| name | huggingface-video-generation |
| description | Use Hugging Face for video generation workflows. Use when the user asks to build, integrate, debug, or deploy text-to-video or image-guided video features with Hugging Face, including Inference Providers, `@huggingface/inference`, `huggingface_hub` `InferenceClient`, Diffusers-based local pipelines, frame and duration controls, prompt and negative prompt tuning, seeds, and handling returned video bytes. |
Hugging Face Video Generation
Use this skill for product-facing video generation features, not for model training.
Check current docs first
Video generation changes fast across providers and model families. Verify the current official docs before coding:
text-to-video: https://huggingface.co/docs/inference-providers/tasks/text-to-video- JS client docs:
https://huggingface.co/docs/huggingface.js/main/en/inference/modules
- Local video pipelines:
https://huggingface.co/docs/diffusers/main/using-diffusers/text-img2vid
Default approach
Prefer hosted inference first for app integrations and prototypes:
- Web or Node apps:
@huggingface/inference
- Python apps or scripts:
huggingface_hub.InferenceClient
Switch to local Diffusers when the user explicitly wants self-hosting, image-to-video support that is missing from hosted routes, custom schedulers, or lower-level pipeline control.
Auth
- Read the token from
HF_TOKEN.
- Never hardcode tokens.
- Keep provider-specific secrets separate and document them clearly.
App integration checklist
- Confirm whether the requested flow is
text-to-video, image-guided video, or a local Diffusers pipeline.
- Let callers pass
model, prompt, negative_prompt, seed, num_frames, guidance_scale, and num_inference_steps when supported.
- Treat model capability as dynamic. Do not assume a
text-to-video model also supports image-to-video.
- Prefer async jobs or queued work for generation-heavy paths instead of blocking synchronous HTTP handlers.
- Store large video outputs in object storage, temp files, or a media bucket rather than database rows.
Good defaults
- Keep early clips short and low-cost while validating the UX.
- Expose
seed whenever reproducibility matters.
- Surface duration, frame count, and resolution as explicit tradeoffs because latency and cost rise quickly with each of them.
- Frame model names as examples, not permanent defaults.
JavaScript note
@huggingface/inference exposes textToVideo. Verify the exact task and output handling in the current docs before wiring an endpoint, especially for image-guided or provider-specific flows.
Python note
Use InferenceClient.text_to_video for hosted workflows. Use Diffusers pipelines for local or advanced video generation.
