🟣 Deploy on Runpod
🟡 Deploy on Vast.ai

ComfyUI-QwenVL-Mod — Enhanced Vision-Language with LTX 2.5 Version 2.5.1 (2026/08/25) — 🎬 LTX 2.5 Native Video+Audio + DeepNeuralNerd Uncensored Gemma 4 + Qwen3.5 Auto-Prompting

⚠️ Requirements — Read First!

GPU & VRAM

🟢 Recommended — RTX 5090 / 4090 (24 GB+) → INT8 ConvRot transformer + text encoder → Fast, best quality
🟡 Enthusiast — RTX 3090 / 4080 (16-24 GB) → offload text encoder / VAE → Slower, good quality
🟠 Budget — RTX 4060 Ti 16 GB → aggressive offload + lower resolution → Experimental, slow

12 GB GPUs: Not recommended for LTX 2.5 with the full DeepNeuralNerd setup. The text encoder alone is ~13.2 GB INT8.

Software

ComfyUI: v0.34.2 (required for LTX 2.5 native audio, UNETLoader, and Sage Attention flags)
Python: 3.10+
CUDA: 13.0 recommended
Storage: 120 GB+ SSD (models ~50 GB + outputs)

Qwen3-VL Prompt Enhancer

GGUF: Q4_K_S (~4.8 GB) or Q5_K_S (~5.5 GB) for 8B model
HF: Qwen3-VL-8B-Heretic-Stable (~16 GB) or Qwen3-VL-4B (~8 GB)

🌟 What is ComfyUI-QwenVL-Mod?

A powerful enhanced vision-language node for ComfyUI that combines Qwen3-VL models with LTX 2.5 video generation workflows. Features multilingual support, visual style detection, native synchronized audio, and NSFW capabilities for professional AI content creation.

Think: "Your all-in-one solution for intelligent prompt enhancement and uncensored video+audio generation with LTX 2.5!"

🎬 Key Features

🚀 LTX 2.5 Native Video+Audio Generation

I2VA (Image-to-Video+Audio): Animate a first-frame image with synchronized audio
FL2VA (First-Last-Frame): Generate the transition between two keyframes with audio — Qwen3-VL sees both frames

LTX 2.5 produces video and audio in the same forward pass: no separate MMAudio or audio-VAE stitching is required.

🧠 Qwen3.5 Auto-Prompting

Multilingual: Write your prompt in any language — Qwen3-VL translates and converts it
Auto-format: Generates optimized LTX 2.5 prompt structure
Multi-reference: Qwen3-VL sees all connected images via image + image2 inputs
Visual style detection: 12+ artistic styles (photorealistic, cinematic, anime, 3D CG, claymation, vintage film, watercolor, fantasy, etc.)
Smart caching: Performance optimization with Fixed Seed Mode
GGUF backend: Efficient local model inference with quantization support
Qwen3.5 support: Thinking mode disabled via /no_think for fast prompt generation

🔊 Native Synchronized Audio

No separate audio node needed — LTX 2.5 generates video and audio jointly
Voice, sound effects, ambience and music modeled together
Describe sounds and dialogue in your prompt and the model generates them natively

🎨 NSFW Support

Comprehensive content generation without restrictions
Dedicated NSFW presets tuned for LTX 2.5 I2V and FL2V
Natural progression, consistent characters, explicit diegetic soundscape

📦 What's Included — 2 Workflows

🖼️ I2VA/T2VA — LTX25-I2VA-T2VA-Qwen3.5.json — text + optional first-frame image (image) — Image-to-Video+Audio or Text-to-Video+Audio. A boolean toggle switches between I2VA (with image input) and T2VA (text-only). DeepNeuralNerd uncensored Gemma 4 text encoder + Gemma 4 E2B prompt enhancer, dual-stage with latent spatial upscaler.
🔄 FL2VA — LTX25-FL2VA-Qwen3.5.json — text + first-frame (image) + last-frame (image2) — First-Last-Frame to video with audio. Qwen3-VL sees both frames and describes the transition. Includes TensorRT upscale + RIFE frame interpolation for high-resolution output.

I2VA and FL2VA include TensorRT upscaling (RealESRGAN x4) and RIFE frame interpolation.

🖼️ Multi-Reference Input (image2)

The QwenVL-Mod node has two image inputs:

I2VA: image = first frame
FL2VA: image = first frame, image2 = last frame, frame_count = 1

Qwen3-VL sees all connected images as individual images (not as a video sequence), enabling proper multi-reference analysis for FL2VA.

🎯 QwenVL-Mod NSFW Presets

The workflows include built-in NSFW presets for the Qwen3-VL prompt enhancer:

🎬 LTX 2.5 NSFW I2V — Image-to-video with mandatory audio instructions
🔄 LTX 2.5 NSFW FL2V — First-last-frame transition with audio instructions

What the presets produce

🎬 I2VA: Rich scene description, camera motion, subject action, explicit diegetic soundscape, non-diegetic music flag
🔄 FL2VA: Describes the transition path between first and last frames (not the scene — images fix the scene). Favors single continuous shot.
All presets: smooth, continuous camera motion (no abrupt or stepped changes), explicit audio notes

SFW presets are also available. Edit the preset dropdown in the QwenVL node to switch.

