Wan Start-End Frame Video with VACE
Provide a first frame, a last frame and a prompt, and Wan creates the clip between them. VACE control options and the self-forcing model add consistency.
- L40S
- Image to video
- Video editing
- Free template
How it works
The workflow is built on ComfyUI-WanVideoWrapper. The model is Wan2.1-T2V-1.3B-Self-Forcing-DMD-FP16, combined with the Wan2_1-VACE_module_1_3B_bf16 module. T5 text encoding uses umt5-xxl-enc-bf16 and decoding uses Wan2_1_VAE_fp32.
Start and end images are loaded, resized with ImageResizeKJv2 and optionally cleaned with BiRefNetRMBG. WanVideoPhantomEmbeds and ImageBatchMulti combine them into the conditioning, while a control video can be run through DepthAnything_V2 (depth_anything_v2_vitl_fp16) or the AIO_Preprocessor from comfyui_controlnet_aux for VACE guidance. WanVideoLoraSelect offers LoRAs such as AccVid, CausVid and Fun-Reward. WanVideoSampler uses a CFG schedule, with SLG and Enhance-A-Video available.
Afterward, ColorAdjustment and LTXVFilmGrain polish the clip and VHS_VideoCombine writes it. Block swap, torch compile and VRAM management nodes are included, and the notes explain that torch compile only helps if you run the same settings repeatedly.
Tips
- Pick start and end frames with similar framing to get a smooth transition.
- Describe the camera movement in the action prompt.
- Keep 1.3B models, LoRAs and VACE modules matched.
- The default is 81 frames at 24 FPS.
What you can use it for
- Morphing between two keyframes
- Camera moves between a start and end composition
- Product or mech reveals from two stills
- Controlled shots using depth or pose from a reference clip
Run it in three steps
Open the template
Click Run on ComfyDeploy. The workflow, its custom nodes and models are set up for you on a cloud machine.
Add your inputs
Provide start frame and end frame in the playground or the ComfyUI editor.
Run, share or deploy
Run it on an L40S, share it with your team as a playground, or deploy it as an API endpoint.
Frequently asked questions
What GPU does the start-end frame VACE workflow need?
The recommended GPU is the L40S. ComfyDeploy provisions it for you, so there is nothing to install.
Why do the model sizes matter?
The workflow notes warn that text encoders and LoRAs are not interchangeable, 1.3B LoRAs do not work on 14B, and VACE only works with T2V models of the matching size.
Can the prompt be written automatically?
Yes. An LLM Toolkit chain with OpenAIProviderNode can create prompts from your references, using secrets set on the machine or in your environment.
