<h2 id="description-(what-this-workflow-is)">Description (What this workflow is)</h2>This is my Wan2.2 I2V → V2V “Video Extend” cascade workflow built around Wan2.2 14B quant (GGUF) High/Low noise pair, designed to generate a first clip, then extend it in multiple stages while keeping motion continuity and prompt control.Core idea: “make a strong first clip → re-inject last-frame → extend again”. You get a practical pipeline for longer videos without turning your generation into a 1-hour single sampling job.<ul><li>Uses GGUF quant 14B High/Low noise models (<code>wan2.2_i2v_high_noise_14B_Q3_K_M.gguf</code> + <code>wan2.2_i2v_low_noise_14B_Q3_K_M.gguf</code>).Wan2.2 I2V Quant 14B Cascade 1…</li><li>Output is saved via VideoHelperSuite (<code>VHS_VideoCombine</code>) at 16 fps, H.264 MP4, CRF 19, yuv420p, metadata on.</li><li>Built around the common Wan2.2 I2V “HighNoise/LowNoise” split that ComfyUI also documents in their official Wan2.2 workflows.</li></ul><hr /><h2 id="what's-inside-(features-&quot;what's-cool-here&quot;)">What’s inside (Features / “what’s cool here”)</h2><h3 id="1)-true-multi-stage-extension-(cascade)">1) True multi-stage extension (cascade)</h3>Instead of trying to brute-force one long clip in a single run, this workflow:<ul><li>generates the start clip,</li><li>extracts/uses the last frame as the seed for the next clip,</li><li>repeats extension multiple times,</li><li>saves each stage as its own MP4 with separate filename prefixes (so you can compare and pick the best stage quickly).</li></ul><h3 id="2)-pinned-&quot;don't-touch-anything-else&quot;-controls">2) Pinned “don’t touch anything else” controls</h3>The key nodes are pinned so you don’t hunt through spaghetti:<ul><li>GGUF loaders for High/Low.Wan2.2 I2V Quant 14B Cascade 1…</li><li>Multiple LoRA loader slots (stack-style) so youWan2.2 I2V Quant 14B Cascade 1…uickly.Wan2.2 I2V Quant 14B Cascade 1…</li><li>Pinned video combine/export nodes per cascade sWan2.2 I2V Quant 14B Cascade 1…</li></ul><h3 id="3)-fast-mode:-lightx2v-logic-(4-step-mindset)">3) Fast mode: LightX2V logic (4-step mindset)</h3>This workflow is designed to support the Light2X / LightX2V low-step style that people use to make Wan2.2 I2V practical. Community workflows show the same “4 steps with Light2X LoRA” approach as the main speed hack.Your practical rule (what you wrote, translated into “how to actually run it”):<ul><li>ON (LightX2V-ish) → low steps, fast iterations.</li><li>OFF (no LightX2V) → you can use random custom LoRAs, but step count jumps to ~20–40 and render time explodes.</li></ul><h3 id="4)-export-quality-defaults-that-don't-sabotage-you">4) Export quality defaults that don’t sabotage you</h3><ul><li>CRF 19 + yuv420p for compatibility.</li><li>16 fps as a stable baseline for Wan2.2 I2V style clips.</li><li>Metadata saving enabled.</li></ul><hr /><h2 id="usage-(how-to-run-it)">Usage (How to run it)</h2><h3 id="0)-required-baseline-(comfyui-side)">0) Required baseline (ComfyUI side)</h3>You need a working Wan2.2 I2V setup (models in the right folders). The official ComfyUI docs list the standard Wan2.2 model components and structure.<h3 id="1)-load-workflow">1) Load workflow</h3>Drag &amp; drop the JSON into ComfyUI.<h3 id="2)-set-the-models-(gguf-quant)">2) Set the models (GGUF quant)</h3>Select your HighNoise + LowNoise 14B GGUF in the pinned loaders. The workflow is already pointed at:<ul><li><code>wan2.2_i2v_high_noise_14B_Q3_K_M.gguf</code></li><li><code>wan2.2_i2v_low_noise_14B_Q3_K_M.gguf</code>Wan2.2 I2V Quant 14B Cascade 1…</li></ul><h3 id="3)-pick-your-speed-strategy">3) Pick your speed strategy</h3>A) Fast motion mode (recommended for iteration)<ul><li>Keep the LightX2V approach (4-step style). This is the same core trick people use in other Wan2.2 workflows to get “minutes not hours.”Wan2.2 I2V Quant 14B Cascade 1…RA freedom mode**</li><li>Disable LightX2V and run your own LoRAs (character/style/etc).