<h3 id="deploy-on-runpod"><a target="_blank" rel="ugc" href="https://console.runpod.io/hub/template/06xkj4x3dt?ref=ioakclrv">🟣 Deploy on Runpod</a></h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b73645e5-d236-4282-b9f6-0470c01b5e6c/original=true/b73645e5-d236-4282-b9f6-0470c01b5e6c.jpeg" /><h3 id="deploy-on-vast.ai"><a target="_blank" rel="ugc" href="https://cloud.vast.ai/?ref_id=188688&amp;creator_id=188688&amp;name=OneClick%20-%20ComfyUI%20-%20LTX%202.3%20Uncensored%20-%20Qwen3.5%20AutoPrompt">🟡</a><a target="_blank" rel="ugc" href="https://cloud.vast.ai/?ref_id=188688&amp;creator_id=188688&amp;name=OneClick%20-%20ComfyUI%20-%20LTX%202.3%20Uncensored%20-%20Qwen3.5%20AutoPrompt"> Deploy on Vast.ai</a></h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/86023401-62bf-461b-ad31-8382e3b99c0e/original=true/86023401-62bf-461b-ad31-8382e3b99c0e.jpeg" />ComfyUI-QwenVL-Mod — Enhanced Vision-Language with LTX 2.3 Version 2.3.1 (2026/08/25) — 🎬 LTX 2.3 Native Video+Audio + 10Eros Uncensored + Qwen3.5 Auto-Prompting<hr /><h2 id="requirements-read-first!">⚠️ Requirements — Read First!</h2><h3 id="gpu-and-vram">GPU &amp; VRAM</h3><ul><li>🟢 Recommended — RTX 5090 / 4090 (24 GB+) → FP8 text encoder + checkpoint FP8 → Fast, best quality</li><li>🟡 Enthusiast — RTX 3090 / 4080 (16-24 GB) → offload text encoder / VAE → Slower, good quality</li><li>🟠 Budget — RTX 4060 Ti 16 GB → aggressive offload + lower resolution → Experimental, slow</li></ul><blockquote>12 GB GPUs: Not recommended for LTX 2.3 with the full 10Eros setup. The text encoder alone is ~12.8 GB FP8.</blockquote><h3 id="software">Software</h3><ul><li>ComfyUI: v0.32.0+ (required for LTX 2.3 native audio and Sage Attention flags)</li><li>Python: 3.10+</li><li>CUDA: 13.0 recommended</li><li>Storage: 100 GB+ SSD (models ~40 GB + outputs)</li></ul><h3 id="qwen3-vl-prompt-enhancer">Qwen3-VL Prompt Enhancer</h3><ul><li>GGUF: Q4_K_S (~4.8 GB) or Q5_K_S (~5.5 GB) for 8B model</li><li>HF: Qwen3-VL-8B-Heretic-Stable (~16 GB) or Qwen3-VL-4B (~8 GB)</li></ul><hr /><h2 id="what-is-comfyui-qwenvl-mod">🌟 What is ComfyUI-QwenVL-Mod?</h2>A powerful enhanced vision-language node for ComfyUI that combines Qwen3-VL models with LTX 2.3 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 2.3!"<hr /><h2 id="key-features">🎬 Key Features</h2><h3 id="ltx-2.3-native-video+audio-generation">🚀 LTX 2.3 Native Video+Audio Generation</h3><ul><li>I2VA (Image-to-Video+Audio): Animate a first-frame image with synchronized audio</li><li>FL2VA (First-Last-Frame): Generate the transition between two keyframes with audio — Qwen3-VL sees both frames</li></ul>LTX 2.3 produces video and audio in the same forward pass: no separate MMAudio or audio-VAE stitching is required.<h3 id="qwen3.5-auto-prompting">🧠 Qwen3.5 Auto-Prompting</h3><ul><li>Multilingual: Write your prompt in any language — Qwen3-VL translates and converts it</li><li>Auto-format: Generates optimized LTX 2.3 prompt structure</li><li>Multi-reference: Qwen3-VL sees all connected images via <code>image</code> + <code>image2</code> inputs</li><li>Visual style detection: 12+ artistic styles (photorealistic, cinematic, anime, 3D CG, claymation, vintage film, watercolor, fantasy, etc.)</li><li>Smart caching: Performance optimization with Fixed Seed Mode</li><li>GGUF backend: Efficient local model inference with quantization support</li><li>Qwen3.5 support: Thinking mode disabled via <code>/no_think</code> for fast prompt generation</li></ul><h3 id="native-synchronized-audio">🔊 Native Synchronized Audio</h3><ul><li>No separate audio node needed — LTX 2.3 generates video and audio jointly</li><li>Voice, sound effects, ambience and music modeled together</li><li>Describe sounds and dialogue in your prompt and the model generates them natively</li></ul><h3 id="nsfw-support">🎨 NSFW Support</h3><ul><li>Comprehensive content generation without restrictions</li><li>Dedicated NSFW presets tuned for LTX 2.3 I2V and FL2V</li><li>Natural progression, consistent characters, explicit diegetic soundscape</li></ul><hr /><h2 id="what's-included-2-workflows">📦 What's Included — 2 Workflows</h2><ol><li>🖼️ I2VA — <code>LTX23-I2VA-Qwen3.5.json</code> — text + first-frame image (<code>image</code>) — Image-to-video+audio. 