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.
Why this exists: 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).
See for Details: https://huggingface.co/DeepNeuralNerd/Gemma-4-12B-it-uncensored-heretic-DeepNeuralNerd-LTX_2.5_ComfyUI
Available versions:
BF16
INT8 ConvRot
Drop the .safetensors into ComfyUI/models/text_encoders/ and select it in any LTX-2.5 text-encoder loader.
First model I’ve put real work into and released. If it helps you, a like is appreciated.
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.
Description
FAQ
Comments (9)
ATTENTION: The creator of the heretic model posted on reddit saying using these text encoders will only make your gens worse. So user beware! https://www.reddit.com/r/StableDiffusion/comments/1vmdxzk/psa_im_the_creator_of_heretic_and_i_advise_you_to/
Correct, Heretic used as a text encoder produces worse prompt following, artifacts.
@nobodiesbusiness Agree. See response. I limited the modification to the text encoder portion, and not the other projectors. It was a VERY surgical modification. See the huggingface repo for this for detail. Also, added your notice in the description itself. People should test this out to see if it helps their particular use case. My intent is NOT to mislead people. Also, I encourage people to post what their results have been . Because if this is useless and/or causes more problems, I can take this model down. Don't need to clutter the space with useless things.
I'll share my opinion too. I've noticed that with many models, this text encoder dramatically improves the quality of the output video, and most importantly, it dramatically enhances detail and style. The video comes out truly lifelike. The anatomy of the fingers, toes, and body is correct, with subtle blurring. I especially appreciate the fusion models by joeygambino. But understand, the encoder is only part of the process. The basis is the model and the correct prompt. If you have difficulty composing prompts, connect the text generator https://huggingface.co/Comfy-Org/gemma-4/blob/refs%2Fpr%2F4/text_encoders/gemma4_e4b_it_int8_convrot.safetensors or https://huggingface.co/Comfy-Org/gemma-4/blob/refs%2Fpr%2F4/text_encoders/gemma4_e2b_it_int8_convrot.safetensors to your workflow. Following your instructions, they'll create a well-written prompt for you, and the video will be truly lively and rich. If for any reason your prompt isn't accepted or doesn't meet your requirements, edit the nodes_textgen.py file in the comfy_extras folder, and you'll be happy. The author showed enthusiasm and put in a tremendous amount of time. So, don't blame them. Learn to criticize yourself, and then, once you've gained experience, you can criticize others.
Any chances for bf16?
its there, available in drop-down
@orzechowy3334318 - Sorry, it was a slow upload and kept messing up. But it is up now. Just wanted to contribute in some small way.
Didn't you find the Heretic's dev post.. weird? He's pretty much saying "this thing I made, it doesn't work, don't use it"... Couldn't he be covering himself legally by making a post like that?
Works for me as predicted and I'm not seeing a degradation in A/B tests against the LTX 2.5 original Gemma4 12B. Paired it with a Sikaworld Gemma3 (ablit) for the prompt enhancer, and the combo is giving me the outputs I asked for (i2v).
As to why use it (them)... My understanding is that it removes the refusal, but then the model understands (or not) specific anatomy, actions, etc., and that's what Lora's are for. Just my thoughts...