IC-Custom is an open-source AI image customization model that transforms your photos by intelligently blending reference images with target backgrounds or generating completely new scenes while preserving specific visual elements.
Developed by TencentARC, this model uses advanced in-context learning to handle both precise object placement and creative image generation tasks.
The model addresses two major challenges in AI image editing: maintaining object identity while changing contexts and providing fine-grained control over where elements appear in the final composition.
Unlike simple AI photo editing tools, IC-Custom understands visual relationships between different image elements and can adapt objects to new environments while preserving their essential characteristics.
Features
- Position-aware customization: Place objects exactly where you want them using masks or drawings to specify locations.
- Position-free generation: Create new images based on text descriptions while maintaining reference object identity.
- Dual masking modes: Choose between precise automated masking (SAM) or manual brush-based mask creation.
- VLM-powered prompts: Auto-generate or polish text prompts using built-in vision-language model assistance.
- Mask refinement tools: Adjust masks with dilate, erode, and bounding box operations for better results.
Use Cases
- Product Placement: A marketer can use the position-aware mode to place a new product, like a handbag or a bottle of perfume, into various lifestyle images to create promotional content. This saves the cost and effort of organizing photoshoots.
- Virtual Try-On: Fashion brands can leverage the position-aware feature to allow customers to virtually try on clothing. By uploading a photo of a garment and a model, they can generate realistic images of the clothing being worn.
- IP Customization and Creation: An artist or designer can use the position-free mode to create a series of images featuring a consistent character or object in different scenarios. This is ideal for storytelling, branding, or developing intellectual property.
- Creative Content Generation: Anyone can use the position-free mode to generate creative and unique images. For instance, you could take a picture of your pet and create a new image of them in a fantastical setting described in your text prompt.
How to Use It
Getting started with IC-Custom requires choosing between two main workflows, depending on your specific needs.
Position-aware mode
Upload your reference image containing the object you want to customize. Next, upload your target background image where you want to place the object.
Choose your masking approach: either click areas for automatic SAM masking or use the brush tool to manually draw where the object should appear.
Add an optional text prompt to guide the generation process, though the VLM buttons can auto-generate suitable prompts if you prefer.
Review your mask using the preview tools and refine it with the dilate, erode, or bbox operations if needed.
Finally, click Run to generate your customized image.
Position-free mode
Position-free mode follows a simpler workflow. Upload your reference image containing the subject you want to preserve.
Write a detailed text prompt describing the target scene or use the VLM buttons for automatic prompt generation and polishing.
The AI will generate new images featuring your reference subject in the described context.
This mode works particularly well for creating multiple variations of the same subject in different scenarios.
Tips
Both modes benefit from experimenting with different seeds to generate varied outputs.
The Alt + “-” or ⌘ + “-” keyboard shortcuts help zoom out for better operation visibility when working with larger images.
Advanced options provide additional control over generation parameters for users who want to fine-tune their results.
Pros
- No training required: Start using the tool immediately without model training or complex setup procedures.
- Dual functionality: Handles both precise placement tasks and creative generation in a single interface.
- Quality preservation: Maintains object identity and visual characteristics when adapting to new contexts.
- Intuitive masking: Offers both automatic and manual masking options to suit different user preferences and skill levels.
- Open source availability: Full code and model weights are available for developers who want to integrate or modify the system.
- Regular updates: Active development with planned improvements for speed and instruction following capabilities.
Cons
- Limited object categories: Currently performs best with bags, perfumes, clothing, and rigid objects rather than complex organic subjects.
- Processing speed: Inference times can be longer compared to simpler image editing tools, particularly for high-resolution outputs.
- Face customization gaps: Face-related data and style transfer capabilities are still being developed and refined.
- Resolution constraints: Optimal performance requires images of at least 800px resolution, which may limit mobile photography usage.
- Position-free limitations: The mask-free customization mode is still being improved and may not match the precision of position-aware workflows.
Related Resources
- GitHub Repository: Access the complete source code, installation instructions, and development updates.
- Research Paper: Read the technical details about the in-context learning approach and architectural innovations.
- ComfyUI Integration: Use IC-Custom within ComfyUI workflows for advanced automation and integration.
- Model Checkpoints: Download pre-trained models for local deployment and customization projects.
FAQs
Q: What types of images work best with IC-Custom?
A: The tool performs optimally with high-resolution images (800px or larger) featuring clear, well-defined objects. Current strengths include fashion items, accessories, perfumes, and rigid products. The system handles objects with distinct boundaries better than subjects with soft edges or complex textures.
Q: Can I use IC-Custom for commercial projects?
A: Yes, the tool is released under an open-source license that supports commercial use. However, users must comply with local laws and use the tool responsibly. Review the complete license terms in the GitHub repository before deploying in commercial applications.
Q: How does IC-Custom compare to traditional photo editing software?
A: Unlike traditional editing tools that rely on manual layer manipulation, IC-Custom uses AI to understand visual relationships and automatically adapt objects to new contexts. This approach preserves lighting, shadows, and perspective relationships that would require significant manual work in conventional editing software.
Q: What hardware requirements are needed to run IC-Custom locally?
A: The system requires CUDA-capable GPUs for optimal performance, with testing conducted on CUDA 12.4 and PyTorch 2.6.0. Local installation also requires Python 3.10 and several GB of storage for model checkpoints. The web demo provides an alternative for users without suitable hardware.
