☁️BizyAir Wan_ImageToVideoPipeline

Where BizyAir WanImageToVideo gives you the same graph-building shape as core ComfyUI's Wan node - conditioning in, latent out, needs a sampler after it - this one skips all of that. Prompt and image go in, a finished video comes out, in a single node call. It's the "diffusers pipeline" style: less control over the individual sampling steps, a lot less setup.

The model underneath is fixed to Wan-AI/Wan2.1-I2V-14B-480P-Diffusers - the HuggingFace diffusers-format release of Wan 2.1's flagship 14B image-to-video model at 480p. This is the checkpoint that put open video generation on the map in the first place; running it here means BizyAir's cloud is holding the 14B weights and doing the actual sampling, and your local GPU does nothing.

What it does

You supply an image, a prompt, and some generation settings, and get back a VIDEO object ready to preview or save - no separate KSampler or VAE decode step required. It trades flexibility for convenience: fine, since most of the knobs you'd actually want to touch (steps, CFG, TeaCache, LoRA) are still exposed as inputs on this one node.

Inputs and outputs that matter

• image - required. Worth flagging honestly: the node's own tooltip describes it as an "optional reference image to guide video generation," but the schema marks it required regardless. Treat it as required in practice; don't expect to skip it just because the tooltip says otherwise.
• model_id - only one choice currently, the 14B 480P diffusers checkpoint described above. No 2.2 variant exposed here yet.
• prompt - text description of the video, multiline.
• negative_prompt (optional) - what to steer away from.
• steps (default 30, capped at 40) - sampling steps.
• cfg (default 6, range up to 100) - that ceiling is a lot higher than you'd ever actually want; Wan generally wants CFG in the 3.5–6+ range on the high-noise pass, and cranking this toward the ceiling doesn't buy you anything, it just makes results worse. Stay near the default.
• seed - set a number for reproducibility, or leave at 0 and expect variety.
• use_teacache - enable / disable. TeaCache skips redundant computation across similar denoising steps, and it's a real, meaningful speedup for Wan generations - but the tradeoff is real too: it costs some quality, and specifically hurts face consistency more than other detail. If a clip's face is drifting or morphing across the run, disabling TeaCache is the first thing to try, before touching anything else.
• lora_config (optional) - plug in the output of BizyAir Wan_LoraLoader here to apply a custom LoRA to the generation.

Output is a single VIDEO.

How to install it

• ComfyUI Manager - search "BizyAir", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git, restart.
• Comfy CLI - comfy node install bizyair.

Set your API key first - the in-app login flow or bizyair.siliconflow.cn. Every call to this node runs against BizyAir's cloud and spends credits, even though the underlying model is open-weight.

Common issues & troubleshooting

Faces wobble or lose consistency across the clip. Turn use_teacache off first. It's the single most common cause of exactly this symptom in the wider Wan ecosystem, not something specific to this node - the speedup comes from skipping work, and skipped work occasionally means reusing something that should've changed.

Generation is slower or more expensive than expected. use_teacache set to disable and a high steps value together will get you the best quality and the slowest, priciest runs. If you're iterating on a prompt, flip TeaCache on and drop steps toward the lower end of its range; save the expensive combination for a final render.

LoRA doesn't seem to apply. Confirm lora_config is actually wired from BizyAir Wan_LoraLoader and that the LoRA is registered to your BizyAir account first - this pipeline reads from BizyAir's hosted model library, not a local loras/ folder.