DreamFast/gemma-3-12b-it-heretic-v2
DreamFast/gemma-3-12b-it-heretic-v2 is an abliterated version of Google's Gemma 3 12B IT model, created using Heretic v1.3.0. This model significantly reduces refusals while maintaining quality, making it suitable as an uncensored text encoder for video generation models like LTX-2. It is available in multiple quantization formats including FP8, INT8, INT4, NVFP4, and MXFP8, all optimized with SVD-guided learned rounding for high fidelity across various GPU architectures. The model is primarily designed to provide more faithful prompt encoding for creative content by removing soft censorship in embeddings.
Loading preview...
Model Overview
DreamFast/gemma-3-12b-it-heretic-v2 is an "abliterated" version of Google's Gemma 3 12B IT model, processed using Heretic v1.3.0. The primary goal of this modification is to reduce model refusals and soft censorship, making it a more flexible text encoder, particularly for video generation applications like LTX-2. The abliteration process involved 200 optimization trials, with Trial 174 selected for achieving 8/100 refusals (down from 100/100) while maintaining a low KL Divergence of 0.0801, indicating minimal damage to the base model's quality.
Key Capabilities & Features
- Reduced Refusals: Significantly lowers the model's tendency to refuse prompts, enabling more faithful encoding of creative content.
- Vision Preserved: All ComfyUI variants retain
vision_modelandmulti_modal_projectorkeys, supporting I2V (image-to-video) prompt enhancement. - Extensive Quantization: Offered in five ComfyUI-compatible quantization formats (FP8, INT8 ConvRot, INT4 W4A4 ConvRot, NVFP4, MXFP8) and GGUF, optimized with SVD-guided learned rounding for maximum fidelity across diverse hardware.
- ComfyUI & LTX-2 Integration: Designed for seamless use with ComfyUI, especially for LTX-2 text-to-video (T2V) and image-to-video (I2V) workflows.
Ideal Use Cases
- Uncensored Text Encoding: For applications requiring a text encoder with reduced censorship, particularly for creative or niche content generation.
- Video Generation (LTX-2): Functions as a text encoder for LTX-2, aiming to produce more faithful visual outputs by removing embedded sanitization.
- Resource-Constrained Environments: Various quantization options (e.g., INT4 at 7.7 GB) allow deployment on GPUs with limited VRAM, while learned rounding ensures high quality.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.