base_model: mlabonne/gemma-3-12b-it-abliterated-v2
tags:
- quantized
- fp8
- mixed_precision
- gemma3
- mlabonne
- ltx2
license: apache-2.0
Gemma-3-12B-it-Abliterated (Sikaworld High-Fidelity Edition)
This is a specialized, fully uncensored (abliterated) text encoder for the LTX-2 audiovisual model.
While standard FP8 conversions often lead to "frozen" videos, facial drifting, or anatomical asymmetry in Image-to-Video (I2V) workflows, this version was surgically optimized to preserve the intelligence and stability of the original model.
🚀 Key Features
Uncensored Freedom: Based on the abliteration technique by Maxime Labonne. This model follows complex or "sensitive" prompts without refusals, ensuring a strong vector signal for high-motion video generation.
High-Fidelity Layer Protection: Unlike radical FP8 quants, this version uses a Mixed Precision Strategy. Critical input layers (0-1) and final output layers (44-47), as well as all LayerNorms and Biases, are kept in BF16. This specifically fixes the "face shifting" and "asymmetry" issues common in LTX-2.
True Standalone (.safetensors): Includes the embedded spiece_model tensor. It works as a single-file plug-and-play solution in ComfyUI without requiring external tokenizer.model files.
FP32 Sourced: Converted directly from the original 47GB FP32 shards to ensure maximum rounding precision during the FP8/BF16 hybrid conversion.
🛠 Usage in ComfyUI
Place the .safetensors file in your ComfyUI/models/text_encoders/ folder.
In your LTX-2 workflow, use the DualCLIPLoader or the specific LTXV Text Encoder Loader.
Tip: For best motion results, leave the negative prompt empty and focus your positive prompt on actions and dynamics.
📊 Technical Background
Standard 8-bit quantization often "muffles" the subtle signals needed for temporal consistency in video models. By protecting the "navigation" layers (the beginning and end of the 48-layer stack) in BF16, this encoder provides a much "louder" and more stable movement command to the LTX-2 Transformer.
Credits
Abliteration: mlabonne https://huggingface.co/mlabonne/gemma-3-12b-it-abliterated-v2
Optimization & Quantization: Sikaworld