Standard Gemma has residual refusals that can still gut perfectly innocent prompts. This removes those limitations so LTX-2.5 actually follows what you write (NSFW is still better handled by LoRAs — this is mainly for prompt fidelity).

First model I’ve put real work into and released. If it helps you, a like is appreciated.

Model Details

First community-made uncensored Gemma 4 12B text encoder for the brand-new LTX-2.5 in ComfyUI. Built from llmfan46/gemma-4-12B-it-uncensored-heretic and carefully converted to single-file ComfyUI format while fully preserving the original LTX-2.5 video + audio projection weights.

Model Description

This is a BF16 and Int8 ContRot Gemma 4 Unified 12B text encoder for LTX-2.5 workflows in ComfyUI.

The conversion preserves the Gemma 4 language, vision, audio, and multimodal components while preserving the LTX-specific video/audio text projections.

Note: See this before you use: https://old.reddit.com/r/StableDiffusion/comments/1vmdxzk/psa_im_the_creator_of_heretic_and_i_advise_you_to/

My Response: Fully agree that swapping in a Heretic/abliterated LLM as a text encoder does not “uncensor” the visual or video output of a diffusion model. The generation model itself still has whatever biases and limitations it was trained with.

My use case with the LTX-2.5 Gemma encoder was narrower. I was running into residual refusal behavior in the stock text encoder that was still interfering with perfectly ordinary prompts (not NSFW-focused). Even after LTX’s own conditioning, some benign requests were getting weakened or partially ignored.

So I took an abliterated Gemma 4 backbone and carefully remapped the original LTX-2.5 projection layers (language, vision, and audio) into it so the conditioning path stayed intact. Goal was just higher prompt fidelity and fewer silent failures on normal requests, not magic uncensoring of the video model.

This was something I made for myself, and thought I would share for others and see if it useful to them. I make NO PROMISES.

Developed by: DeepNeuralNerd
Funded by: Not applicable
Shared by: DeepNeuralNerd
Model type: Gemma 4 Unified 12B LTX-2.5 ComfyUI encoder
Language(s): Multilingual; see the base model card
License: See the upstream Gemma 4 and LTX-2 license terms
Finetuned from model: llmfan46/gemma-4-12B-it-uncensored-heretic

Model Sources

Base model: https://huggingface.co/llmfan46/gemma-4-12B-it-uncensored-heretic
ComfyUI: https://github.com/comfyanonymous/ComfyUI
Paper: https://arxiv.org/abs/2601.03233

Uses

Direct Use

This model is intended for local LTX-2.5 video and audiovisual generation in ComfyUI.

Place the file in:

ComfyUI/models/text_encoders/

Then select:

Gemma-4-12B-it-uncensored-heretic-DeepNeuralNerd-LTX_2.5_ComfyUI-bf16.safetensors

in the LTX-2.5 text-encoder loader.

Downstream Use

Intended uses include:

LTX-2.5 text-to-video generation
LTX-2.5 audiovisual generation
Prompt conditioning
Supported LTX-2.5 image-to-video workflows

Out-of-Scope Use

This model is not intended as:

A generic replacement for every Gemma 4 checkpoint
A replacement for Gemma 3 LTXV encoders
A standalone general-purpose Transformers checkpoint
A replacement for the complete LTX-2.5 diffusion, VAE, audio, or upscaler models
A safety-certified model for high-stakes applications

Bias, Risks, and Limitations

The base model is an uncensored/decensored Gemma 4 model and may produce content that standard instruction-tuned checkpoints refuse.

Recommendations

Use appropriate safeguards and review generated content before publication or distribution.

For comparisons, use the same prompt, seed, resolution, frame count, sampler, and LTX diffusion model.

How to Get Started with the Model

Copy the .safetensors file into ComfyUI/models/text_encoders/.
Restart ComfyUI or refresh the model list.
Open an LTX-2.5 workflow.
Select the DeepNeuralNerd encoder.
Generate a test clip with a fixed seed.

There is both a BF16 and a INT8 ConvRot encoder.

Conversion Details

Source tensors were remapped as follows:

text

model.language_model.*                         -> model.*
model.vision_embedder.*                        -> vision_model.*
model.embed_vision.embedding_projection.weight -> multi_modal_projector.embedding_projection.weight
model.embed_audio.embedding_projection.weight  -> audio_projector.embedding_projection.weight

The source tokenizer.json was embedded as:

tokenizer_json

Training Details

Training Data

No additional training was performed.

The base model’s training and abliteration details are documented in:

https://huggingface.co/llmfan46/gemma-4-12B-it-uncensored-heretic

Training Procedure

This model was produced by deterministic safetensors repackaging and tensor-layout conversion.

Speeds, Sizes, Times

Precision: BF16 and INT8 ConvRot
Format: Safetensors

Evaluation

Testing Data, Factors & Metrics

Testing Data

Runtime validation was performed in a local ComfyUI LTX-2.5 workflow.

This was not a standardized benchmark dataset.

Factors

Validation checked:

Model discovery
Text-encoder loading
LTX projection loading
Prompt conditioning
End-to-end generation

Metrics

The primary metric was successful end-to-end generation without model-loading or tensor-shape errors.

Results

The converted encoder successfully loaded and generated output in the tested LTX-2.5 ComfyUI workflow. Runtime validation passed in the tested LTX-2.5 ComfyUI workflow.

This is a community conversion and is not an official Lightricks release.

Model Examination

The output contains these principal tensor groups: