kkxao/Qwen-Image-2.1-PE-I2I-Heretic

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

The kkxao/Qwen-Image-2.1-PE-I2I-Heretic is a 9 billion parameter image-editing prompt rewriter, derived from Qwen/Qwen-Image-2.1-PE-I2I. This model is a fine-tuned Qwen3.5-VL variant that transforms short edit instructions and input images into detailed English edit prompts. Its primary differentiator is the removal of refusal behavior through Heretic directional ablation, making it suitable for generating prompts for a wider range of image edits. It is designed for non-commercial use in image-to-image editing workflows.

Loading preview...

Overview

This model, kkxao/Qwen-Image-2.1-PE-I2I-Heretic, is a 9 billion parameter image-editing prompt rewriter. It is a community derivative of the Qwen/Qwen-Image-2.1-PE-I2I model, which is itself a fine-tuned Qwen3.5-VL. Its core function is to take a concise image editing instruction along with one or more input images and expand them into a detailed English edit prompt.

Key Differentiator: Refusal Behavior Ablation

The most significant feature of this "Heretic" variant is the removal of refusal behavior using the Heretic directional ablation technique. While the original model was already relaxed about edgy image subjects, this ablation specifically targets and reduces refusals for chat-style harmful instructions. Benchmarking shows a reduction in refusals from 100/100 to 9/100 on mlabonne/harmful_behaviors with minimal impact on ordinary behavior (KL divergence of 0.0368).

Capabilities

  • Prompt Rewriting: Converts brief image edit instructions into comprehensive, detailed prompts.
  • Image-to-Image Editing Focus: Specifically designed to enhance prompts for image editing applications.
  • Reduced Refusals: Generates prompts for a broader range of content by mitigating refusal behaviors.

Usage Notes

  • Requires system_prompt.txt (included) for defining the output format.
  • Load using AutoModelForImageTextToText and AutoProcessor.
  • This model is released under the Qwen Research License and is intended for non-commercial use only.

Finetunes
1 models