WAN 2.2 14B Multi-Phase I2V/T2V Workflow

WAN 4-Phase V1.4.3 UpdateWhat's New in V1.4.3This update adds GGUF model support and fixes model unloader connections in the final processing stage.New FeaturesGGUF Model Support<ul><li>Added SmartSequentialWANLoaderGGUFPhase1 node for loading quantized GGUF models</li><li>Added 2x CR Model Input Switch nodes to select between safetensors and GGUF loaders</li><li>New "GGUF Models" and "safesensor models" groups for organization</li><li>Users can now run the workflow with GGUF models for lower VRAM usage</li></ul>Model Input Switching<ul><li>CR Model Input Switch nodes allow you to choose which loader type to use</li><li>Switch between safetensors (standard) and GGUF (quantized) models</li><li>Manual selector - set to input 1 for safetensors, input 2 for GGUF</li></ul>Fixed IssuesModel Unloader Connections<ul><li>Fixed Node 376's 'model for unload' connection (was disconnected in V1.4)</li><li>Model unloaders now properly chain in the Final Processing section</li><li>These unload WAN models after all 4 phases complete, freeing VRAM for final upscaling/interpolation operations</li></ul>System RequirementsMinimum Requirements<ul><li>GPU: NVIDIA GPU with 16GB+ VRAM</li><li>System RAM: 32GB+</li><li>Storage: SSD with 100GB+ free space</li><li>OS: Windows 10/11, Linux, or macOS</li><li>ComfyUI: Latest version</li></ul>GGUF Model Benefits<ul><li>Lower VRAM requirements compared to safetensors models</li><li>Faster loading times</li><li>Same generation quality with proper quantization</li></ul>Required ComfyUI ExtensionsInstall these via ComfyUI Manager:<ol><li>ComfyUI-VideoHelperSuite (VHS) - Video generation and processing</li><li>ComfyUI-Impact-Pack - Execution control and utilities</li><li>ComfyUI-LoRA-Manager - LoRA loading and management</li><li>rgthree's ComfyUI Nodes - Power Prompt, Fast Groups Bypasser</li><li>pysssss Custom Scripts - ShowText, ReroutePrimitive nodes</li><li>ComfyUI-KJNodes - ImageBatchMulti, JoinStringMulti nodes</li><li>ComfyUI-GGUF (city96) - GGUF model loading support</li><li>ComfyUI LayerPeople Nodes - ModelUnloader, SoftFullCleanRAMAndVRAM</li><li>ComfyRoll Custom Nodes - CR Model Input Switch, CR Image Input Switch</li></ol>Required Custom NodesDownload these Python files and place in <code>ComfyUI/custom_nodes/</code>:<ol><li>smart_sequential_wan_<a target="_blank" rel="ugc" href="http://loader.py">loader.py</a> - Safetensors model loader (Phase 1)</li><li>smart_sequential_wan_loader_<a target="_blank" rel="ugc" href="http://gguf.py">gguf.py</a> - GGUF model loader (new in V1.4.2)</li><li>multi_phase_prompt_manager_4phase_<a target="_blank" rel="ugc" href="http://v2.py">v2.py</a> - Multi-phase prompt management</li><li>advanced_wan_settings_<a target="_blank" rel="ugc" href="http://controller.py">controller.py</a> - Video settings control</li></ol>Restart ComfyUI after adding these files.Installation InstructionsModel Setup<ul><li>Safetensors models: Place in <code>ComfyUI/models/diffusion_models/</code></li><li>GGUF models: Place in <code>ComfyUI/models/unet/</code> or <code>ComfyUI/models/diffusion_models/</code></li><li>VAE: Place in <code>ComfyUI/models/vae/</code></li><li>CLIP: Place in <code>ComfyUI/models/clip/</code></li></ul>Workflow Installation<ol><li>Download <code>WAN_4-Phase_V1_4_2.json</code></li><li>Open ComfyUI</li><li>Drag and drop the JSON file into the ComfyUI interface</li><li>ComfyUI will alert you if any required extensions are missing</li></ol>Quick Start GuideStep 1: Choose Your Model TypeThe workflow includes two model loader groups:<ul><li>"safesensor models" - For standard safetensors WAN models</li><li>"GGUF Models" - For quantized GGUF WAN models</li></ul>To use safetensors models:<ol><li>Configure your models in the "safesensor models" group</li><li>Use Fast Groups Bypasser to bypass/disable the "GGUF Models" group</li><li>Set both CR Model Input Switch nodes to input 1</li></ol>To use GGUF models:<ol><li>Configure your models in the "GGUF Models" group</li><li>Use Fast Groups Bypasser to bypass/disable the "safesensor models" group</li><li>Set both CR Model Input Switch nodes to input 2 (current default setting)</li></ol>Step 2: Configure Video SettingsIn the Advanced WAN Settings Controller:<ul><li>Set base resolution (default: 896x1152 - 7:9 Portrait)</li><li>Set frame count (default: 160 frames)</li><li>Set generation FPS (default: 16)</li><li>Set final output FPS (default: 32)</li><li>Choose number of active phases (default: 1, range 1-4)</li><li>Configure upscale multiplier (default: 1.35)</li></ul>Step 3: Set Workflow ModeIn your Phase 1 loader:<ul><li>I2V Mode: Load an image and generate video from it</li><li>T2V Mode: Generate video from text prompts only</li></ul>Step 4: Configure