<h1 id="v2">v2 </h1>The workflow allows ultra-fast iterations over different seeds/prompts/images/settings before rendering the final video at full resolution, while removing the need to constantly swap between models (which is ideal for low-RAM setups).<h2 id="how-it-works">How it works</h2>There are 2 stages:<ol><li>Generate a draft version of your video at low resolution, using the high-noise model only.</li><li> Upscale the video latent to render it at your target resolution, using the low-noise model only.</li></ol><h2 id="process">Process</h2><h3 id="stage-1">Stage 1</h3><ul><li>Disable the groups "LN" and "Post".</li><li>Upload your image.</li><li>Write your prompt.</li><li>Set the low-resolution width (typically, 256, 288 or 320px, depending on the image aspect ratio).</li><li>Click "Run".</li></ul><h3 id="stage-2">Stage 2</h3><ul><li>Drop the video draft created at Stage 1 into the ComfyUI canvas.</li><li>Disable the group "HN"; enable the groups "LN" and "Post".</li><li>In the "Load Draft" node, select the same video again.</li><li>Set the high-resolution width (typically, x2 or x2.5 the width set at stage 1).</li><li>Click "Run".</li></ul>⚠️Important constraint: make sure the video width and height are multiples of 16, at both stages of the process, or the final output will be blurred.The workflow was intentionnally made as simple as possible to maximize compatibility. Up to you to adapt it further to your needs.For a simpler but less efficient version, see v1 below.<h1 id="v1">v1</h1>The workflow significantly mitigates the "speed VS quality" dilemma, allowing users with low-end hardware to generate videos in HD resolution nearly twice as fast!<h2 id="how-it-works">How it works</h2>The principle is dumb simple: we run the high-noise model at very low resolution, then upscale the latent before injecting it into the low-noise sampler.Since the original image is also reinjected, with a new Wan wrapper node at the low-noise sampling step, visual details are preserved.<h3 id="limitations">Limitations</h3><ul><li>The motion does lose a little in subtlety, but the speed gain is totally worth it in most of cases.</li><li><s>Not tested on T2V, probably won't work.</s> Works on T2V too. Simply replace both I2V nodes by a EmptyHunyuanLatentVideo node. Thanks to <a target="_blank" rel="ugc" href="https://civarchive.com/user/axicec">@axicec</a> for sorting this out.</li></ul><h2 id="getting-started">Getting Started</h2><ol><li>Replace the models by yours, or follow the links below to download them.</li><li>Install the required custom nodes listed below if they are missing from your installation.</li><li>Load an image and write your prompt.</li><li>Click Run.</li></ol><h2 id="speed-benchmark">Speed benchmark</h2><ul><li>settings: 65 frames, using Q5_K_M I2V models, 4 steps on high-noise with lightx2v 1030, 4 steps on low-noise with lightx2v 1022 and Fun HPS2.1 loras, euler/beta sampler/scheduler on both samplers.</li><li>hardware: RTX 3060 with 12GB VRAM and 32GB RAM.</li></ul><h3 id="768*1152px-(2:3)">768*1152px (2:3)</h3><ul><li>768*1152, no upscale: 20′46″</li><li>256*384 x2 then x1.5: 11′16″ (-46%)</li><li>256*384 x1.5 then x2: 10′48″ (-48%)</li></ul><h3 id="720*1280px-(9:16)">720*1280px (9:16)</h3><ul><li>720*1280, no upscale : 23′19″</li><li>288*512 x2.5 : 15′57″ (-32%)</li><li>288*512 x2 then x1.25: 11′57″ (-49%) &lt;- this is the showcased video</li></ul><h2 id="target-resolution-vs-hardware-requirements">Target resolution VS Hardware Requirements</h2><ul><li>HD : &gt;= 12GB VRAM v</li><li>FHD: &gt;= 16GB VRAM ? (not tested, feedbacks appreciated)</li></ul>Initial sampling resolutions and step settings recommendations are included in the workflow.<h2 id="custom-nodes">Custom Nodes</h2>Required<ul><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">ComfyUI-KJNodes</a></li><li><a target="_blank" rel="ugc" href="https://github.com/ClownsharkBatwing/RES4LYF">RES4LYF</a></li></ul>Optional<ul><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-GIMM-VFI">ComfyUI-GIMM-VFI</a> (for interpolation)</li><li><a rel="ugc" href="https://github.com/rgthree/rgthree-comfy">rgthree</a> (for fast group bypassing) (v2)</li></ul><h2 id="models-used">Models Used</h2>Wan 2.2 14B I2V, Quantized:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main">https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main</a></li></ul>Lightx2v LoRas :<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2361379">https://civarchive.com/models/1585622?modelVersionId=2361379</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2337903">https://civarchive.com/models/1585622?modelVersionId=2337903</a></li></ul>Fun LoRa:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/alibaba-pai/Wan2.2-Fun-Reward-LoRAs/tree/main">https://huggingface.co/alibaba-pai/Wan2.2-Fun-Reward-LoRAs/tree/main</a></li></ul>Edit (v1): the last sampler's scheduler is set on linear_quadratic by default but it should be beta.

Files with SHA256: d12e2a5f2f314dab247d670ecfa94b51223da76adfa7d652b63f40be76256a27