LTX2 Video Nudify Workflow Tutorial

LTX2 is a fairly heavy model and as of right now, there’s no compression methods available for inpainting of LTX2. Expect this workflow to use a lot of VRAM and RAM. I suggest using a GPU-cloud service like Runpod to run this workflow. Follow the latest Runpod tutorial for LTX2 here.
https://gofile.io/d/YRv2oy

Workflow Download:

Download the workflow depending on you’re preference here:
https://gofile.io/d/yYRC9T

Model Downloads:

Important Notes about model types:
In many of these model links, you may notice the same model listed under different names such as Q8, Q6, Q2, fp16, fp8, or fp4. These are all the same base model, but provided at different sizes, each with its own trade-offs.

Models labeled with “fp” are typically faster than GGUF models, but they are less space-efficient. GGUF models focus on better storage efficiency and often make it easier to achieve higher quality results. You can identify a GGUF model by the number that follows the letter Q, such as Q8. This number represents the quantization level, or how much the model has been compressed.

In simple terms, a higher Q number generally means better output quality, but it also requires more GPU resources. Lower Q numbers reduce hardware requirements at the cost of quality. The same idea applies to fp models, where fp16 is usually the largest and highest-quality option, and quality decreases as the number gets smaller. As a rough comparison, a GGUF Q8 model is fairly equivalent to an fp16 model in overall quality.

LTX2 Models (Both workflows):

LTX2 Loras (Both Workflows):

Notes for LTX2 Models and Loras:
The main model has dev and distilled versions. You can use distilled versions, but I prefer using dev versions with a distilled lora at a low strength, which often produces the best ‘non-plasticy’ and non-shiny result. If you use a distilled main model, you will not need the distilled lora.

The embeddings connector has two versions, I didn’t find a huge difference of each version, but I recommend using the dev version for the dev main model and the distilled version for the distilled main model.

I added three of the current best NSFW loras that worked for me. I used them all at the same time, but you can use only one at a time if needed. Using only 1 may help reduce VRAM usage. You can also change these loras as better NSFW loras come out.

The IC Lora Detailer generally helps with detail, but is not needed and may be removed for the sake of saving VRAM.

File Structure for both LTX2 Models and Loras:

Code:

📁 ComfyUI/
 ┣ 📁 models/
 ┃ ┣ 📁 unet/
 ┃ ┃ ┗ 📄 LTXV2_GGUF_Model_Here.gguf
 ┃ ┣ 📁 text_encoders/
 ┃ ┃ ┣ 📄 LTXV2_Text_Encoder.safetensors
 ┃ ┃ ┗ 📄 LTXV2_Embeddings_Connector.safetensors
 ┃ ┣ 📁 vae/
 ┃ ┃ ┣ 📄 LTX2_audio_vae.safetensors
 ┃ ┃ ┗ 📄 LTX2_video_vae.safetensors
 ┃ ┣ 📁 loras/
 ┃ ┃ ┣ 📄 LTXV2_Distilled_Lora.safetensors
 ┃ ┃ ┣ 📄 ltx-2-19b-ic-lora-detailer.safetensors
 ┃ ┃ ┣ 📄 LTX2_BestBreasts_lora.safetensors
 ┃ ┃ ┣ 📄 SexGod_Nudity_LTX2.safetensors
 ┗ ┗ ┗ 📄 LTX2-i2v-SexyMove.safetensors

QWEN Models (Only qwen workflow version)

Notes for QWEN models:
The text encoder models have two models, the base model and a .mmproj model. Just keep the .mmproj file in the same folder as the base model.

File Structure for QWEN Image Edit AIO models:

Code:

📁 ComfyUI/
 ┣ 📁 models/
 ┃ ┣ 📁 unet/
 ┃ ┃ ┗ 📄 QWEN_Main_GGUF.gguf
 ┃ ┣ 📁 text_encoders/
 ┃ ┃ ┣ 📄 QWEN_Text_Encoder.gguf
 ┃ ┃ ┗ 📄 QWEN_Text_Encoder.mmproj.gguf or .mmproj.safetensors
 ┃ ┣ 📁 vae/
 ┗ ┗ ┗ 📄 Qwen_Image-VAE.safetensors

Video Masking Models (Both workflows)
The workflow uses three types of video masking techniques. MatAnyone, SeC4B, and SAM3. MatAnyone and SAM3 are set up where you will not need to manually download these models. The only one which you’ll need to download is SeC4B.

Video Masking Models

Notes for Video Masking Models:
It’s not required to download all three video masking models. There is a switch node which depending on which version you want to use, it will only load that specific video masking model and ignore the others.

The SeC-4B model has fp32, fp16, bf16, and fp8 model sizes. I suggest using either fp16 or bf16. The fp32 is a bit too large and doesn’t provide a huge difference in quality. The fp8 model is severly worse compared to the fp16 and bf16 models and the nodes do not support this model.

Settings for the Workflow

LTX2 Model Loaders (Both workflows)

These loaders are self explanatory except the Clip loader. You will need to load both the Gemma model and the embeddings connector model in the dual clip loader.

QWEN Model Loaders

These loaders are self explanatory.

Settings Area
LTX2 Settings (Both workflows)

LTX2 handles more resolutions, frame rates, and lengths than WAN VACE.

VideoHelperSuite node auto-adjusts input width, height, and total seconds to fit LTX2 needs. Resolutions snap to multiples of 32. FPS adjusts in steps of 8 plus base 1. Initial tests show 20 seconds at 24 FPS works well. No full stress test yet on ideal lengths.

Grow mask expands video mask coverage area.

Masking Option Selector picks from three methods: MatAnyone, SeC4B, SAM3. MatAnyone is lightest but struggles when masked objects vanish and return. SeC4B is reliable for persistent objects even if off-screen, though heavier. SAM3 is slow now but supports text prompts for masking.

Section Video For Masking lets you pick multiple spots in the input video for selective masking. The values are in seconds. Useful with MatAnyone for objects that disappear from frames.

LTX2 Lora Loaders handle model-specific loras. Include distilled lora if using dev model. Mix NSFW loras freely as needed.

QWEN Settings

The QWEN workflow includes QWEN model loaders and QWEN-specific settings alongside LTX2 settings.

Model loaders are straightforward. QWEN settings are clearly labelled. Keep them at default values unless you have a reason to adjust. Change the QWEN prompt freely if needed.

The QWEN NSFW AIO model already includes its speed lora. Do not add that lora again in the lora loaders.

Running the Workflow

After adjusting settings, click Run in the top right.
The workflow loads models and your input video.
If you chose SeC4B or MatAnyone, it pauses after an image appears in the Preview Bridge (Image) node.
If you chose SAM3, it runs fully without needing manual mask input.

Right-click the Preview Bridge (Image) node and choose Open in MaskEditor Image Canvas. Draw your mask. Run the workflow again. It will build the video mask and generate the final output.