IP Adapter

Use images as prompts to transfer subject matter, style, and composition to new generations.

IP Adapter revolutionizes image generation by allowing images to serve as prompts, thus seamlessly transferring subject matter or style to new generations. IP Adapter can be accessed through ENDPOINT_URL https://image.octoai.run/generate/ip-adapter-sdxl with the following API parameters:

  • ip_adapter_method, supported values full, style, composition
  • ip_adapter_scale, supported range between 0.1 to 1
  • ip_adapter_image : BASE_64 image

OctoAI’s IP Adapter integrates with SDXL models, enhancing its adaptability and functionality when used with style presets, checkpoints, and LoRAs. However, it currently does not support ControlNets.

1. Full Transfer of Style and Composition

For our first example, we will take an embroidery pattern from a fabric and impose a similar pattern on a garment.

  • Parameter Settings:
    • ip_adapter_method: full
    • ip_adapter_scale: Adjust for optimal results.
    • ip_adapter_image : BASE_64 image.
    • prompt: Desired text prompt.

Output: A garment featuring the transferred embroidery pattern, retaining both the style and composition of the original fabric.

IP Adapter Method: full

Sample Code Snippet

Python
1{
2    "prompt": "A mannequin wearing a long v-neck dress, sleeveless. The mannequin is black, minimalistic, and headless, set against a neutral background.",
3    "negative_prompt": "deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
4    "sampler": "DPM_PLUS_PLUS_2M_KARRAS",
5    "checkpoint": "octoai:RealVisXL",
6    "cfg_scale": 7,
7    "num_images": 1,
8    "width": 1024,
9    "height": 1024,
10    "steps": 30,
11    "use_refiner": "False",
12    "ip_adapter_scale": 0.9,
13    "ip_adapter_method": "full",
14    "ip_adapter_image": <BASE_64>
15    }

2. Style Transfer Only

Next, we’ll generate apparel that embodies the style of a golden spherical ball with intricate patterns, without replicating its composition.

  • Parameter Settings:
    • ip_adapter_method: style
    • ip_adapter_scale: Adjust for optimal results.
    • ip_adapter_image : BASE_64 image.
    • prompt: Desired text prompt.

Output: A garment reflecting the intricate style and colors of the golden sphere.

IP Adapter Method: style

Sample Code Snippet

Python
1{
2    "prompt": "A mannequin wearing a long v-neck dress, sleeveless. The mannequin is black, minimalistic, and headless, set against a neutral background.",
3    "negative_prompt": "deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
4    "sampler": "DDIM",
5    "checkpoint": "octoai:default-sdxl",
6     "sampler": "DPM_PLUS_PLUS_2M_KARRAS",
7    "checkpoint": "octoai:RealVisXL",
8    "cfg_scale": 7,
9    "num_images": 1,
10    "width": 1024,
11    "height": 1024,
12    "steps": 30,
13    "use_refiner": "False",
14    "ip_adapter_scale": 0.9,
15    "ip_adapter_method": "style",
16    "ip_adapter_image":<BASE_64>
17    }

3. Color Palette Transfer

For our third example, we’ll apply a color palette to generate apparel by transferring only the colors.

  • Parameter Settings:
    • ip_adapter_method: style
    • ip_adapter_scale: Adjust for optimal results.
    • ip_adapter_image : BASE_64 image.
    • prompt: Desired text prompt.

Output: Garments displaying colors from the provided palette, with variations based on the ip_adapter_scale.

IP Adapter Method: style

Sample Code Snippet

Python
1{
2    "prompt": "A mannequin wearing a long v-neck dress, sleeveless. Background is ((neutral)).",
3    "negative_prompt": "deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
4    "sampler": "DPM_PLUS_PLUS_2M_KARRAS",
5    "checkpoint": "octoai:RealVisXL",
6    "cfg_scale": 7,
7    "num_images": 1,
8    "width": 1024,
9    "height": 1024,
10    "steps": 30,
11    "use_refiner": "False",
12    "ip_adapter_scale": 0.99,
13    "ip_adapter_method": "style",
14    "ip_adapter_image":<BASE_64>
15    }

4. Composition Transfer

Lastly, let’s focus on transferring the composition from a fluffy ballroom gown and fairy wings to generate new apparel.

  • Parameter Settings:
    • ip_adapter_method: composition
    • ip_adapter_scale: Adjust to achieve the desired outline and structure.
    • ip_adapter_image : BASE_64 image.
    • prompt: Desired text prompt.

Output: New designs showcasing the transferred composition while allowing for style variations.

IP Adapter Method: composition

Sample Code Snippet

Python
1{
2    "prompt": "A mannequin wearing a blue dress. Background is ((neutral)).",
3    "negative_prompt": "deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
4    "sampler": "DPM_PLUS_PLUS_2M_KARRAS",
5    "checkpoint": "octoai:RealVisXL",
6    "cfg_scale": 7,
7    "num_images": 1,
8    "width": 1024,
9    "height": 1024,
10    "steps": 30,
11    "use_refiner": "False",
12    "ip_adapter_scale": 0.8,
13    "ip_adapter_method": "composition",
14    "ip_adapter_image":<BASE_64>
15    }

5. Generating Variations Without Text Prompts

To generate variations of a reference apparel without using text prompts:

  • Use the reference image directly as a prompt and experiment with different parameters for diverse outputs.
  • Parameter Settings:
    • ip_adapter_method: full
    • ip_adapter_scale: Adjust for optimal results.
    • ip_adapter_image : BASE_64 image
    • prompt: None

Output: Multiple variations of the original apparel, reflecting different styles or minor tweaks.

IP Adapter Method: full, No text prompt

Sample Code Snippet

Python
1{
2    "sampler": "DPM_PLUS_PLUS_2M_KARRAS",
3    "checkpoint": "octoai:RealVisXL",
4    "cfg_scale": 7,
5    "num_images": 1,
6    "width": 1024,
7    "height": 1024,
8    "steps": 30,
9    "use_refiner": "False",
10    "ip_adapter_scale": 0.8,
11    "ip_adapter_method": "full",
12    "ip_adapter_image":<BASE_64>
13    }

By leveraging the IP Adapter, businesses can generate consistent, visually appealing images across various applications—from fashion and gaming to marketing and beyond.