Key Features

1. Mitigation of Identity Collapse

Our reduces identity collapse. The model is trained to generate highly distinct individuals across a full demographic spectrum (age, gender, ethnicity), effectively eliminating the "same face syndrome" and ensuring unique character representation in every output.

2. High Stylistic Integrity

The model resists the "style bleed" common in other models, which often collapse into a generic, overly polished "influencer" aesthetic. It maintains strict stylistic control, enabling true versatility across genres—from anime and classical art to documentary photography and new wave cinema—without unwanted aesthetic contamination.

3. Enhanced Output Diversity

The model features a significant expansion in output diversity from a single prompt across different seeds. This improvement not only fosters greater creative exploration by reducing output repetition but also provides a richer, more superior foundation for high-quality fine-tuning or distillation.

How to use

Use the default workflow for Qwen Image (not 2512). (Note: Automatic quantization in ComfyUI is not currently supported and must be disabled.) The workflow is the 23kb checkpoint available on this page. rename its extension to .json

How to Create Good Images: A Prompting Guide

We've found that the QwenImage series, including this model, performs best with short to medium-length prompts. You don't need to write overly artistic or abstract prose; often, describing your idea directly and clearly yields the best results.

Here are some best practices:

Be Direct and Concise: Simply inputting your core idea can often generate a good image without complex modifications.
Handling Long Prompts: If you want to use a very long prompt from another source (which are often rewritten and expanded by LLMs), you can ask an LLM to simplify it. A good instruction would be: "Rewrite this prompt to be precise and detailed, but without redundant keyword stuffing."
Focus on What Matters:
- Precision over Quantity: Prioritize precise descriptions for key elements like pose and lighting.
- Avoid Redundancy: For stylistic controls, use only one or two strong, distinct terms. For example, fashion photography, film grain is more effective than stacking many similar keywords

Note:

1. This is an experimental version and it may produce a higher rate of artifacts. For better results, I recommend using a sampling step count of 50 or higher cfg. I am actively working to address this issue.

QwenImage-SuperAesthetic(preview) - preview