This is BY FAR the best AI image generator. Free & offline

Key Concepts

• Quen Image: An open-source AI image model by Alibaba, excelling in text generation and prompt understanding.
• Flux Create Dev: Another open-source image model, known for photorealistic image generation.
• GPT-4o: A proprietary, closed-source image generator, considered one of the best overall.
• ComfyUI: A node-based interface for Stable Diffusion, used for running and customizing AI image generation workflows.
• VRAM: Video RAM, the memory on a graphics card, crucial for running AI models.
• Quantization: A technique to compress AI models, reducing VRAM requirements.
• GGUF: A file format for quantized models, allowing them to run on lower VRAM.
• Text Encoder: A component of the AI model that processes text prompts.
• VAE (Variational Autoencoder): A component responsible for encoding and decoding images.
• K Sampler: A ComfyUI node that performs the iterative denoising process to generate images.
• CFG Scale: A parameter that controls how closely the AI follows the prompt.
• Steps: The number of iterations the AI performs during image generation.
• Seed: A number that determines the initial state of the random number generator, influencing the generated image.

Quen Image: The Best Open-Source AI Image Model

The video introduces Quen Image, an open-source AI image model developed by Alibaba, which the presenter claims is currently the best available, surpassing even recently released models like Flux Create Dev. The model's strengths lie in its ability to generate text within images accurately and its superior prompt understanding.

Key Features and Capabilities

• Text Generation: Quen Image excels at generating text within images, accurately reproducing specified fonts, styles, and content. Examples include generating bookstore window displays with correct signage and book titles, creating PowerPoint slides with accurate text and icons, and producing images with Chinese text.
• Image Editing: Quen Image can edit existing images, allowing for style transfer (e.g., turning a Pikachu into Ghibli style), object removal, adding elements (e.g., adding a hat and sunglasses), and changing perspectives.
• Outpainting: The model can extend existing images beyond their original boundaries, creating a zoomed-out effect.
• Pose and Perspective Manipulation: Quen Image can alter the pose and perspective of subjects within an image while preserving their facial features.
• Art Style Preservation: The model can maintain the art style of an original image while making modifications.

Comparison with Other Models

The presenter conducts a series of tests comparing Quen Image with Flux Create Dev (a photorealistic open-source model) and GPT-4o (a top-tier proprietary model). The tests involve complex prompts with specific text, objects, and styles.

• Bakery Window Example: Quen Image accurately generates a bakery window display with a wooden sign ("freshly baked today" in Pacifico font), a chalkboard listing prices, a baking workshop poster, and specific pastries. Flux Create Dev struggles with text accuracy, while GPT-4o produces a more cartoony image.
• Urban Street View Example: Quen Image accurately depicts an urban street view featuring Bella's Cafe (with a neon sign), Golden Lotus Chinese Restaurant (with lanterns and a dragon design), and Urban Boutique (with a mannequin and geometric patterns). Flux Create Dev mixes up elements, while GPT-4o produces a good result but with a cartoony aesthetic.
• Bali Beach Example: Quen Image accurately generates a tropical beach scene in Bali with a woman doing yoga, a monkey stealing a coconut, a surfboard with a hibiscus, and a Bali sunset sign. Flux Create Dev lacks specific details, while GPT-4o produces a cartoony image.
• PowerPoint Slide Example: Quen Image creates a visually appealing PowerPoint slide with accurate text and icons, surpassing Flux Create Dev and GPT-4o in visual quality.
• UI Design Example: Quen Image generates a mobile fitness app UI with correct text, icons, and progress bar colors, outperforming Flux Create Dev and GPT-4o.
• YouTube Search Results Page Example: GPT-4o excels in generating a YouTube search results page with accurate text and thumbnails, while Quen Image has some text errors.
• Pokemon Card Example: GPT-4o generates a more accurate Baby Yoda Pokemon card with correct stats and energy logos, while Quen Image has text errors.
• Realistic Photo Examples: All three models perform well in generating realistic photos, with Flux Create Dev being known for the most natural-looking results.
• Anatomy Test (Feet): Quen Image accurately generates a photo of a woman showing her palms and soles of her feet, while Flux Create Dev has anatomical errors and GPT-4o produces a cartoony image.
• Anatomy Test (Hands): Quen Image and GPT-4o both generate five hands making a star shape, while Flux Create Dev fails.
• Anime Example: Quen Image excels at generating anime-style images, accurately depicting characters like Naruto, Nezuko, Goku, and Doraemon eating at McDonald's.
• 3D Pixar Style Example: Quen Image accurately generates a busy marketplace in 3D Pixar style, while Flux Create Dev's style is off and GPT-4o struggles to produce a 3D animation style.
• Art Style Example (Monet): None of the models perfectly replicate a Monet-style impressionist painting, but Quen Image and GPT-4o come close.
• Uncommon Species Example (Tarsiers): Quen Image and GPT-4o generate creatures that resemble tarsiers, while Flux Create Dev fails to understand the prompt.

