PE-Heretic-INT8-ConvRot-for-Qwen-Image-2.1

PE Heretic INT8 ConvRot for Qwen Image 2.1

ComfyUI-native INT8 ConvRot conversions of the Heretic Qwen Image 2.1 T2I and I2I Prompt Enhancers.

Community conversion. Not affiliated with or endorsed by Alibaba / Qwen, pottokao, darrellbest, Comfy-Org, or the Heretic project.

The model weights are distributed under the Qwen Research License. Non-commercial use only. See LICENSE.

Models

This repository contains two models:

T2I

File:

qwen3.5_9b_qwen_image_2.1_pe_t2i_heretic.int8_convrot.safetensors

Size:

9,471,072,252 bytes

The T2I Prompt Enhancer takes a short or simple text-to-image request and rewrites it into a detailed prompt for Qwen Image 2.1, including the recommended aspect ratio.

I2I

File:

qwen3.5_9b_qwen_image_2.1_pe_i2i_heretic.int8_convrot.safetensors

Size:

9,471,072,252 bytes

The I2I Prompt Enhancer takes one or more source images together with an editing instruction and produces a detailed image-editing prompt for Qwen Image 2.1.

What was modified

The Heretic directional-ablation modifications were performed by the respective upstream model authors:

• T2I: pottokao/Qwen-Image-2.1-PE-T2I-Heretic
• I2I: darrellbest/Qwen-Image-2.1-PE-I2I-Heretic

No additional abliteration, fine-tuning, or behavioral modification was performed in this repository.

The modification performed here is:

BF16 Heretic model → ComfyUI INT8 ConvRot

The provenance chains are therefore:

Qwen/Qwen-Image-2.1-PE-T2I
→ pottokao/Qwen-Image-2.1-PE-T2I-Heretic
→ this INT8 ConvRot conversion

and:

Qwen/Qwen-Image-2.1-PE-I2I
→ darrellbest/Qwen-Image-2.1-PE-I2I-Heretic
→ this INT8 ConvRot conversion

Source revisions

The following upstream revisions were retained in the local Hugging Face cache (refs/main) associated with the source downloads used for these conversions:

• T2I source revision: 96e92081f46cd0ca2ac9679432c5b8b02c12cfdb
• I2I source revision: 915c4c14750fbcc73cb181afaa0efbc41ef10e36

The original local download metadata was deleted after conversion, so these revisions are reported from the surviving Hugging Face cache references rather than from retained source snapshots.

Quantization

The models were converted using Comfy-Org's comfy-model-tools, specifically:

quant_int8_convrot.py

Conversion configuration:

• INT8 tensor-wise quantization
• ConvRot enabled
• default absmax scaling
• lm_head explicitly included
• no MSE clipping
• embedding quantized separately by the converter

Each merged BF16 source contained:

• 760 tensors
• all source tensors: BF16
• merged size: 18,819,722,360 bytes

Each resulting INT8 ConvRot model contains:

• 1,380 total tensors
• 310 quantized weights
• 310 I8 tensors
• 310 U8 tensors
• 310 F32 tensors
• 450 BF16 passthrough tensors
• 310 comfy_quant metadata entries
• 310 weight_scale tensors
• quantization format: int8_tensorwise
• 310 ConvRot-quantized weights

Group-size distribution:

This consists of 309 regular quantized layers + 1 quantized embedding table.

Structural verification

Both converted Heretic models were structurally compared against their corresponding official ComfyUI INT8 ConvRot Prompt Enhancer format.

For both T2I and I2I:

• tensor count matches exactly
• tensor keys match exactly
• tensor shapes match exactly
• tensor dtypes match exactly
• comfy_quant metadata matches structurally
• quantized-weight count matches exactly
• ConvRot group-size distribution matches exactly

This verifies compatibility of the file structure and quantization schema. It does not mean that the numerical model weights are identical to the official non-Heretic models.

Both converted models were also tested successfully through actual ComfyUI Prompt Enhancer inference.

SHA256 checksums

The following SHA-256 checksums identify the exact INT8 ConvRot .safetensors files distributed in this repository.

T2I

91b9ba42539fb662775181eabce845befe2c1441658a736fadf0033dd799d8f3

I2I

979b9063dd16645f0191c9ff55aa5059105bfe84096d96a817a85deed22231f0

Installation

Place the .safetensors files in your ComfyUI text-encoder directory.

For example:

ComfyUI/models/text_encoders/Qwen/2.1_Prompt_Enhancer/

Then restart ComfyUI or refresh its model list.

The models are intended to be used through the same native ComfyUI Prompt Enhancer workflow as the corresponding official Qwen Image 2.1 INT8 ConvRot Prompt Enhancers.

Which model should I use?

PE-T2I-Heretic

Use the T2I model when your input is a normal text-to-image prompt.

Typical flow:

short T2I prompt
→ PE-T2I-Heretic
→ rewritten detailed prompt
→ Qwen Image 2.1 text encoder
→ Qwen Image 2.1
→ image

PE-I2I-Heretic

Use the I2I model for image editing, where the Prompt Enhancer receives source image information together with an edit request.

Typical flow:

source image(s) + edit instruction
→ PE-I2I-Heretic
→ rewritten detailed edit instruction
→ Qwen Image 2.1
→ edited image

System prompts

The Prompt Enhancers rely on their corresponding upstream system prompts.

The T2I and I2I system prompts are not interchangeable.

This repository includes them separately as:

system_prompts/t2i/system_prompt.txt

and:

system_prompts/i2i/system_prompt.txt

These files should be copied unchanged from the corresponding upstream Heretic repositories.

Important distinction

These files are Prompt Enhancers / prompt rewriters.

They are not:

• the Qwen Image 2.1 diffusion transformer
• the standard Qwen3-VL text encoder
• standalone image-generation models

Their job is to rewrite instructions before those instructions are used by the rest of the Qwen Image 2.1 pipeline.

Conversion tooling

INT8 ConvRot conversion was performed with:

Comfy-Org/comfy-model-tools

using quant_int8_convrot.py.

The conversion software is distributed separately under its own software license. The model weights distributed here remain subject to the applicable Qwen Research License.

Credits

Original Qwen Image 2.1 Prompt Enhancers:

• Qwen / Alibaba

Heretic T2I model:

• pottokao/Qwen-Image-2.1-PE-T2I-Heretic

Heretic I2I model:

• darrellbest/Qwen-Image-2.1-PE-I2I-Heretic

Heretic directional-ablation tooling:

• p-e-w/heretic

INT8 ConvRot conversion tooling and ComfyUI support:

• Comfy-Org/comfy-model-tools
• ComfyUI

This repository provides the INT8 ConvRot conversion, structural validation, ComfyUI inference testing, and packaging of the two upstream Heretic Prompt Enhancers.

License

The model weights in this repository are distributed under the Qwen Research License.

See the complete license text in LICENSE.

These models are for non-commercial use unless the user has obtained the appropriate separate commercial license from Qwen.

A separate NOTICE file contains attribution and modification information.

Disclaimer

This is an independent community conversion.

It is not an official Qwen release and is not affiliated with or endorsed by Alibaba / Qwen, pottokao, darrellbest, the Heretic project, or Comfy-Org.