ComfyUI workflow options

I see the ComfyUI screenshot you uploaded — it looks like a WAN2.2 T2V workflow with UMT5 text encoders, dual WAN models, CLIP conditioning, an advanced KSampler, VAE decode, and a Video Combine node. What would you like me to do with this? I can: ✅ rebuild this into a clean, minimal workflow ✅ rewrite it as a step-by-step setup guide ✅ generate a JSON workflow template ✅ explain each node and how it connects ✅ help fix errors or optimize for speed ✅ convert this into a fresh workflow for WAN2.2 T2V + UMT5 XXL Just tell me what you want next.

rebuild into clean minimal workflow

Below is the **cleanest possible minimal WAN2.2-T2V workflow** you can build in ComfyUI for *text → video* using: - **wan2.2-t2v-14b-high-noise-q4_k_m** - **wan2.2-t2v-14b-low-noise-q4_k_m** - **wan2.2_vae.safetensors** - **umt5_xxl_fp8_e4m3fn_scaled** This removes all unnecessary duplicate nodes, merges text encoders, and keeps only the core graph required for WAN2.2 generation. --- # ✅ Minimal WAN2.2 T2V Workflow (Clean Rebuild) Below is the structure written in a ComfyUI-friendly node order. --- ## **1. Prompt Input** ``` Text Input → "Prompt" Text Input → "Negative Prompt" ``` --- ## **2. UMT5 XXL Text Encoder** **Node:** `WanVideo T5 Text Encoder Loader` Inputs: - `t5_model`: `umt5_xxl_fp8_e4m3fn_scaled` - `quantization`: `fp8_e4m3fn` Outputs: - `text_embeds` (used by WAN model) - `pooled_embeds` (optional) **Node:** `WANVideo TextEncode` Inputs: - `text`: from Prompt input - `negative_text`: from Negative Prompt input - `t5_model`: from the loader Outputs: - `positive_embeds` - `negative_embeds` --- ## **3. WAN 2.2 T2V Model Loader** You only need **one loader** — the model auto-switches high/low noise internally. **Node:** `WanVideo Model Loader` Inputs: - `model`: `wan2.2-t2v-14b-high-noise-q4_k_m` - `vae`: `wan2.2_vae.safetensors` - `attention_mode`: `xformers` or `sag` depending on GPU Outputs: - `model` - `vae` - `config` *(If you want low-noise, use the other file — but you don't use both in the same graph.)* --- ## **4. Empty Latent Video** **Node:** `Empty Latent Video` Settings: - `width`: 512 - `height`: 512 - `frames`: 16 (or whatever you want) - `batch`: 1 --- ## **5. KSampler (WAN-Compatible)** **Node:** `KSampler (Advanced)` Inputs: - `model`: from WAN Model Loader - `positive`: from UMT5 Encoder - `negative`: from UMT5 Encoder - `latent_image`: from Empty Latent Video Recommended settings: ``` Sampler: euler Scheduler: wan2 Steps: 20 (WAN default) CFG: 8 Strength: 1.0 Seed: 12345 ``` Outputs: - `latent_video` --- ## **6. VAE Decode (WAN 2.2 VAE)** **Node:** `VAE Decode` Inputs: - `samples`: from KSampler - `vae`: WAN 2.2 VAE from loader Output: - decoded video frames --- ## **7. Combine Frames Into Video** **Node:** `Video Combine` Inputs: - `images`: from VAE Decode - `fps`: 24 - `format`: mp4 or webm --- ## **8. Save Video** **Node:** `Save Video` --- # 🎉 Workflow Summary (Minimal Graph) ``` Prompt → UMT5 Encoder → Positive Embeds → → KSampler → VAE Decode → Video Combine → Save Negative Prompt → UMT5 Encoder → Negative Embeds → Empty Latent Video → KSampler (WAN2 Scheduler) → ... WAN2.2 Model Loader (model + vae) → KSampler / VAE Decode ``` This is the smallest valid WAN2.2-T2V pipeline possible. --- If you want, I can: 📦 **export this as a .json ComfyUI workflow file** 📊 **draw the node graph** ⚙️ **add automatic high/low noise blending** 🎬 **optimize for higher FPS or resolution** 💨 **optimize for 4070, 3090, or M-series Mac** What’s next?