license: apache-2.0 language:

• en pipeline_tag: image-text-to-text library_name: comfyui base_model: huihui-ai/Qwen2.5-VL-7B-Instruct-abliterated tags:
• comfyui
• qwen2.5-vl
• qwen-image
• qwen-image-edit
• qwen-image-edit-2511
• abliterated
• uncensored
• int8
• convrot

Qwen2.5-VL-7B Huihui Abliterated — ComfyUI BF16 and ConvRot INT8

Single-file ComfyUI text encoders for Qwen Image and Qwen Image Edit 2511, built from huihui-ai/Qwen2.5-VL-7B-Instruct-abliterated.

The upstream model card describes that model as an abliterated/uncensored Qwen2.5-VL-7B-Instruct variant and states that only the language part was abliterated; its vision tower was not altered.

Files

Recommended: learned-rounding INT8 ConvRot

qwen_2.5_vl_7b_huihui_abliterated_int8_convrot.safetensors

• Size: 10,064,106,602 bytes (9.37 GiB)
• SHA-256: 3dc4aae7dc34000c95de546cb220f1b67e51c86d7095fb9e3d19cec7032f5df7
• 1,647 tensors
• 306 row-wise INT8 weights
• 306 FP32 row-wise weight scales
• 306 ComfyUI comfy_quant descriptors
• Every INT8 layer has:
  • format: int8_tensorwise
  • per_row: true
  • convrot: true
  • convrot_groupsize: 256
• 423 tensors remain BF16

Sensitive embeddings, the first and last language blocks, the first vision block, vision patch/position components, the vision merger, norms, and biases remain BF16. Vision MLP down-projections also remain BF16 because their input width (3420) is not divisible by the selected ConvRot group size; this avoids silently mixing ordinary non-ConvRot INT8 layers into a ConvRot-labeled file.

Full precision reference: BF16

qwen_2.5_vl_7b_huihui_abliterated_bf16.safetensors

• Size: 16,584,415,728 bytes (15.45 GiB)
• SHA-256: 82343dee991fe55532d6cd6b5eeae86f16803dfb38f1382b86dbe3c478f4cafe
• 729 tensors, all BF16
• Direct single-file merge of the pinned upstream shards

ComfyUI installation

Place either file under:

ComfyUI/models/text_encoders/

Subfolders are supported. In CLIPLoader, select the file and set the type to qwen_image.

The encoder is compatible with Qwen Image Edit 2511. It contains both the Qwen2.5-VL language model and the vision tower used for image-conditioned prompt encoding.

This is only a text encoder. It does not require different sampler steps, CFG, scheduler, or LoRA strength; keep the settings recommended for your diffusion model or acceleration LoRA.

Use a current ComfyUI build with its pinned comfy-kitchen dependency. ConvRot metadata requires a recent quantization-aware ComfyUI loader.

Runtime verification

Both files were tested with:

• ComfyUI commit: 961212abc8bdcd74514dff389c682672be312711
• comfy-kitchen==0.2.22
• NVIDIA GeForce RTX 5090
• ComfyUI QWEN_IMAGE text-encoder loader
• A real image-conditioned encode, exercising both the Qwen2.5-VL vision tower and language model

Results:

• BF16: loaded and encoded successfully
• INT8: loaded exactly 306 quantized modules
• INT8: all 306 modules reported ConvRot parameters
• Both produced finite conditioning with shape (1, 21, 3584)

The INT8 file also passed structural validation of every weight, scale shape, per-layer descriptor, and global quantization metadata entry. All 177 protected BF16 weight tensors were checked against the merged source and were unchanged.

Provenance

Upstream source:

repository: huihui-ai/Qwen2.5-VL-7B-Instruct-abliterated
revision:   fa935a7958b3669b194c7ba4d1cfcebbe222641d

The four downloaded source shards matched the SHA-256 values published by Hugging Face before they were merged.

The source uses the Apache-2.0 license. Abliteration reduces refusal behavior but does not guarantee that every refusal or safety behavior has been removed.

Conversion

Converted locally with silveroxides/convert_to_quant 1.3.1:

env PYTHONPATH=.deps /media/p5/miniforge3/bin/python .deps/bin/ctq \
  -i qwen_2.5_vl_7b_huihui_abliterated_bf16.safetensors \
  -o qwen_2.5_vl_7b_huihui_abliterated_int8_convrot.safetensors \
  --int8 \
  --scaling_mode row \
  --convrot \
  --convrot-group-size 256 \
  --comfy_quant \
  --save-quant-metadata \
  --qwen35 \
  --exclude-layers '(model\.layers\.27\.|visual\.blocks\.[0-9]+\.mlp\.down_proj\.)' \
  --low-memory \
  --device cuda \
  --manual-seed 42 \
  --num-iter 4000

The run used Prodigy AdaRound optimization with plateau-based early stopping, not the converter's --simple mode.