# Wan 2.2 I2V – Start + End Frame, Last-Frame Saver (8GB VRAM)
A clean, lightweight Wan 2.2 image-to-video workflow built for 8GB VRAM cards.
Quantized GGUF models + LightX2V 4-step LoRAs = a 7-second 480×832 clip in just 4 steps.
## ✨ Features
- 🎬 Start frame – animate any image
- 🎯 Optional end frame – works as an anchor: face and pose are pulled toward it by the end of the clip (bypass it with Ctrl+B to run start-frame only)
- 💾 Automatic last-frame saving – the final frame of every clip is saved to your output folder, ready to be the start of the next clip
- 🔗 Built for long videos – chain 3–4 clips into 20+ seconds. Two chaining methods are explained inside the workflow notes (keyframes, or continuation with face refresh)
- ⚡ 4 steps – 2 high-noise + 2 low-noise (8-step setup explained in the notes)
- 🧊 8GB VRAM friendly – Q4_K_M GGUF models, Q5_K_M text encoder
- 📖 Full instructions inside – Notes in the workflow explain everything
- ⟳ Quick frame loading – the load nodes can grab the newest file from your output folder with one click
## ⚙️ Default settings
- Resolution: 480×832 (portrait)
- Length: 113 frames ≈ 7 s at 16 fps (81 ≈ 5 s, 97 ≈ 6 s)
- Output: H.264 MP4, CRF 17
## 📥 What to download
Custom nodes (install via ComfyUI Manager)
- [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF) by city96
- ComfyUI-VideoHelperSuite (VHS) by Kosinkadink
- The "Load Image (from Outputs)" node is built into recent ComfyUI – if it shows red, update ComfyUI
Models
| File | Put in | Link |
|---|---|---|
| Wan2.2-I2V-A14B-HighNoise-Q4_K_M.gguf (9.65 GB) | ComfyUI/models/unet | [QuantStack/Wan2.2-I2V-A14B-GGUF](https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF) → HighNoise folder |
| Wan2.2-I2V-A14B-LowNoise-Q4_K_M.gguf (9.65 GB) | ComfyUI/models/unet | same repo → LowNoise folder |
| wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors | ComfyUI/models/loras | [lightx2v/Wan2.2-Distill-Loras](https://huggingface.co/lightx2v/Wan2.2-Distill-Loras) |
| wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors | ComfyUI/models/loras | same repo |
| umt5-xxl-encoder-Q5_K_M.gguf (4.15 GB) | ComfyUI/models/text_encoders (or clip) | [city96/umt5-xxl-encoder-gguf](https://huggingface.co/city96/umt5-xxl-encoder-gguf) |
| wan_2.1_vae.safetensors | ComfyUI/models/vae | Wan 2.1 VAE (e.g. Comfy-Org "Wan_2.1_ComfyUI_repackaged" → split_files/vae) |
If you use a different quant or file name, just re-select it in the loader node.
## 🚀 How to use
1. Load your START frame (⟳ picks the newest image from your output folder).
2. (Optional) Enable the END frame to anchor where the clip finishes.
3. Write a prompt describing motion and camera only (Wan doesn't generate audio).
4. Set the length in the VIDEO node (frames must be 4n+1).
5. Run. The video and the last frame are saved automatically.
## 💡 Tips
- Wan takes the face only from the start/end frames – there is no face reference input. For long chains, refresh the face between clips (see the CHAIN note inside the workflow).
- At the join, trim the first frame of the next clip in your video editor to avoid a micro-pause.
- Keep the motion speed steady and don't end a clip on a standstill, so the next clip continues smoothly.
Enjoy, and feel free to share your results! 🙌
