V1.0a <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Experimental!!Important, pls read!!<h1 id="gemma-3-12b-heretic-x-(sikaworld-high-fidelity-edition)">Gemma-3-12B-Heretic-X (Sikaworld High-Fidelity Edition)</h1>This is the ultra-dynamic, fully uncensored text encoder for LTX-2, based on the experimental <a target="_blank" rel="ugc" href="https://www.google.com/url?sa=E&amp;q=https%3A%2F%2Fhuggingface.co%2FLastRef%2Fgemma-3-12b-it-heretic-x">Heretic-X</a> fine-tune by LastRef.While the standard abliterated version removes the "refusal" mechanism, Heretic-X was actively steered with a custom dataset to be proactively descriptive and uninhibited. In LTX-2 video generation, this translates to significantly stronger motion vectors, helping to "unfreeze" static videos and generate more intense dynamics in complex scenes.This edition applies the Sikaworld High-Fidelity Quantization method to tame the aggressive nature of Heretic-X, ensuring that the increased dynamics do not come at the cost of facial symmetry or anatomical coherence.<h2 id="key-features">🚀 Key Features</h2><ul><li>Aggressive Uncensoring (Heretic-X): Unlike standard abliteration (which just deletes the refusal direction), this model uses modified weights (attn.o_proj, mlp.down_proj) derived from x-rated dataset training. It delivers a "louder" and more confident signal to the video transformer, which is often the cure for "frozen" I2V generations.</li><li>High-Fidelity Layer Protection (The Stabilizer): Aggressive fine-tunes can often lead to "melting" faces in video. This version uses a Mixed Precision Strategy: The critical input layers (0-1) and the final output layers (44-47), as well as all LayerNorms and Biases, are kept in BF16. This acts as a safety rail, keeping facial features symmetric while allowing the body and background to move dynamically.</li><li>True Standalone (.safetensors): Includes the embedded spiece_model tensor. It works as a single-file plug-and-play solution in ComfyUI (LTX-2) without requiring external tokenizer.model files or complex folder structures.</li><li>Surgical Extraction: Stripped of the 20GB+ Vision-Tower weights (which LTX-2 does not use) to save VRAM and loading time, while retaining the full 48-layer text intelligence of the 24GB BF16 source.</li></ul><h2 id="usage-in-comfyui">🛠 Usage in ComfyUI</h2><ol><li>Place the .safetensors file in your ComfyUI/models/text_encoders/ folder.</li><li>In your LTX-2 workflow (DualCLIPLoader), select this model.</li><li>Recommended Dtype: Set weight_dtype to fp8_e4m3fn (the critical layers remain BF16 automatically).</li><li>Prompting Tip: This model reacts very well to "action verbs" at the very beginning of the prompt. It requires less CFG scale than standard models to produce motion.</li></ol><h2 id="technical-background">📊 Technical Background</h2>Why Heretic-X for Video? LTX-2 (especially the Dev version) often suffers from "motion collapse" (frozen video) when the text embedding is too neutral. Heretic-X provides a higher variance in its embeddings.Why this Quantization? Standard FP8 conversions of Heretic models often result in "weird" artifacts because the aggressive weights clip during quantization. By protecting the last 4 layers (44-47) in BF16, we ensure that the final instructions sent to the Video Transformer retain their high-precision spatial alignment, preventing the "uncanny valley" effect often seen in dynamic clips.<h2 id="credits">Credits</h2><ul><li>Base Model: Google Gemma 3</li><li>Heretic Fine-tune: <a target="_blank" rel="ugc" href="https://www.google.com/url?sa=E&amp;q=https%3A%2F%2Fhuggingface.co%2FLastRef">LastRef</a></li><li>Optimization &amp; Architecture Fixes: Sikaworld</li></ul>v1.0<h1 id="gemma-3-12b-it-abliterated-(sikaworld-high-fidelity-edition)">Gemma-3-12B-it-Abliterated (Sikaworld High-Fidelity Edition)</h1>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.<h2 id="key-features">🚀 Key Features</h2><ul><li>Uncensored Freedom: Based on the abliteration technique by <a target="_blank" rel="ugc" href="https://www.google.com/url?sa=E&amp;q=https%3A%2F%2Fhuggingface.co%2Fmlabonne%2Fgemma-3-12b-it-abliterated-v2">Maxime Labonne</a>. This model follows complex or "sensitive" prompts without refusals, ensuring a strong vector signal for high-motion video generation.</li><li>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.</li><li>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.</li><li>FP32 Sourced: Converted directly from the original 47GB FP32 shards to ensure maximum rounding precision during the FP8/BF16 hybrid conversion.</li></ul><h2 id="usage-in-comfyui">🛠 Usage in ComfyUI</h2><ol><li>Place the .safetensors file in your ComfyUI/models/text_encoders/ folder.</li><li>In your LTX-2 workflow, use the DualCLIPLoader or the specific LTXV Text Encoder Loader.</li><li>Tip: For best motion results, leave the negative prompt empty and focus your positive prompt on actions and dynamics.</li></ol><h2 id="technical-background">📊 Technical Background</h2>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.<h2 id="credits">Credits</h2><ul><li>Abliteration: <a target="_blank" rel="ugc" href="https://www.google.com/url?sa=E&amp;q=https%3A%2F%2Fhuggingface.co%2Fmlabonne">mlabonne</a></li><li>Optimization &amp; Quantization: Sikaworld</li></ul><hr />

Files with SHA256: eac548f527dadfba25380c62f73965c951ad0768d615ca666fc866b749ac9cf1