10Eros v1.5 + DMD hybrid v2 LoRA, dual-stage with spatial upscaler.</li><li>🔄 FL2VA — <code>LTX23-FL2VA-Qwen3.5.json</code> — text + first-frame (<code>image</code>) + last-frame (<code>image2</code>) — 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.</li></ol><blockquote>Both workflows include TensorRT upscaling (RealESRGAN x4) and RIFE frame interpolation.</blockquote><hr /><h2 id="multi-reference-input-(image2)">🖼️ Multi-Reference Input (image2)</h2>The QwenVL-Mod node has two image inputs:<ul><li>I2VA: <code>image</code> = first frame</li><li>FL2VA: <code>image</code> = first frame, <code>image2</code> = last frame, <code>frame_count = 1</code></li></ul><blockquote>Qwen3-VL sees all connected images as individual images (not as a video sequence), enabling proper multi-reference analysis for FL2VA.</blockquote><hr /><h2 id="qwenvl-mod-nsfw-presets">🎯 QwenVL-Mod NSFW Presets</h2>The workflows include built-in NSFW presets for the Qwen3-VL prompt enhancer:<ul><li><code>🎬 LTX 2.3 NSFW I2V</code> — Image-to-video with mandatory audio instructions</li><li><code>🔄 LTX 2.3 NSFW FL2V</code> — First-last-frame transition with audio instructions</li></ul><h3 id="what-the-presets-produce">What the presets produce</h3><ul><li>🎬 I2VA: Rich scene description, camera motion, subject action, explicit diegetic soundscape, non-diegetic music flag</li><li>🔄 FL2VA: Describes the transition path between first and last frames (not the scene — images fix the scene). Favors single continuous shot.</li><li>All presets: smooth, continuous camera motion (no abrupt or stepped changes), explicit audio notes</li></ul><blockquote>SFW presets are also available. Edit the preset dropdown in the QwenVL node to switch.</blockquote><hr /><h2 id="usage-examples">🎮 Usage Examples</h2><h3 id="image-to-video-(i2va)">Image-to-Video (I2VA)</h3><ol><li>Load <code>LTX23-I2VA-Qwen3.5.json</code></li><li>Upload your first-frame image to <code>image</code></li><li>Select preset <code>🎬 LTX 2.3 NSFW I2V</code></li><li>Write what happens next (in any language)</li><li>Generate animated video with synchronized audio</li></ol><h3 id="first-last-frame-(fl2va)">First-Last-Frame (FL2VA)</h3><ol><li>Load <code>LTX23-FL2VA-Qwen3.5.json</code></li><li>Upload first-frame to <code>image</code>, last-frame to <code>image2</code>, set <code>frame_count=1</code></li><li>Select preset <code>🔄 LTX 2.3 NSFW FL2V</code></li><li>Describe the transition between the two frames</li><li>Generate the interpolated video with TensorRT upscale + RIFE</li></ol><hr /><h2 id="technical-specifications">🔧 Technical Specifications</h2><h3 id="performance">⚡ Performance</h3><ul><li>Output: Native LTX 2.3 resolution, 24 fps</li><li>Audio: Native synchronized stereo audio</li><li>Upscale: TensorRT RealESRGAN x4 (both workflows)</li><li>Frame interpolation: RIFE rife49 → higher fps (both workflows)</li><li>Sage Attention: FP16 accumulation, async offload</li><li>Smart caching: Reuse prompts with same inputs, Fixed Seed Mode for text-only caching</li></ul><h3 id="model-support">🎨 Model Support</h3><ul><li>Qwen3-VL 4B: 7 GGUF variants (2.38 GB – 4.28 GB)</li><li>Qwen3-VL 8B: 7 GGUF variants (4.8 GB – 8.71 GB)</li><li>Qwen3.5: 4B / 9B / 27B (uncensored, heretic, unsloth) — thinking mode disabled</li><li>HF Models: Josiefed, official, Heretic-Stable variants</li><li>Quantization: Q4_K_S, Q5_K_S, FP16, INT8</li></ul><h3 id="multilingual-capabilities">🌐 Multilingual Capabilities</h3><ul><li>Input languages: Any language supported</li><li>Auto-translation: Automatic translation to optimized English</li><li>Style detection: Works with multilingual