About SDXL Refiner
SDXL Refiner is a specialized refinement model that focuses on enhancing image details and quality based on the SDXL base model. This process is like having an art retoucher:- First, it receives preliminary images or text descriptions generated by the base model
- Then, it guides the refinement process through precise aesthetic scoring and dimensional parameters
- Finally, it focuses on processing high-frequency image details to improve overall quality
- As a standalone refinement step for post-processing images generated by the base model
- As part of an expert integration system, taking over processing during the low-noise phase of generation
Inputs
Parameter Name | Data Type | Input Type | Default Value | Value Range | Description |
---|---|---|---|---|---|
clip | CLIP | Required | - | - | CLIP model instance used for text tokenization and encoding, the core component for converting text into model-understandable format |
ascore | FLOAT | Optional | 6.0 | 0.0-1000.0 | Controls the visual quality and aesthetics of generated images, similar to setting quality standards for artwork: - High scores(7.5-8.5): Pursues more refined, detail-rich effects - Medium scores(6.0-7.0): Balanced quality control - Low scores(2.0-3.0): Suitable for negative prompts |
width | INT | Required | 1024 | 64-16384 | Specifies output image width (pixels), must be multiple of 8. SDXL performs best when total pixel count is close to 1024×1024 (about 1M pixels) |
height | INT | Required | 1024 | 64-16384 | Specifies output image height (pixels), must be multiple of 8. SDXL performs best when total pixel count is close to 1024×1024 (about 1M pixels) |
text | STRING | Required | - | - | Text prompt description, supports multi-line input and dynamic prompt syntax. In Refiner, text prompts should focus more on describing desired visual quality and detail characteristics |
Outputs
Output Name | Data Type | Description |
---|---|---|
CONDITIONING | CONDITIONING | Refined conditional output containing integrated encoding of text semantics, aesthetic standards, and dimensional information, specifically for guiding SDXL Refiner model in precise image refinement |
Notes
- This node is specifically optimized for the SDXL Refiner model and differs from regular CLIPTextEncode nodes
- An aesthetic score of 7.5 is recommended as the baseline, which is the standard setting used in SDXL training
- All dimensional parameters must be multiples of 8, and total pixel count close to 1024×1024 (about 1M pixels) is recommended
- The Refiner model focuses on enhancing image details and quality, so text prompts should emphasize desired visual effects rather than scene content
- In practical use, Refiner is typically used in the later stages of generation (approximately the last 20% of steps), focusing on detail optimization