WAN 2.2 GGUF — FLF Loop &amp; Video Workflow (8GB VRAM)This workflow is a complete WAN 2.2 video generation pipeline designed to run on consumer GPUs with 8GB VRAM. It supports seamless looping, normal video generation and image/video reinterpretation inside one graph.The system is based on a start-frame / end-frame method.You can use it in three ways:• Seamless FLF loop (same start &amp; end image) • Normal video (different start &amp; end image) • Animate / restyle / extend existing footageAutomatic prompting using Qwen is included but optional. If you don’t want AI prompts, just bypass the Qwen node and write a prompt manually.The workflow also contains:<ul><li>frame interpolation</li><li>temporal color matching</li><li>optional upscaling</li><li>negative prompt handling (NAG compatible)</li></ul>The Radial Attention Patch is included ONLY to increase speed. ( 6 minutes for 5 seconds 720p) It is optional and the workflow works without it.<hr /><h2 id="model-installation-(very-important)">MODEL INSTALLATION (VERY IMPORTANT)</h2>WAN GGUF models DO NOT go into the normal UNet folder.Place them here: ComfyUI/models/diffusers/WAN 2.2 works using a High-Noise + Low-Noise expert pair system (Mixture-of-Experts video diffusion). (<a target="_blank" rel="ugc" href="http://docs.comfy.org">docs.comfy.org</a>) Both models are required or motion generation will break.Recommended model pair used in this workflow:SmoothMix WAN 2.2 I2V GGUF <a target="_blank" rel="ugc" href="https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF">https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF</a>Alternative repository: <a target="_blank" rel="ugc" href="https://huggingface.co/BigDannyPt/WAN-2.2-SmoothMix-GGUF">https://huggingface.co/BigDannyPt/WAN-2.2-SmoothMix-GGUF</a><hr />VAE: wan_2.1_vae.safetensors Place into: ComfyUI/models/vae/<hr />TEXT ENCODER: umt5_xxl_fp8_e4m3fn_scaled.safetensorsor NSFW version<a target="_blank" rel="ugc" href="https://huggingface.co/zootkitty/nsfw_wan_umt5-xxl_bf16_fixed/resolve/main/nsfw_wan_umt5-xxl_bf16_fixed.safetensors?download=true">https://huggingface.co/zootkitty/nsfw_wan_umt5-xxl_bf16_fixed/resolve/main/nsfw_wan_umt5-xxl_bf16_fixed.safetensors?download=true</a> Place into: ComfyUI/models/text_encoders/<hr /><h2 id="qwen-auto-prompt-(optional)">QWEN AUTO PROMPT (OPTIONAL)</h2>This workflow can automatically generate prompts using Qwen-VL.Qwen ComfyUI nodes: <a target="_blank" rel="ugc" href="https://github.com/1038lab/ComfyUI-QwenVL">https://github.com/1038lab/ComfyUI-QwenVL</a> (<a target="_blank" rel="ugc" href="https://github.com/1038lab/ComfyUI-QwenVL?utm_source=chatgpt.com">GitHub</a>)or custom node with wan mode <a target="_blank" rel="ugc" href="https://github.com/huchukato/ComfyUI-QwenVL-Mod">https://github.com/huchukato/ComfyUI-QwenVL-Mod</a>Download a supported Qwen model (example): <a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen2.5-VL">https://huggingface.co/Qwen/Qwen2.5-VL</a>Place model into: ComfyUI/models/LLM/Qwen is a vision-language model capable of understanding images and generating descriptive prompts. (<a target="_blank" rel="ugc" href="https://github.com/1038lab/ComfyUI-QwenVL?utm_source=chatgpt.com">GitHub</a>)If not installed: → simply bypass the Qwen node.<hr /><h2 id="painter-inpaint-helper">PAINTER / INPAINT HELPER</h2>The workflow uses a painter/inpaint helper for temporal consistency between frames.<a target="_blank" rel="ugc" href="https://github.com/princepainter/ComfyUI-PainterI2VforKJ">https://github.com/princepainter/ComfyUI-PainterI2VforKJ</a>Install in Comfy/custom nodesThese node packs include mask and enhancement helpers used for automatic detail correction and repainting. (<a target="_blank" rel="ugc" href="http://ltdrdata.github.io">ltdrdata.github.io</a>)<hr /><h2 id="required-nodes">REQUIRED NODES</h2>Install with ComfyUI Manager:• VideoHelperSuite • KJNodes • WAS Node Suite • RGThree Nodes • Custom Scripts (pysssss) • ComfyUI GGUF Loader<hr /><h2 id="optional-speed-boost-(radial-attention-patch)">OPTIONAL SPEED BOOST (RADIAL ATTENTION PATCH)</h2>The radial attention patch accelerates video diffusion.It is OPTIONAL.If you do not want additional dependencies: → disable/bypass the node and the workflow still runs normally.