prompts</li><li>Cultural adaptation: Context-aware prompt enhancement</li></ul><hr /><h2 id="installation">📦 Installation</h2><h3 id="quick-install">Quick Install</h3><ol><li>Download: <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-QwenVL-Mod">ComfyUI-QwenVL-Mod</a> (latest version)</li><li>Extract to <code>ComfyUI/custom_nodes/ComfyUI-QwenVL-Mod</code></li><li>Install requirements: <code>pip install -r requirements.txt</code></li><li>Restart ComfyUI</li><li>Load included workflows from <code>ltx/23/</code> folder</li></ol><h3 id="custom-nodes-required">Custom Nodes Required</h3><ul><li>ComfyUI-QwenVL-Mod — All workflows (Qwen3-VL prompt enhancer) — <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-QwenVL-Mod">huchukato/ComfyUI-QwenVL-Mod</a></li><li>ComfyUI-LTXVideo — LTX 2.3 native nodes — <a target="_blank" rel="ugc" href="https://github.com/Lightricks/ComfyUI-LTXVideo">Lightricks/ComfyUI-LTXVideo</a></li><li>ComfyUI-RIFE-TensorRT-Auto — FL2VA, I2VA (frame interpolation) — <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-RIFE-TensorRT-Auto">huchukato/ComfyUI-RIFE-TensorRT-Auto</a></li><li>ComfyUI-Upscaler-TensorRT-Auto — FL2VA, I2VA (upscaling) — <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-Upscaler-TensorRT-Auto">huchukato/ComfyUI-Upscaler-TensorRT-Auto</a></li><li>ComfyUI-VideoHelperSuite — (VHS_VideoCombine) — <a target="_blank" rel="ugc" href="https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite">Kosinkadink/ComfyUI-VideoHelperSuite</a></li><li>ComfyUI-Easy-Use — (easy showAnything) — <a target="_blank" rel="ugc" href="https://github.com/yolain/ComfyUI-Easy-Use">yolain/ComfyUI-Easy-Use</a></li><li>comfyui-find-perfect-resolution — All workflows (ResolutionSelector) — <a target="_blank" rel="ugc" href="https://github.com/ashtar1984/comfyui-find-perfect-resolution">ashtar1984/comfyui-find-perfect-resolution</a></li><li>was-node-suite-comfyui — (ComfyMathExpression) — <a target="_blank" rel="ugc" href="https://github.com/ltdrdata/was-node-suite-comfyui">ltdrdata/was-node-suite-comfyui</a></li></ul><h3 id="models-required">Models Required</h3>LTX 2.3 I2VA / FL2VA use the 10Eros v1.5 uncensored setup:<ul><li><code>models/checkpoints/</code> → <code>10Eros_v1.5_fp8mixed_experimental_learned.safetensors</code> (~30 GB) — <a target="_blank" rel="ugc" href="https://huggingface.co/LokkenJP/10EROS_1.5_fp8_exp_learned">LokkenJP/10EROS_1.5_fp8_exp_learned</a></li><li><code>models/text_encoders/</code> → <code>gemma-3-12b-it-ablit-norms-biproj-fp8mixed.safetensors</code> (~12.8 GB) — <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3-10Eros/resolve/main/text_encoders/gemma-3-12b-it-ablit-norms-biproj-fp8mixed.safetensors">TenStrip/LTX2.3-10Eros</a></li><li><code>models/loras/ltx23/</code> → <code>LTX2.3_DMD_hybrid_v2.safetensors</code> (~662 MB) — <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3_DMD_Lora/resolve/main/LTX2.3_DMD_hybrid_v2.safetensors">TenStrip/LTX2.3_DMD_Lora</a></li><li><code>models/latent_upscale_models/</code> → <code>ltx-2.3-spatial-upscaler-x2-1.1.safetensors</code> (~996 MB) — <a target="_blank" rel="ugc" href="https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-spatial-upscaler-x2-1.1.safetensors">Lightricks/LTX-2.3</a></li><li><code>models/latent_upscale_models/</code> → <code>ltx-2.3-temporal-upscaler-x2-1.0.safetensors</code> (~262 MB) — <a target="_blank" rel="ugc" href="https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-temporal-upscaler-x2-1.0.safetensors">Lightricks/LTX-2.3</a></li><li><code>models/vae/</code> → <code>pruna_ltx2.3_vae_comfy_bf16.safetensors</code> (~500 MB) — <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/LTX2.3_comfy">Kijai/LTX2.3_comfy</a></li></ul>Qwen3-VL Prompt Enhancer<ul><li><code>models/LLM/</code> → <code>Qwen3-VL-8B-Heretic-Stable</code> (GGUF or HF)</li></ul>TensorRT Engines<ul><li><code>models/upscale_models/</code> → <code>RealESRGAN_x4</code> (TensorRT engine)</li><li><code>models/rife/</code> → <code>rife49_ensemble_True_scale_1_sim</code> (TensorRT engine)</li></ul><blockquote>TensorRT engines must be built for your specific GPU. See <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-RIFE-TensorRT-Auto">ComfyUI-RIFE-TensorRT-Auto</a> and <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-Upscaler-TensorRT-Auto">ComfyUI-Upscaler-TensorRT-Auto</a> for build instructions.