Wan2.2 VACE Fun is a creative, prompt-driven reference-to-video workflow for ComfyUI. Provide a single reference image and a text description, and the graph animates your subject into a coherent video while preserving identity and style. Built on the Wan 2.2 VACE module with a staged sampler, it balances motion, fidelity, and runtime, making it ideal for concept reels, character tests, and short storytelling clips.
This ComfyUI Wan2.2 VACE Fun workflow focuses on three things: strong subject adherence from the reference image, expressive motion guided by your prompt, and reliable export to an MP4 video. Use it when you need quick iterations that still feel cinematic, or when you want to turn a still image into a dynamic scene without complex keyframing.
The workflow runs in stages: you set global controls, prepare the reference image, write prompts, generate a VACE-conditioned latent video, refine it through staged sampling, then decode and export. Groups are organized so you can work top to bottom with minimal friction.
This group centralizes the controls the rest of the graph reads: width, height, length (frames), fps, steps, sampling_shift, and seed. Change these once and every downstream node picks up the values via SetNode/GetNode pairs. Resolution and length influence both quality and VRAM, while fps controls perceived motion in the final MP4. Keep width and height aspect-aligned to your reference image to avoid stretching. The seed is exposed for reproducibility across runs.
Load your subject with LoadImage (#118), then the image is resized in ImageResizeKJv2 (#112) to match your target resolution. RMBG (#73) removes the background so VACE can lock onto the foreground subject more reliably, helping identity consistency across frames. A preview node lets you quickly inspect the cutout before generation. The processed image is stored as the reference and fed downstream.
Prompts are encoded with CLIP Text Encode (Positive Prompt) (#56) and CLIP Text Encode (Negative Prompt) (#54) using the UMT5-XXL encoder. Write clear action verbs, camera language, and scene context in the positive prompt to direct motion and composition. Use the negative prompt to suppress unwanted artifacts, styles, or clutter; multilingual phrasing works well. The outputs provide rich conditioning for the VACE step and the samplers that follow.
The graph loads Wan 2.2 T2V A14B weights and applies the Wan2.2 VACE Fun module, then augments attention and scheduling for stability. The HIGH branch passes through PathchSageAttentionKJ (#8) and ModelSamplingSD3 (#57), while the LOW branch uses LoraLoaderModelOnly (#61), PathchSageAttentionKJ (#66), and ModelSamplingSD3 (#20). This split gives you a detail-first early pass and a motion-focused refinement. All model choices are prewired; you simply run the graph once your settings and prompts are ready.
WanVaceToVideo (#43) injects your reference_image, positive/negative conditioning, and VAE, then generates an initial video latent sequence sized by width, height, and length. Think of this as the moment the still image “learns” to move according to your prompt. The node returns both conditioning streams for reuse and an integer for latent trimming to keep the pipeline frame-consistent. No manual mask or control video is required unless you want to experiment.
A three-stage sampler stack shapes the result. First pass KSamplerAdvanced (#108) seeds the latent sequence for overall composition and motion cues. Second pass KSamplerAdvanced (#107) deepens detail and temporal stability using the same conditioning while preserving the scene layout. Final pass KSamplerAdvanced (#109) runs on the LOW variant to polish motion and reduce artifacts, striking a practical balance between speed and quality. TrimVideoLatent (#65) aligns the frames to the target length before decoding.
VAEDecode (#19) turns the refined latents into RGB frames. VHS_VideoCombine (#69) then assembles those frames into an MP4 at your chosen fps, saving with a sensible filename pattern. This group is optimized for quick review loops, so you can iterate on prompts, length, or resolution without touching the rest of the graph. When satisfied, keep the same seed for repeatability or change it to explore variations.
WanVaceToVideo (#43)
The heart of Wan2.2 VACE Fun: it binds prompt semantics to your reference image and produces the initial video latents. Adjust width, height, and length here via the shared settings to match your creative target and VRAM budget. Keep the reference subject centered and well-lit for best identity retention. If motion feels off, revise the positive prompt to emphasize actions, camera moves, and timing words.
KSamplerAdvanced (#108, #107, #109)
A staged sampler chain that progressively improves composition, detail, and motion smoothness. Increase steps when you need more detail or temporal stability, and reuse the same seed to compare changes fairly. The final pass on the LOW variant often cleans subtle artifacts; if results look too soft, shift some steps to the earlier passes. sampling_shift is exposed to nudge the schedule toward either motion emphasis or finer texture.
RMBG (#73)
Automatic background removal improves subject adherence for Wan2.2 VACE Fun, especially with busy or low-contrast scenes. Use high-quality, non-blurry references to minimize cutout errors. If you want natural backgrounds, replace them later in post or composite a plate behind the decoded frames.
TrimVideoLatent (#65)
Keeps the latent sequence aligned to the requested length. If you extend or shorten a clip, let this node handle the bookkeeping; it prevents subtle frame drift across the sampler stages.
VHS_VideoCombine (#69)
Encodes the final frames to MP4. Tune frame_rate to control clip duration relative to length (duration equals frames divided by fps). Raise quality for final delivery or lower it for rapid previews; higher quality increases file size and encode time.
width and height to avoid stretching and unwanted crops.length and fps in sync with your goal: higher fps looks smoother but shortens total duration for the same frame count.RMBG and add a brief identity clause in the prompt (clothing, color, or gear).This workflow implements and builds upon the following works and resources. We gratefully acknowledge @BenjisAIPlayground for the “Wan2.2 VACE Fun Demo” workflow and for their contributions and maintenance. For authoritative details, please refer to the original documentation and repositories linked below.
Note: Use of the referenced models, datasets, and code is subject to the respective licenses and terms provided by their authors and maintainers.
RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Playground, enabling artists to harness the latest AI tools to create incredible art.