PromptsIn the Multi-Phase Prompt Manager:<ul><li>Enter positive prompts for each phase</li><li>Add LoRA trigger words if needed</li><li>Configure negative prompts per phase</li><li>Enable/disable individual phases</li></ul>Step 5: Configure LoRAs (Optional)Each phase has its own LoRA loader section:<ul><li>Phase 1-4 each support multiple LoRAs</li><li>Use the LoRA Manager visual interface to select LoRAs</li><li>Configure LoRA strengths per phase</li></ul>CRITICAL - Avoid Double-Loading LoRAs:LoRAs loaded in Phase 1 remain active throughout all subsequent phases. When configuring Phases 2-4:<ul><li>DO NOT re-enable/load LoRAs that were already loaded in previous phases</li><li>If you load the same LoRA in multiple phases, you're stacking its weight, which causes severe anomalies and distortions</li><li>Only use trigger words in later phases to activate LoRAs that are already loaded</li><li>Example: If you load <code>LoRA_A</code> with strength 1.0 in Phase 1, just add the trigger word in Phase 2 prompts - don't load <code>LoRA_A</code> again in Phase 2's LoRA loader</li></ul>The LoRAs stay in VRAM across all phases. Re-loading them multiplies their effect and creates very bad artifacts.Step 6: Generate<ol><li>Queue your prompt</li><li>Phase 1 generates initial video</li><li>Phases 2-4 extend the video sequentially</li><li>Final Processing unloads models and combines all segments</li><li>Outputs are saved in your configured output directory</li></ol>Using GGUF ModelsWhat Are GGUF Models?GGUF (GPT-Generated Unified Format) models are quantized versions of standard models that use less VRAM while maintaining quality.ConfigurationSmartSequentialWANLoaderGGUFPhase1 accepts:<ul><li>Workflow mode selection (I2V or T2V)</li><li>Separate I2V high/low model selection</li><li>Separate T2V high/low model selection</li><li>Quantization settings (dequant_dtype, patch_dtype)</li></ul>The loader automatically loads the appropriate model pair based on your selected workflow mode.Model SwitchingThe CR Model Input Switch nodes have a selector widget:<ul><li>Input 1: Routes safetensors loader output</li><li>Input 2: Routes GGUF loader output (current default)</li></ul>You need two switches (one for high_model, one for low_model). Both should be set to the same input number.TroubleshootingGGUF models not appearing in loader:<ul><li>Verify ComfyUI-GGUF extension is installed</li><li>Place GGUF files in <code>models/unet/</code> or <code>models/diffusion_models/</code></li><li>Restart ComfyUI</li></ul>Wrong models loading:<ul><li>Check CR Model Input Switch selector values (both should match)</li><li>Verify the correct loader group is enabled in Fast Groups Bypasser</li><li>Check that the unused loader group is bypassed</li></ul>Out of VRAM errors:<ul><li>Try GGUF models instead of safetensors</li><li>Reduce number of active phases</li><li>Lower base resolution</li><li>Reduce batch size</li><li>Disable upscaling phases</li></ul>Workflow won't load:<ul><li>Install all required extensions via ComfyUI Manager</li><li>Ensure all custom .py files are in custom_nodes directory</li><li>Restart ComfyUI after installing extensions</li></ul>Technical NotesModel Loading ArchitectureV1.4.2 uses a dual-loader system:<ol><li>SmartSequentialWANLoaderPhase1 (safetensors)</li><li>SmartSequentialWANLoaderGGUFPhase1 (GGUF)</li></ol>Both loaders run in parallel but only one set of outputs is used via the CR Model Input Switch nodes.Model UnloadingModelUnloader nodes are located in the "Final Processing" group. They:<ol><li>Unload WAN models after Phase 4 completes</li><li>Free VRAM before upscaling/interpolation operations</li><li>Chain together (Node 376 → Node 394 → SoftFullCleanRAMAndVRAM)</li></ol>This ensures maximum VRAM is available for RIFE upscaling and FPS interpolation.Phase ArchitectureAll 4 phases use the same models loaded once in Phase 1. There is no model switching between phases - the loaded models remain in VRAM throughout all generation phases, then get unloaded during final processing.CreditsWorkflow Design: DarthRidonkulous Custom Nodes: DarthRidonkulous ComfyUI-GGUF Extension: city96 ComfyRoll Nodes: RockOfFire (ComfyRoll Custom Nodes) LayerPeople Nodes: LayerPeopleFind me on CivitAI: @DarthRidonkulousVersion HistoryV1.4.2 (Current)<ul><li>Added GGUF model support via SmartSequentialWANLoaderGGUFPhase1</li><li>Added CR Model Input Switch nodes for loader selection</li><li>Fixed model unloader connections in Final Processing</li><li>Added "GGUF Models" and "safesensor models" organization groups</li></ul>V1.4<ul><li>Previous version (see previous documentation)</li></ul>

wan2214BMultiPhaseI2V_wan22I2VT2V14B4PV142.zip

Files with SHA256: c8a8e4128c94e033b2195a41f16bef68d4f304224be92df02116d8f3eb340ebc