</blockquote><hr /><h2 id="docker-cloud-ready">🐳 Docker / Cloud Ready</h2><h3 id="oneclick-runpod-template">OneClick RunPod Template</h3>Prefer a ready-to-go environment? Use the OneClick - ComfyUI - LTX 2.3 Uncensored - CU13 RunPod template:<ul><li>Docker image: <code>huchukato/comfyui-qwenvl-runpod:cu13-ltx23</code></li><li>Base: <code>runpod/comfyui:cuda13.0</code></li><li>All custom nodes pre-installed</li><li>Both workflows auto-downloaded at boot</li><li>Models auto-downloaded at first boot (~40 GB, persistent)</li><li>ComfyUI v0.32.0+ forced at boot</li><li>Sage Attention, FP16 accumulation, async offload</li><li>TensorRT upscaling + RIFE frame interpolation</li></ul><h3 id="comfyui-args-(pre-configured)">ComfyUI Args (pre-configured)</h3><pre><code>--disable-auto-launch
--fast fp16_accumulation
--use-sage-attention
--cuda-malloc
--async-offload
</code></pre><hr /><h2 id="why-choose-comfyui-qwenvl-mod-+-ltx-2.3">🚀 Why Choose ComfyUI-QwenVL-Mod + LTX 2.3?</h2><h3 id="for-content-creators">🎬 For Content Creators</h3><ul><li>Native audio: Video and audio in one pass — no separate audio node</li><li>Multilingual: Write in any language, Qwen3-VL handles translation</li><li>Professional: Smooth camera vocabulary and natural motion</li><li>Quality: 10Eros v1.5 uncensored + TensorRT upscale</li></ul><h3 id="for-nsfw-content">🔥 For NSFW Content</h3><ul><li>Explicit: Uncensored generation with dedicated NSFW presets</li><li>Detailed: Rich scene descriptions with explicit diegetic soundscape</li><li>Natural: Realistic progression, consistent characters</li><li>Audio: Native moans, breaths, skin contact, ambient sounds</li></ul><h3 id="for-power-users">⚡ For Power Users</h3><ul><li>Customizable: Easy to modify presets and system prompts</li><li>Extendable: Add your own Qwen3-VL models (GGUF or HF)</li><li>Integrable: Works with existing ComfyUI setups</li><li>Optimized: Sage Attention, FP16, async offload, smart caching</li><li>Multi-reference: <code>image2</code> input for FL2VA workflow</li></ul><hr /><h2 id="what-makes-this-special">🌟 What Makes This Special?</h2><ul><li>First: Complete LTX 2.3 uncensored workflow pack with Qwen3-VL auto-prompting</li><li>Native audio: No separate audio node — LTX 2.3 does it all</li><li>2 workflows: I2VA + FL2VA covering the main LTX 2.3 use cases</li><li>Multi-reference: Qwen3-VL sees all connected images (not just the first)</li><li>TensorRT: Built-in upscaling and frame interpolation</li><li>NSFW presets: Dedicated presets tuned for LTX 2.3</li><li>Multilingual: Any input language, auto-translated and formatted</li><li>Ready: Works out-of-the-box with included workflows</li></ul><hr /><h2 id="what's-new-in-v2.3.1">🎯 What's New in v2.3.1</h2><h3 id="ltx-2.3-10eros-setup">✅ LTX 2.3 10Eros Setup</h3><ul><li>Switched to 10Eros v1.5 uncensored checkpoint</li><li>Added TenStrip FP8 abliterated text encoder (<code>gemma-3-12b-it-ablit-norms-biproj-fp8mixed.safetensors</code>)</li><li>Removed obsolete Gemma 3 FP4 + abliterated LoRA combo</li><li>Added DMD hybrid v2 LoRA for fast sampling</li></ul><h3 id="native-audio-workflows">✅ Native Audio Workflows</h3><ul><li>Both I2VA and FL2VA workflows generate synchronized native audio</li><li>Presets include mandatory audio instructions</li></ul><h3 id="qwen3.5-rename">✅ Qwen3.5 Rename</h3><ul><li>Workflow filenames updated from <code>Qwen3VL</code> to <code>Qwen3.5</code> for consistency</li></ul>

Files with SHA256: a835e03e5cfa28d747e8ff63ba5c63cf49f9ce015680fc44d46435f448f794bf