🎮 Usage Examples

Image-to-Video (I2VA)

Load LTX25-I2VA-Qwen3.5.json
Upload your first-frame image to image
Select preset 🎬 LTX 2.5 NSFW I2V
Write what happens next (in any language)
Generate animated video with synchronized audio

First-Last-Frame (FL2VA)

Load LTX25-FL2VA-Qwen3.5.json
Upload first-frame to image, last-frame to image2, set frame_count=1
Select preset 🔄 LTX 2.5 NSFW FL2V
Describe the transition between the two frames
Generate the interpolated video with TensorRT upscale + RIFE

🔧 Technical Specifications

⚡ Performance

Output: Native LTX 2.5 resolution, 24 fps
Audio: Native synchronized stereo audio
Upscale: TensorRT RealESRGAN x4 (both workflows) + latent spatial upscaler x2 (I2VA)
Frame interpolation: RIFE rife49 → higher fps (both workflows)
Sage Attention: FP16 accumulation, async offload
Smart caching: Reuse prompts with same inputs, Fixed Seed Mode for text-only caching

🎨 Model Support

Qwen3-VL 4B: 7 GGUF variants (2.38 GB – 4.28 GB)
Qwen3-VL 8B: 7 GGUF variants (4.8 GB – 8.71 GB)
Qwen3.5: 4B / 9B / 27B (uncensored, heretic, unsloth) — thinking mode disabled
HF Models: Josiefed, official, Heretic-Stable variants
Quantization: Q4_K_S, Q5_K_S, FP16, INT8

🌐 Multilingual Capabilities

Input languages: Any language supported
Auto-translation: Automatic translation to optimized English
Style detection: Works with multilingual prompts
Cultural adaptation: Context-aware prompt enhancement

📦 Installation

Quick Install

Download: ComfyUI-QwenVL-Mod (latest version)
Extract to ComfyUI/custom_nodes/ComfyUI-QwenVL-Mod
Install requirements: pip install -r requirements.txt
Restart ComfyUI
Load included workflows from ltx/25/ folder

Custom Nodes Required

ComfyUI-QwenVL-Mod — All workflows (Qwen3-VL prompt enhancer) — huchukato/ComfyUI-QwenVL-Mod
ComfyUI-LTXVideo — LTX 2.5 native nodes — Lightricks/ComfyUI-LTXVideo
ComfyUI-RIFE-TensorRT-Auto — FL2VA, I2VA (frame interpolation) — huchukato/ComfyUI-RIFE-TensorRT-Auto
ComfyUI-Upscaler-TensorRT-Auto — FL2VA, I2VA (upscaling) — huchukato/ComfyUI-Upscaler-TensorRT-Auto
ComfyUI-VideoHelperSuite — (VHS_VideoCombine) — Kosinkadink/ComfyUI-VideoHelperSuite
ComfyUI-Easy-Use — (easy showAnything) — yolain/ComfyUI-Easy-Use
comfyui-find-perfect-resolution — All workflows (ResolutionSelector) — ashtar1984/comfyui-find-perfect-resolution
was-node-suite-comfyui — FL2VA, I2VA (GetImageSize, ResizeImageMaskNode) — ltdrdata/was-node-suite-comfyui

Note: ComfyMathExpression is built into ComfyUI core (v0.24.1+) — no custom node needed.

Models Required

LTX 2.5 I2VA / FL2VA use the DeepNeuralNerd uncensored Gemma 4 setup:

models/diffusion_models/ → ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors (~14 GB) — Lightricks/LTX-2.5
models/text_encoders/ → gemma4-12b-uncensored-heretic-ltx2.5-comfy-int8-convrot.safetensors (~13.2 GB) — DeepNeuralNerd/Gemma-4-12B-it-uncensored-heretic-DeepNeuralNerd-LTX_2.5_ComfyUI
models/text_encoders/ → gemma4_e2b_it_bf16.safetensors (~5 GB, I2VA prompt enhancer only) — TrevorJS/gemma-4-E2B-it-uncensored
models/vae/ → ltx-2.5-video-vae-bf16.safetensors (~500 MB) — Lightricks/LTX-2.5
models/vae/ → ltx-2.5-audio-vae-bf16.safetensors (~500 MB) — Lightricks/LTX-2.5
models/latent_upscale_models/ → ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors (~990 MB, I2VA only) — Lightricks/LTX-2.5

Qwen3-VL Prompt Enhancer

models/LLM/ → Qwen3-VL-8B-Heretic-Stable (GGUF or HF)

TensorRT Engines

models/upscale_models/ → RealESRGAN_x4 (TensorRT engine)
models/rife/ → rife49_ensemble_True_scale_1_sim (TensorRT engine)

TensorRT engines must be built for your specific GPU. See ComfyUI-RIFE-TensorRT-Auto and ComfyUI-Upscaler-TensorRT-Auto for build instructions.

