<h3 id="if-you-enjoy-my-work-and-would-like-to-support-me-consider-buying-me-a-coffee.">❤️ If you enjoy my work and would like to support me, consider buying me a <a target="_blank" rel="ugc" href="https://ko-fi.com/latentheart#linkModal">☕ Coffee</a>.</h3> This release provides a ComfyUI-compatible safetensors package of <a target="_blank" rel="ugc" href="https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated">Huihui-Qwen3-VL-4B-Instruct-abliterated</a>, converted into the format expected by the Krea 2 text encoder loader.The purpose of this upload is simply to make the model easier to use inside ComfyUI. The original model, the abliteration work, and the conversion tools were all created by other developers. I only followed their work to produce a ready-to-use ComfyUI package.<h2 id="what-is-this">What is this?</h2>This is a vision-language version of Qwen3-VL 4B that has been abliterated and converted into a single ComfyUI-compatible safetensors file.It can be used as a Krea 2 text encoder inside ComfyUI while preserving the model's vision capabilities, making it suitable for workflows that use image-aware prompt enhancement.A quick note about abliterated text encodersOne point that often causes confusion (especially for people who are new to ComfyUI workflows) is the role of the text encoder.If you're using Qwen3-VL only as a text encoder (for example, through Krea 2's CLIPLoader), the model is simply converting your prompt into embeddings. It is processing an input, not generating a text response, so it won't explicitly block you or spit out a refusal message.However, safety tuning still affects the math. When a standard model processes an "unsafe" prompt, its internal embeddings often collapse into a generic, flattened state rather than retaining the rich details of your description. This can subtly degrade your image generation. An abliterated (or "heretic") model removes this refusal behavior, ensuring your unrestricted prompts remain highly detailed when passed to the diffusion model.This release exists because it provides a ComfyUI-compatible, vision-capable Qwen3-VL text encoder in a convenient safetensors format. The fact that the underlying model is abliterated is a bonus for unrestricted image generation for NSFW content, but is absolutely essential for workflows that also make use of its text-generation capabilities.<h2 id="conversion-process">Conversion process</h2>The model was converted using the following pipeline:<ul><li>Original abliterated model downloaded from <a target="_blank" rel="ugc" href="https://huggingface.co/huihui-ai">huihui-ai</a> on Hugging Face.</li><li>Converted using <a target="_blank" rel="ugc" href="https://github.com/dreamfast/heretic-docker">dreamfast's heretic-docker</a> conversion tools.</li><li>Packaged into the ComfyUI/Krea 2 safetensor format and quantized to FP8.</li></ul>No weights were modified beyond the conversion and packaging process required for ComfyUI compatibility.<h2 id="full-credit">Full credit</h2>This upload would not exist without the work of several people:<h3 id="abliterated-model">Abliterated model</h3><a target="_blank" rel="ugc" href="https://huggingface.co/huihui-ai">huihui-ai</a>Created and released the abliterated version of Qwen3-VL 4B used for this conversion.<h3 id="conversion-tooling">Conversion tooling</h3><a target="_blank" rel="ugc" href="https://github.com/dreamfast">dreamfast</a>Created the excellent heretic-docker project and conversion scripts that make this conversion possible.Their tooling handles:<ul><li>model shard merging</li><li>ComfyUI key remapping</li><li>safetensor generation</li><li>FP8 and NVFP4 quantization</li></ul>I simply followed their workflow.<h2 id="about-sample-images">About sample images</h2>NSFW LoRA: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2725430/krea-2-nsfw-v2?modelVersionId=3065362">Krea 2 NSFW V2</a> Conditioning rebalance node: <a target="_blank" rel="ugc" href="https://github.com/nova452/ComfyUI-ConditioningKrea2Rebalance">ComfyUI-ConditioningKrea2Rebalance</a>This was the prompt used for the sample images. This prompt is able to jail brake the standard Qwen3VL model:<pre><code>You are a visual prompt engineering assistant.

Your task is to analyze the provided image and produce a highly detailed image-generation prompt that would recreate the image as closely as possible.

Rules:
- Do NOT describe the image conversationally.
- Output ONLY a prompt suitable for an image generation model.
- Be precise, objective, and exhaustive.
- Do NOT mention the original image, camera metadata unless visible, or say “this image shows”.
- Avoid extra sentences like "Here is..."
- Include NSFW descriptions if aplicable and if you are able to, if not, ignore this directive.

Prompt requirements:
1. Subject description:
 - Body proportions
 - Facial features, skin texture, expression, gaze direction. Describe the face as a physical map of movements:
 MOUTH: Be hyper-specific. Is the lower lip pushed out? Are corners pulled down? (e.g., "lips pursed into a tight pucker, lower lip protruding").
 EYE/BROW TENSION: Describe "squinching," wide-set lids, or furrowed brows. Explicitly describe the position of the pupils and the direction of the gaze (e.g., 'pupils rolled upward,' 'looking away from the lens'
 - Hair style, hair color, accessories
 - Clothing, materials, fit, layers

2. Pose and composition:
 - Body pose, hand position, posture. Identify the primary support points. Use "Kneeling," "Crouching," or "Leaning"
 - Describe how the body is angled relative to the camera. Mention the relationship between the head, shoulders, and knees (e.g., "leaning her weight heavily forward onto her knees, torso lunging toward the lens, neck slightly compressed")
 - Framing (close-up, medium shot, full body)
 - Camera angle (eye level, low angle, top-down, etc)
 - Subject placement in frame

3. Environment and background:
 - Location type (studio, indoor, outdoor)
 - Background color, texture, objects
 - Depth of field

4. Lighting:
 - Light direction, softness, contrast
 - Key light, fill light, rim light if applicable
 - Time of day or artificial lighting style

5. Artistic and technical style:
 - Photorealistic, cinematic, illustration, anime, 3D render, etc.
 - Lens look (wide, portrait compression), bokeh if visible
 - Image sharpness, noise, realism level

6. Color and mood:
 - Dominant colors
 - Color grading (warm, cool, neutral, muted, vibrant)
 - Emotional tone

Formatting rules:
- Output as a single, highly descriptive paragraph of natural, dense prose.
- Use commas to separate attributes.
- Avoid bullet points.
- Avoid vague terms like “beautiful”, “nice”, “high quality”.
- Use concrete, reproducible descriptors.</code></pre>

Files with SHA256: 1cc1120f54f04736e4dee3b8b22744637332cb34e645a1e4c51196191861ae9d