The best free AI image generator is here!

Key Concepts

• Zimage: A new, highly efficient, and powerful open-source image generation model.
• Zimage Turbo: The text-to-image generation component of Zimage.
• Zimage Edit: The image editing component of Zimage (not yet released).
• World Understanding: Zimage's ability to comprehend and accurately represent real-world entities, celebrities, fictional characters, and locations.
• Efficiency: Zimage's small model size (6 billion parameters) and fast generation times, even on low VRAM.
• Quantized Models (GGUF): Compressed versions of Zimage that allow it to run on GPUs with as little as 4GB of VRAM.
• ComfyUI: A popular platform for running open-source image, video, and audio generators locally, used for demonstrating Zimage.
• Image-to-Image: Using an existing image as a reference to generate a new, modified image.
• LoRAs (Low-Rank Adaptation): Smaller, fine-tuned models that can be added to Zimage to achieve specific styles, characters, or effects.
• VRAM (Video Random Access Memory): The memory on a graphics card, crucial for running AI models.
• Parameters: The variables within a machine learning model that are learned during training.
• CFG (Classifier-Free Guidance): A parameter that controls how strictly the AI follows the prompt.
• Denoising Strength: In image-to-image, this parameter controls how much of the original image is retained or altered.

Zimage: The New Leading Open-Source Image Generator

This video introduces Zimage, an open-source image generation model that is presented as the current best option due to its exceptional world understanding, ability to generate highly realistic photos, efficiency, and uncensored capabilities. The presenter aims to demonstrate its features, compare it with other leading models like Flux 2 and Quen image, and provide a comprehensive guide on how to download and run it.

Zimage Turbo: Text-to-Image Generation Capabilities

Zimage Turbo, the text-to-image component, is highlighted for its ability to produce incredibly realistic images, including portraits, scenery, and even text within images.

• Realism and Resolution: Zimage can generate images in 2K, 3K, and up to 4K resolution, with impressive detail in skin textures, hands, and fingers. The presenter notes a slight quality degradation at 4K and beyond.
• Text Rendering: It excels at rendering text in both English and Chinese, making it suitable for creating posters with specific fonts and text styles.
• World Understanding: A key strength is its deep understanding of hundreds of celebrities, existing characters, and global locations. Examples shown include realistic depictions of the Great Wall of China and other landmarks.
• Artistic Styles: The model can handle diverse design and art styles, making it versatile for social media posts, marketing materials, and posters.
• Uncensored Content: Zimage is noted for its ability to generate uncensored content out-of-the-box, though explicit examples cannot be shown on YouTube.

Zimage Edit: Image Editing Capabilities (Future Release)

While not yet released, Zimage Edit is described as an image editor that uses natural language to modify existing images. Potential applications include changing hair color, clothing, text, expressions, and even entire scenes. The presenter emphasizes its ability to preserve original image details.

Model Efficiency and Technical Specifications

• Model Size: Zimage has a relatively small size of 6 billion parameters.
• VRAM Requirements: The full model can fit within 16GB of VRAM, and quantized versions are available for as low as 4GB of VRAM.
• Generation Speed: Images can be generated in seconds, with the presenter achieving results in under 10 seconds on their setup.

Comparison with Leading Open-Source Models

Zimage Turbo is directly compared against Flux 2 Dev and Quen image across various challenging prompts.

• Celebrity and Character Generation: Zimage accurately generated multiple celebrities (Anne Hathaway, Jackie Chan, Messi) and fictional characters (Miku, Nezuko, Gojo, Sasuke) with correct likenesses and outfits, outperforming both Flux 2 Dev and Quen image, which struggled with accuracy and consistency.
• Complex Prompts and Text Rendering: Zimage successfully rendered a complex bakery scene with specific text and pastries, while Flux 2 Dev and Quen image had issues with realism and accuracy of the pastries.
• Character Design Sheets: Zimage produced consistent character model sheets across front, side, and back views, whereas Flux 2 Dev and Quen image had inconsistencies and anatomical errors.
• Poster Generation: Zimage created a well-designed holiday cookie swap poster with correct text and logos, while Flux 2 Dev had text errors and Quen image was considered a tie with Zimage in terms of aesthetic appeal.
• Expressions and Hand Gestures: Zimage accurately depicted a woman winking and making a peace sign, outperforming the other models which had awkward or overly cinematic results.
• Wildlife and Tricky Scenes: Zimage generated a realistic fisheye photo of a sloth and accurately depicted a complex scene with a ballerina, rabbit, and elephant, demonstrating superior realism and detail compared to the others.
• Human Anatomy: Zimage produced realistic and anatomically correct depictions of human palms and soles, and crucially, a woman doing a handstand with correct leg and foot positioning, a feat none of the other models achieved.
• Art Styles: Zimage excelled at generating a minimalist Chinese watercolor painting, capturing the brush strokes and artistic style better than Flux 2 Dev and Quen image.
• Complex Scenes: Zimage generated a detailed and coherent busy marketplace scene with intact people, while Flux 2 Dev and Quen image showed more artifacts and less realism.
• Long Text Generation: While not perfect, Zimage produced the most accurate long handwritten text, with only a minor repetition, whereas Flux 2 Dev and Quen image had significant errors and gibberish.
• Manga Generation: All models struggled with generating a manga page with multiple panels, but Zimage was considered the least flawed, despite some gibberish text and minor inaccuracies.

