<h2 id="overview">🔥 Overview</h2>This workflow enables stable, coherent image‑to‑video generation using Kandinsky5 Lite I2V, even on GPUs with only 4GB of VRAM (e.g., RTX 3050 Ti mobile). Through targeted optimizations (SAGEAttention, 4‑step LoRA, Hunyuan Video VAE, SD3 sampling, latent normalization), the workflow achieves impressive performance:<ul><li>5‑second video duration</li><li>675×900 resolution</li><li>8 steps → ~15 minutes</li><li>12 steps → ~27 minutes on first run, ~23 minutes on subsequent runs</li><li>High visual quality via YUV422p12le (FFV1)</li><li>Consistent motion, no drift, no temporal artifacts</li></ul>This makes Kandinsky5 Lite one of the most efficient video models available for low‑VRAM systems.<h1 id="key-features">✨ Key Features</h1><h3 id="runs-on-4gb-vram-gpus">✔ Runs on 4GB VRAM GPUs</h3>The entire model is offloaded to system RAM while maintaining stability and temporal consistency.<h3 id="optimized-resolution:-675900">✔ Optimized Resolution: 675×900</h3>Balanced resolution that preserves detail while keeping latent sizes manageable.<h3 id="professional-output-codec">✔ Professional Output Codec</h3>Videos are exported using FFV1 MKV – YUV422p12le, ideal for editing, grading, and upscaling.<h3 id="natural-motion-and-high-coherence">✔ Natural Motion &amp; High Coherence</h3>Achieved through:<ul><li><code>NormalizeVideoLatentStart</code></li><li><code>ReplaceVideoLatentFrames</code></li><li><code>Kandinsky5ImageToVideo</code></li><li><code>PathchSageAttentionKJ</code></li><li>Hunyuan 1.5 4‑step LoRA</li></ul><h3 id="advanced-negative-prompt">✔ Advanced Negative Prompt</h3>Prevents anatomical drift, motion loops, frame duplication, and temporal artifacts.<h1 id="model-components">📦 Model Components</h1>From the workflow:<ul><li>UNET: <code>kandinsky5lite_i2v_5s.safetensors</code></li><li>VAE: <code>hunyuan_video_vae_bf16.safetensors</code></li><li>Text Encoders:<ul><li><code>qwen_2.5_vl_7b_fp8_scaled.safetensors</code></li><li><code>clip_l.safetensors</code></li></ul></li><li>LoRA: <code>hunyuanvideo1.5_t2v_480p_lightx2v_4step_lora_rank_32_bf16.safetensors</code></li></ul><h1 id="hardware-requirements">🧩 Hardware Requirements</h1><ul><li>GPU: 4GB VRAM (tested on RTX 3050 Ti mobile)</li><li>RAM: 64GB recommended</li><li>CPU: any CPU capable of heavy RAM offload</li><li>ComfyUI 0.3.68+</li><li>KJ Nodes + VideoHelperSuite</li></ul><h1 id="realworld-performance">🚀 Real‑World Performance</h1><h3 id="8-steps">8 steps</h3><ul><li>~15 minutes</li><li>Good motion stability</li><li>Ideal for quick tests</li></ul><h3 id="12-steps">12 steps</h3><ul><li>~27 minutes (first run)</li><li>~23 minutes (subsequent runs)</li><li>Higher detail</li><li>More stable motion</li><li>Cleaner temporal consistency</li></ul><h1 id="output-specifications">📁 Output Specifications</h1><ul><li>Format: MKV</li><li>Codec: FFV1</li><li>Color: YUV422p12le</li><li>FPS: 24</li><li>Duration: 5 seconds</li><li>Resolution: 675×900</li></ul>

Files with SHA256: 5b6fcdd7a82265761ee0ac745fef43c684599b182ac70c52fab66f64e44da01a