Conclusion from Comparisons

Quen Image consistently outperforms Flux Create Dev and often matches or surpasses GPT-4o in generating images with complex prompts, accurate text, and specific styles. It excels in anime and 3D Pixar styles.

Using Quen Image

Online Options

• Hugging Face Space: A free online platform for trying Quen Image, but with limited free credits.
• Quen Chat: A free platform by Alibaba for using their AI models, including image generation, but it's uncertain if it uses the latest Quen Image model.

Local Installation with ComfyUI

The video provides a step-by-step guide to installing and running Quen Image locally using ComfyUI.

1. Install ComfyUI: The video assumes the user already has ComfyUI installed.
2. Download Models: Download the diffusion model, text encoder file, and VAE file from the provided Hugging Face links. The official models require 24 GB of VRAM.
3. Place Files in Correct Folders: Place the downloaded files in the appropriate ComfyUI folders (models/diffusion_models, models/text_encoders, models/VAE).
4. Update ComfyUI: Open ComfyUI, click on "Manager," and then "Update ComfyUI." Restart ComfyUI after the update.
5. Download Workflow: Download the official Quen Image workflow file from the provided link.
6. Load Workflow: Drag and drop the downloaded workflow file onto the ComfyUI interface.
7. Select Models: In the ComfyUI workflow, select the downloaded models in the dropdown menus for each corresponding node.
8. Configure Settings: Adjust the width, height, batch size, steps, CFG scale, and other parameters as desired.
9. Run Workflow: Click "Run" to generate the image.

Running on Low VRAM

For users with less than 24 GB of VRAM, the video provides instructions on using quantized models (GGUF) that require as little as 8 GB of VRAM.

1. Download Quantized Model: Download a GGUF model from the provided Hugging Face repository by City96, choosing a version appropriate for your VRAM.
2. Download GGUF Optimized Text Encoder (Optional): If you still encounter out of memory errors, download a GGUF optimized text encoder from the same repository.
3. Download GGUF Workflow: Download the GGUF workflow file.
4. Load Workflow: Drag and drop the downloaded workflow file onto the ComfyUI interface.
5. Install Missing Custom Nodes: If the GGUF node is red, install the "comfy_gguf" custom node using the ComfyUI Manager.
6. Select Models: In the ComfyUI workflow, select the downloaded GGUF model and text encoder.
7. Configure Settings: Adjust the width, height, batch size, steps, and other parameters as desired.
8. Run Workflow: Click "Run" to generate the image.

Image Editing

The presenter notes that the image editing capabilities showcased by Alibaba are not yet available in the open-sourced model. The editing model is planned for a future release.

Conclusion

Quen Image is a groundbreaking open-source AI image model that excels in text generation, prompt understanding, and various image generation tasks. It often outperforms other open-source models and even competes with top-tier proprietary models like GPT-4o. The video provides a comprehensive guide to installing and using Quen Image, including instructions for running it on low VRAM systems. The presenter encourages viewers to experiment with the model and share their creations.