Overall Conclusion of Comparison: Zimage consistently outperformed Flux 2 Dev and Quen image across a wide range of challenging prompts, particularly in realism, accuracy, world understanding, and handling complex details.

Using Zimage Turbo: Installation and Workflow

The video provides detailed instructions on how to use Zimage Turbo.

• Online Demo: A free Hugging Face demo is available for quick testing with daily credits.
• Local Installation with ComfyUI:
  • Prerequisites: ComfyUI needs to be installed.
  • Workflow Download: An official Zimage Turbo workflow JSON file is provided.
  • Model Downloads: Users need to download three specific models:
    • Quen 34B (text encoder, ~7.5 GB)
    • Zimage Turbo model (~11.4 GB)
    • VAE.safe tensor (~300 MB)
  • Workflow Setup: The downloaded JSON file is loaded into ComfyUI, and the downloaded models are selected in the respective nodes.
  • Prompting and Parameters: Users input their desired prompt. Key parameters discussed include:
    • Resolution: Up to 2K/3K/4K, with quality degradation beyond 2048x2048.
    • Shift Value: Affects contrast and detail (default 3, lower for higher contrast, higher for softer images).
    • Steps: Number of steps for generation (sweet spot 8-10 for quality).
    • CFG: Controls prompt adherence (Zimage Turbo works best at CFG 1; higher values can saturate colors and improve text accuracy but may distort images). Negative prompts are generally ignored at CFG 1.
    • Sampler: Default settings are recommended, but variations exist for different outcomes.
  • Generation: The process is initiated by clicking "Run," and generated images are saved automatically.

Running Zimage with Low VRAM (Quantized Models)

For users with limited VRAM (e.g., 4GB), quantized GGUF versions of Zimage Turbo are available.

• Model Sources: A link to a repository with various compressed models is provided.
• Workflow: A separate GGUF workflow JSON file is used.
• Model Downloads: Users need to download:
  • A GGUF model (e.g., Q4 medium GGUF, ~5 GB).
  • A quantized text encoder (e.g., Q4 medium, ~2.5 GB).
• Custom Nodes: The ComfyUI GGUF custom node needs to be installed.
• Parameters: Similar prompting and resolution settings apply, with the ability to generate high-resolution images even with compressed models.

Image-to-Image Editing with Zimage Turbo

While Zimage Edit is not yet released, Zimage Turbo can be adapted for basic image-to-image tasks.

• Workflow Modification: The latent image node in the ComfyUI workflow is replaced with a "Load Image" node.
• VAE Encoding: The uploaded image is encoded into latent space using a VAE encode node.
• Prompting: A descriptive prompt is used to guide the transformation of the input image.
• Denoising Strength: This crucial parameter controls how much the original image influences the output. A lower value retains more of the original, while a higher value allows for more significant changes. The presenter demonstrates converting a 3D animation into a realistic photo by adjusting this value.
• Applications: This allows for retouching existing images, converting between styles (e.g., anime to realistic), or enhancing previous generations.

Adding LoRAs for Customization

LoRAs are smaller, fine-tuned models that can be added to Zimage to achieve specific results.

• Functionality: LoRAs can generate specific art styles, characters, poses, or effects.
• Installation: Downloaded LoRA files are placed in the ComfyUI/models/loras folder.
• Workflow Integration: A "Lora Loader" node is inserted into the ComfyUI workflow, placed between the diffusion model loading and the sampler.
• Strength and Trigger Words: The strength of the LoRA's influence can be adjusted. Many LoRAs require specific "trigger words" in the prompt to activate their effect.
• Multiple LoRAs: Multiple LoRA nodes can be chained together for combined effects.

Conclusion and Future Outlook

Zimage is presented as a groundbreaking open-source image generation model that significantly surpasses its predecessors in realism, world understanding, and efficiency. Its accessibility, even on low-end hardware, and its potential for uncensored content make it a highly valuable tool for creators. The upcoming release of Zimage Edit is anticipated to further enhance its capabilities. The presenter encourages users to experiment with the model and share their creations.

Sponsor Mention: Chat LLM by Abacus AI is highlighted as an all-in-one platform for accessing various AI models, including image and video generators, for a monthly subscription fee.