Open source Nano-banana is here!

Key Concepts:

• Quen ImageEdit 259: Alibaba's latest open-source image editor, positioned as a competitor to Nano Banana and Cadream.
• Character Consistency: The ability of an AI image editor to maintain the appearance of a subject across different scenes or poses.
• Pose Skeleton: Using a skeletal representation of a pose to control the pose of a generated character.
• Deepfakes: Generating realistic images or videos of people doing things they never did.
• Depth Map: A representation of the distance of objects in an image from the viewer.
• ComfyUI: A platform for running open-source image, video, and audio generators offline.
• VRAM: Video RAM, the memory on a graphics card.
• FP8: A compressed model format that reduces VRAM requirements.
• GGUF: A further compressed model format, allowing for use on lower-end GPUs.
• Loras: Fine-tuned models that can be added to a base model to specialize its output.

1. Introduction and Overview of Quen ImageEdit 259

• Alibaba released Quen ImageEdit 259, an open-source image editor.
• The video claims it's as good as or better than Nano Banana.
• The video will cover its features, limitations, usage, and download instructions.
• HubSpot sponsors the video.

2. Key Features and Capabilities

• Image Editing: Upload reference photos and prompt the AI to manipulate them.
  • Example: Generating wedding photos from two input photos.
• Character Consistency: Preserves character details across different backgrounds and poses.
  • Example: Placing a character in a cafe while maintaining their appearance and preserving cafe details.
• Product Placement: Generates images of models using products.
  • Example: Placing an LV bag on a Chinese actress.
• Pose Control: Uses pose skeleton photos to control the pose of the output image, effectively having ControlNet built-in.
  • Example: Uploading an image and a pose skeleton to generate the character in that pose.
• Multi-Character Scenes: Places multiple characters in a background while preserving details.
  • Example: Generating an image of two characters drinking and chatting in a detailed background.
• Outfit Merging: Combines a character, dress, and pose into one output photo.
  • Example: Merging a character with a specific dress and pose.
• Deepfake Generation: Creates realistic images of people doing things they never did.
  • Example: Uploading a photo of someone and generating an image of them doing something else.
• Text Generation: Generates accurate text in images, similar to GPT-4o.
  • Example: Generating an image of a person writing accurate text on a whiteboard.
• Photo Restoration: Restores and colorizes old photos.
  • Example: Restoring a damaged black and white photo.
• Logo Integration: Integrates logos into various scenes and products.
  • Example: Adding a logo to a bag or shirt.
• Product Advertising: Generates ads and posters for products.
  • Example: Creating an ad for a product.
• Text Editing: Edits colors and fonts of text in images.
  • Example: Changing the font of the text in an image.

3. Comparative Analysis: Quen ImageEdit vs. Cadream 4.0 vs. Nano Banana

• Quen ImageEdit is free and open-source, while Cadream and Nano Banana are paid and closed-source.
• Aerial View Test: Quen ImageEdit generates a more accurate aerial view from a satellite image compared to Cadream and Nano Banana.
• Pose Matching Test: Quen ImageEdit accurately matches a character's pose to a pose skeleton, outperforming Cadream and Nano Banana.
• Object Removal Test: Quen ImageEdit accurately removes white geese from an image, matching Cadream but surpassing Nano Banana.
• Outfit Swapping Test: Quen ImageEdit accurately swaps an outfit while preserving details, outperforming Cadream and slightly better than Nano Banana in detail preservation (buttons, sword).
• Depth Map Generation: Quen ImageEdit generates accurate depth maps, while Cadream and Nano Banana struggle.
• Fisheye View Generation: Quen ImageEdit generates a fisheye overhead view while preserving background details, outperforming Cadream and Nano Banana.
• Ad Generation Test: All three generate ads for a drink, but Quen ImageEdit is considered more balanced, while Nano Banana goes "overboard."
• Raindrop Removal Test: Quen ImageEdit and Cadream remove raindrops effectively, while Nano Banana removes people as well.
• Tourist Removal Test: Quen ImageEdit and Cadream remove tourists effectively, while Nano Banana fails to remove all humans.
• Deepfake Test (Will Smith): Quen ImageEdit generates a realistic image of Will Smith with correct text, outperforming Cadream (misspellings) and Nano Banana (different pose).
• Photo Restoration Test: Quen ImageEdit and Cadream perform well, while Nano Banana's colors are desaturated and fails to remove edge fading.
• Anime Conversion Test: Quen ImageEdit and Cadream convert an image to anime style effectively, while Nano Banana fails.
• Text Translation Test: Quen ImageEdit and Cadream perform similarly, translating some text to Chinese, while Nano Banana fails.
• Model Sheet Generation Test: Quen ImageEdit turns a 3D model into 2D, while Cadream and Nano Banana perform slightly better.

