🟣 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

1. 🖼️ 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.

2. 🔄 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)

1. Load LTX25-I2VA-Qwen3.5.json

2. Upload your first-frame image to image

3. Select preset 🎬 LTX 2.5 NSFW I2V

4. Write what happens next (in any language)

5. Generate animated video with synchronized audio

First-Last-Frame (FL2VA)

1. Load LTX25-FL2VA-Qwen3.5.json

2. Upload first-frame to image, last-frame to image2, set frame_count=1

3. Select preset 🔄 LTX 2.5 NSFW FL2V

4. Describe the transition between the two frames

5. 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

1. Download: ComfyUI-QwenVL-Mod (latest version)

2. Extract to ComfyUI/custom_nodes/ComfyUI-QwenVL-Mod

3. Install requirements: pip install -r requirements.txt

4. Restart ComfyUI

5. 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