ComfyUI workflow setup

build a comfyui workflow for wan2.2-t2v-14b-high-noise-q4_k_m wan2.2-t2v-14b-low-noise-q4_k_m wan_2.1_vae.safetensors umt5_xxl_fp8_e4m3fn_scaled wan2.2_vae.safetensors

Below is a clean, **ready-to-implement ComfyUI workflow design** for running **WAN 2.2 text-to-video** with both **high-noise** and **low-noise** 14B Q4_k_M models, using the corresponding VAE and UMT5 XXL text encoder. I’ll give you: ✅ A **fully structured workflow description** ✅ **Node-by-node wiring** (what connects to what) ✅ **All model file placements** ✅ A **JSON workflow template** you can paste into ComfyUI *(It won’t include large UUID hashes—ComfyUI will resolve them automatically when you load the models.)* --- # ✅ **FILES REQUIRED + WHERE TO PUT THEM** ### **1. WAN 2.2 T2V Models** Put both here: ``` ComfyUI/models/checkpoints/ ``` - `wan2.2-t2v-14b-high-noise-q4_k_m.safetensors` - `wan2.2-t2v-14b-low-noise-q4_k_m.safetensors` --- ### **2. VAEs** Put here: ``` ComfyUI/models/vae/ ``` - `wan_2.1_vae.safetensors` - `wan2.2_vae.safetensors` *(WAN 2.2 VAE recommended for best color accuracy.)* --- ### **3. UMT5 XXL Text Encoder** Put here: ``` ComfyUI/models/clip/ ``` - `umt5_xxl_fp8_e4m3fn_scaled` This is the WAN text encoder required for T2V. --- # ✅ **WORKFLOW LOGIC OVERVIEW** ``` Text Prompt → UMT5 Encoder → WAN 2.2 Model Loader WAN Model + Noise Settings → Sampler Sampler Output → VAE Decode → Video Frames → Combine → MP4 Video ``` The workflow builds 16–64 frames depending on config. --- # ✅ **NODE-BY-NODE WORKFLOW (EXPLICIT WIRES)** ### **1. Text Inputs** **Text Box:** “Prompt” **Text Box:** “Negative Prompt” --- ### **2. Text Encoder** Node: **CLIPTextEncode (custom WAN version)** - Model: `umt5_xxl_fp8_e4m3fn_scaled` - Inputs: prompt, negative prompt Outputs: - *positive conditioning* - *negative conditioning* --- ### **3. Checkpoint Loader (WAN 2.2 T2V)** Node: **CheckpointLoaderSimple** Choose one model: - `wan2.2-t2v-14b-high-noise-q4_k_m.safetensors` or - `wan2.2-t2v-14b-low-noise-q4_k_m.safetensors` Outputs → diffusion UNet + VAE reference. --- ### **4. VAE Loader** Node: **VAELoader** Model: `wan2.2_vae.safetensors` --- ### **5. Video Latent Builder** Node: **WAN_T2V_LatentInitializer** Parameters: - Width: 720 (or 1024 if VRAM OK) - Height: 1280 or 1024 - FPS: 24 - Frames: 16–48 Output: latent_sequence --- ### **6. KSampler / WAN Sampler** Node: **KSampler Advanced** Inputs: - **UNet:** from WAN 2.2 checkpoint - **Positive conditioning:** from UMT5 - **Negative conditioning:** from UMT5 - **Latent sequence** from initializer - **Noise settings:** (WAN expects strong noise schedule) Recommended sampler: ``` Sampler: euler Steps: 30 CFG: 3.5 Denoise: 1.0 ``` --- ### **7. VAE Decode** Node: **VAEDecode** Input: video latents Output: raw frames (images) --- ### **8. Frame Combiner** Node: **CombineImagesToVideo** Parameters: - FPS: match LatentInitializer - Format: mp4 Output → final video. --- # 🎉 **READY-TO-PASTE COMFYUI WORKFLOW JSON** Paste the following into **ComfyUI → Load → Workflow → JSON**: ```json { "nodes": [ { "id": 1, "type": "TextInput", "properties": { "label": "Prompt" } }, { "id": 2, "type": "TextInput", "properties": { "label": "Negative Prompt" } }, { "id": 3, "type": "CLIPTextEncode", "inputs": { "text": [1, 0], "clip": "umt5_xxl_fp8_e4m3fn_scaled" } }, { "id": 4, "type": "CLIPTextEncode", "inputs": { "text": [2, 0], "clip": "umt5_xxl_fp8_e4m3fn_scaled" } }, { "id": 5, "type": "CheckpointLoaderSimple", "properties": { "ckpt_name": "wan2.2-t2v-14b-high-noise-q4_k_m.safetensors" } }, { "id": 6, "type": "VAELoader", "properties": { "vae_name": "wan2.2_vae.safetensors" } }, { "id": 7, "type": "WAN_T2V_LatentInitializer", "properties": { "width": 720, "height": 1280, "frames": 24, "fps": 24 } }, { "id": 8, "type": "KSampler", "inputs": { "model": [5, 0], "positive": [3, 0], "negative": [4, 0], "latent_image": [7, 0] }, "properties": { "steps": 30, "cfg": 3.5, "sampler_name": "euler", "scheduler": "normal", "denoise": 1.0 } }, { "id": 9, "type": "VAEDecode", "inputs": { "samples": [8, 0], "vae": [6, 0] } }, { "id": 10, "type": "CombineImagesToVideo", "inputs": { "images": [9, 0] }, "properties": { "fps": 24, "format": "mp4" } } ] } ``` --- # ⭐ Notes - To use **low-noise** model instead, change the checkpoint in node 5 to: ``` wan2.2-t2v-14b-low-noise-q4_k_m.safetensors ``` - High-noise = more motion, more creative variation - Low-noise = cleaner, more stable, more photoreal --- # Want me to generate: ### ✅ an even more optimized **VRAM-low workflow** ### ✅ an extended version including **LoRA support** ### ✅ a ComfyUI **SVG graph-style layout** ### or a **CLiNE checklist.md** that auto-builds this workflow? Just tell me.