4. HubSpot Sponsorship: AI Marketing Agents

• HubSpot sponsors the video and offers a free guide: "22 ChatGPT Marketing Agents You Can Launch in 5 Minutes."
• The guide provides plug-and-play templates for ChatGPT to create AI agents.
• Examples of agents:
  • Competitive Intelligence Agent
  • ICP Reverse Engineer Agent
  • Content Calendar Creator Agent
  • Conversion Rate Optimization Agent
  • Market Trend Analysis Agent
  • Influencer Research Agent
  • Competitive Pricing Research Agent

5. Using Quen ImageEdit Online

• Quen ImageEdit can be used online through Quen's native chat interface, similar to ChatGPT.
• Users can upload images and enter prompts.
• The interface allows selecting from various Quen models, including Quen 3 Max and the ImageEdit model.
• A limited number of free generations are available daily.
• A free Hugging Face space is also available for generating images with free compute credits.

6. Downloading and Installing Quen ImageEdit Locally (ComfyUI)

• The best method is to download and run Quen ImageEdit locally for unlimited use and customization.
• ComfyUI is recommended as the platform for running open-source image generators offline.
• VRAM Requirements:
  • Original model: 40 GB VRAM required.
  • FP8 version: 24 GB VRAM recommended (20 GB model size).
  • GGUF versions: Available for lower VRAM GPUs (e.g., 7 GB model for 8 GB VRAM GPU).
• Installation Steps:
  1. Install ComfyUI.
  2. Download the desired Quen ImageEdit model (FP8 or GGUF).
  3. Place the model in the appropriate ComfyUI directory (models/diffusion_models or models/unet).
  4. In ComfyUI, refresh the model list (press R).
  5. Load the model using either:
     • Load Diffusion Model node (for FP8 version)
     • UNet Loader node (for GGUF versions) - requires installing the "comfy_gguf" custom node.
  6. Configure the workflow with appropriate CLIP and VAE models.
  7. Upload the image to edit and enter the prompt.
  8. Enable Quen Image Lightning for faster generation (set step count to 4 and CFG to 1).
• Using GGUF Models:
  • Install the "comfy_gguf" custom node.
  • Use the UNet Loader node to load the GGUF model.
  • Connect the UNet Loader to the rest of the workflow.
  • Note: GGUF models may have lower quality than the full model due to compression.

7. Adding Loras for Customization

• Loras can be added to fine-tune the model for specialized purposes.
• Use the "Lora Loader Model Only" node.
• Connect the diffusion model or GGUF model to the Lora Loader.
• String multiple Lora Loaders together, but avoid overdoing it (2-3 Loras max).

8. Conclusion

• Quen ImageEdit 259 is a free and open-source competitor to Nano Banana, often performing better.
• It can run on consumer-grade GPUs, especially with compressed models.
• The open-source nature allows for community contributions and customization through Loras.
• The video encourages viewers to try Quen ImageEdit and share their experiences and errors in the comments.
• The video promotes subscribing to the channel and newsletter for more AI news and tools.

9. Notable Quotes:

• "Well, ladies and gentlemen, the moment is finally here." (Introduction of Quen ImageEdit 259)
• "If you're a wedding photographer, I would be slightly worried if I were you." (Regarding the image editing capabilities)
• "How insane is that?" (After demonstrating pose control)
• "God bless the Alibaba team for making this open weights." (Expressing gratitude for the open-source release)