</li><li>Expect 20–40 steps territory and much slower total runtime (your note matches what most people experience: low-step LoRA tricks exist specifically to avoid that).</li></ul><h3 id="4)-prompt-like-you-mean-it-(this-workflow-is-prompt-driven)">4) Prompt like you mean it (this workflow is prompt-driven)</h3>For short clips (1.5–3s), it behaves best. For longer clips (6s+), prompt precision becomes mandatory:<ul><li>Write movement instructions explicitly (camera + subject motion).</li><li>Write the ending motion/pose explicitly (you need to know where the clip ends).</li><li>Treat each cascade stage as its own “shot continuation” and prompt accordingly.</li></ul><h3 id="5)-run-cascade-stages">5) Run cascade stages</h3>Each stage outputs a separate MP4 via <code>VHS_VideoCombine</code> nodes (you’ll see filename prefixes for the extend steps).<hr /><h2 id="settings-(what-to-tweak-first)">Settings (What to tweak first)</h2><h3 id="video-export-(safe-defaults-already-set)">Video export (safe defaults already set)</h3><ul><li>frame_rate: 16</li><li>format: video/h264-mp4</li><li>pix_fmt: yuv420p</li><li>crf: 19</li><li>save_metadata: true</li></ul>If you want smaller files: increase CRF (23–28). If you want higher compatibility: keep yuv420p.<h3 id="lora-slots-(fast-switching)">LoRA slots (fast switching)</h3>This workflow is set up with multiple LoRA loader nodes so you can keep options “parked” and toggle quickly.Wan2.2 I2V Quant 14B Cascade 1…Also included: an AccVid LoRA option is present in the file (<code>Wan21_AccVid_T2V_14B_lora_rank32_fp16.safetensors</code>).Wan2.2 I2V Quant 14B Cascade 1…<hr /><h2 id="troubleshooting-(the-stuff-that-actually-breaks)">Troubleshooting (the stuff that actually breaks)</h2><h3 id="&quot;my-custom-lora-doesn't-work-with-lightx2v&quot;">“My custom LoRA doesn’t work with LightX2V”</h3>That’s expected in practice: LightX2V-style motion LoRAs often don’t like being mixed with other LoRAs in the same run (behavior varies by LoRA and how it was trained). Community workflows frequently treat Light2X as the LoRA for the low-step motion trick, nWan2.2 I2V Quant 14B Cascade 1…k casually. Fix:<ul><li>Disable LightX2V → use custom LoRAs → accept higher steps.</li></ul><h3 id="&quot;wit">“Wit</h3>Wan2.2 I2V Quant 14B Cascade 1…s forever”Yes. That’s literally why Light2X/LightX2V workflows exist: to avoid high-step runtime.Fix:<ul><li>Keep clips shorter (1.5–3s) and extend via cascade.</li><li>Iterate on prompts with fast mode first, then do one “final” higher-step run only when you already know the motion.</li></ul><h3 id="&quot;my-extended-stage-drifts-changes-subject&quot;">“My extended stage drifts / changes subject”</h3>Common for I2V extension pipelines when the prompt is vague. Fix:<ul><li>Add anchor descriptors (identity, outfit, environment, lighting) every stage.</li><li>Add explicit motion endpoints (final pose/camera state).</li></ul><h3 id="&quot;video-won't-play-on-my-phone&quot;">“Video won’t play on my phone”</h3>Use the default export:<ul><li>H.264 MP4, yuv420p, CRF 19.</li></ul><hr /><h2 id="faq">FAQ</h2>Q: What is this workflow optimized for? A: Practical I2V generation + multi-stage extension on Wan2.2 14B quant GGUF, with exports that just work.Q: Can I use other Wan2.2 quant files (Q4/Q5/Q8, etc.)? A: Yes — swap the GGUF model selections in the pinned High/Low loaders. The pipeline logic stays the same.Wan2.2 I2V Quant 14B Cascade 1…Q: Why two models (HighNoise + LowNoise)? A: That’s the standard Wan2.2 pattern: high-noise for establishing motion/structure, low-noise for refinement. ComfyUI’s official Wan2.2 docs follow the same split.Q: Why does the workflow feel “better” on 2–3 seconds? A: Short clips reduce drift and make prompt control tighter. Then you extend through cascade instead of praying that one long run stays coherent.Q: Can I get 32 fps? A: Yes, but that’s typically done with interpolation nodes after generation (many community workflows do exactly that).

Wan2.2 I2V V2V Video Extend Quant  14B

Files with SHA256: 5f4ee1ec53480006903a114a30e321d9005754b5866b671eca3cdd8418071a93