🐳 Docker / Cloud Ready

OneClick RunPod Template

Prefer a ready-to-go environment? Use the OneClick - ComfyUI - LTX 2.5 Uncensored - CU13 RunPod template:

Docker image: huchukato/comfyui-qwenvl-runpod:cu13-ltx25
Base: huchukato/comfyui-base:cu130
All custom nodes pre-installed
Both workflows auto-downloaded at boot
Models auto-downloaded at first boot (~50 GB, persistent)
ComfyUI v0.34.2 baked into base image
Sage Attention, FP16 accumulation, async offload
TensorRT upscaling + RIFE frame interpolation

ComfyUI Args (pre-configured)

--disable-auto-launch
--fast fp16_accumulation
--use-sage-attention
--cuda-malloc
--async-offload

🚀 Why Choose ComfyUI-QwenVL-Mod + LTX 2.5?

🎬 For Content Creators

Native audio: Video and audio in one pass — no separate audio node
Multilingual: Write in any language, Qwen3-VL handles translation
Professional: Smooth camera vocabulary and natural motion
Quality: DeepNeuralNerd uncensored Gemma 4 + TensorRT upscale

🔥 For NSFW Content

Explicit: Uncensored generation with dedicated NSFW presets
Detailed: Rich scene descriptions with explicit diegetic soundscape
Natural: Realistic progression, consistent characters
Audio: Native moans, breaths, skin contact, ambient sounds

⚡ For Power Users

Customizable: Easy to modify presets and system prompts
Extendable: Add your own Qwen3-VL models (GGUF or HF)
Integrable: Works with existing ComfyUI setups
Optimized: Sage Attention, FP16, async offload, smart caching
Multi-reference: image2 input for FL2VA workflow

🌟 What Makes This Special?

First: Complete LTX 2.5 uncensored workflow pack with Qwen3-VL auto-prompting
Native audio: No separate audio node — LTX 2.5 does it all
2 workflows: I2VA + FL2VA covering the main LTX 2.5 use cases
Multi-reference: Qwen3-VL sees all connected images (not just the first)
TensorRT: Built-in upscaling and frame interpolation
NSFW presets: Dedicated presets tuned for LTX 2.5
Multilingual: Any input language, auto-translated and formatted
Ready: Works out-of-the-box with included workflows

🎯 What's New in v2.5.1

✅ LTX 2.5 DeepNeuralNerd Uncensored Setup

Switched to DeepNeuralNerd uncensored heretic Gemma 4 12B text encoder (INT8 ConvRot)
Uses LTX 2.5 22B distilled transformer (INT8 ConvRot) from Lightricks
Separate video VAE and audio VAE loaders (LTX 2.5 architecture)
Added Gemma 4 E2B uncensored prompt enhancer (TrevorJS) for I2VA
Added latent spatial upscaler x2 for I2VA dual-stage upscaling

✅ Architecture Changes from LTX 2.3

LTX 2.5 uses UNETLoader + VAELoader (transformer and VAE are separate files)
No CheckpointLoaderSimple — VAE is not embedded in the checkpoint
No DMD LoRA — LTX 2.5 distilled transformer has built-in distillation

✅ Native Audio Workflows

Both I2VA and FL2VA workflows generate synchronized native audio
Presets include mandatory audio instructions

✅ Qwen3.5 Rename

Workflow filenames updated from Qwen3VL to Qwen3.5 for consistency

🙏 Credits

ComfyUI — comfyanonymous/ComfyUI
LTX 2.5 — Lightricks/LTX-2.5
QwenVL-Mod — huchukato/ComfyUI-QwenVL-Mod
Qwen3-VL — Qwen Team / Alibaba
DeepNeuralNerd Uncensored Gemma 4 — DeepNeuralNerd
Gemma 4 E2B Uncensored — TrevorJS
LTXVideo nodes — Lightricks/ComfyUI-LTXVideo
TensorRT RIFE / Upscaler — huchukato
VideoHelperSuite — Kosinkadink
Easy-Use — yolain
find-perfect-resolution — ashtar1984
was-node-suite — ltdrdata

📄 License

Workflows are released under the same license as the underlying models and custom nodes. See each repository for details.

LTX 2.5 model weights: Lightricks/LTX-2.5 — Lightricks Community License.

Gemma 4 model weights: DeepNeuralNerd — see repository for license details.

Built with ❤️ for the ComfyUI community

Files with SHA256: 2d3e5e8a6c1e38dd6c6e11d70b0a1131f762fb26955f49802c51ce3a7fe14cb5