<h3 id="step-1-install-triton-for-windows">Step 1 — Install Triton for Windows</h3>WAN 2.2 + RadialAttn requires Triton. Download the correct Windows wheel here: <a target="_blank" rel="ugc" href="https://github.com/woct0rdho/triton-windows/releases">https://github.com/woct0rdho/triton-windows/releases</a>Install inside your venv.<h3 id="step-2-install-sparsesageattn">Step 2 — Install SparseSageAttn</h3>Download the Windows wheel from: <a target="_blank" rel="ugc" href="https://github.com/woct0rdho/SpargeAttn/releases">https://github.com/woct0rdho/SpargeAttn/releases</a>Install inside your venv.<h3 id="step-3-install-radialattn-node">Step 3 — Install RadialAttn Node</h3>Download from: <a target="_blank" rel="ugc" href="https://github.com/woct0rdho/ComfyUI-RadialAttn">https://github.com/woct0rdho/ComfyUI-RadialAttn</a>Place inside your ComfyUI custom_nodes folder.<h3 id="step-4-restart-comfyui">Step 4 — Restart ComfyUI</h3>If Sparse / Radial Attn loads correctly, startup log will say: “Using sparse_sage_attn as block_sparse_sage2_attn_cuda”Then the patch is active.<hr /><h2 id="usage">USAGE</h2>Sampler: Euler simple CFG: 1 Steps: smooth mix 3+3Loop: Use identical start and end frame. Enable frame cut nodes and connect it to the next group input.Video: Use different images. Disable frame cut nodes and connect the output to the group next to framecut.WAN prefers scene descriptions and camera movement prompts.Example for timebased promt: (At 0 seconds: The camera is positioned in the neon-lit city at the base of the massive central tower; the vertical blue energy beam rises straight upward through the frame, surrounded by dense futuristic architecture glowing in purple, blue, and warm orange light. No text, symbols, or logos exist anywhere in the scene.)(At 1 seconds: The camera begins a smooth, continuous upward movement; street-level details slide downward and out of view while the beam remains perfectly centered. Traffic lanes are active, small flying vehicles move calmly, windows flicker, and holographic panels animate softly.)(At 2 seconds: Mid-city levels pass by slowly; illuminated windows, platforms, holographic advertisements and distant air traffic drift beneath the camera as atmospheric haze and particles float through the air. The city clearly feels alive but calm.)(At 3 seconds: The camera ascends above the streets and crowds; the tower now dominates the frame and the energy beam grows brighter. Soft pulses of light travel upward inside the beam and the surrounding buildings subtly react with changing reflections.)(At 4 seconds: Near the upper tower section, the beam intensifies and begins emitting faint geometric light patterns into the sky above the city. The patterns slowly organize and form a large transparent holographic structure suspended in the air.)(At 5 seconds: The holographic projection stabilizes into a bright floating logo created by the energy beam itself. The logo is made of luminous particles and soft light, perfectly centered above the city. The camera gently stops and holds; the city continues subtle motion — drifting mist, distant vehicles and softly flickering lights — forming a seamless looping end frame.)Avoid portrait prompts.<hr /><h2 id="notes">NOTES</h2>This workflow is optimized for:<ul><li>temporal stability</li><li>long sequences</li><li>low VRAM usage</li><li>reduced flicker</li><li>consistent motion</li></ul>You can freely modify and extend it.

Files with SHA256: 69a8bfc46a28187f9813d36b3bdf85148b3081785b6b09acad1898d7e18d4795