, showing the ability to create complex scenes with high precision.