AI-HIVE 广告与 TVC 视频

AI-HIVE 广告与 TVC 视频 (class AIHiveAdvertisingVideo) is the ad-flavored twin of the pack's ecommerce video node. Where that one frames a product demo, this one frames a commercial: establish the subject fast, show action, material, and real benefit points, and leave room at the end for brand and call-to-action. It's still AIHiveGenerateVideo underneath - same three families (seedance_2_5, minimax_h3, happyhorse), same five modes (t2v, i2v, r2v, edit, extend), same string-URL outputs.

The platform field is free text, default 抖音 / 小红书 / 视频号 / Instagram / YouTube, and - as with the ecommerce video node - it's literally injected into the prompt. The node prepends advertising guidance (build the subject quickly, show action and real benefits, keep a brand/CTA slot at the end, don't fabricate efficacy, prices, certifications, or brand facts) followed by the target platform and then your own prompt. You can't see that boilerplate on the canvas, but it's doing the work of keeping a TVC-flavored prompt honest and structured.

Why this node exists instead of just one video node

The pack ships three video nodes that are mechanically identical, and the honest take is: the difference is the framing prompt, not the engine. AI-HIVE 视频生成与编辑 is the blank one; the ecommerce and advertising variants exist for people who want the platform framing enforced so their output comes back shaped like a 15-second commercial rather than a looping product b-roll. If you write your own prompts carefully, the base node does all of this - this one is the convenience plus the guardrail. For churning out douyin/information-feed ad variants in volume, that guardrail is worth it.

Inputs and outputs

Inputs carry over from the base video node: prompt, model_family, generation_mode, custom_public_model_id, routing_mode, image_media_ids_json / video_media_ids_json / audio_media_ids_json, first_frame_media_id / last_frame_media_id, model_params_json, wait_for_result, timeout_seconds, plus the optional reference_image IMAGE socket (uploaded and appended automatically). Outputs are result_urls_json, task_id, task_json - URLs again, no video tensor, so budget for a download step. Family/mode rules from the base node hold: minimax_h3 only does t2v/i2v/r2v, and an unsupported combo needs custom_public_model_id starting with public_model_.

Installing

ComfyUI Manager: search AI-HIVE, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/wubin1836/comfyui-ai-hive.git
pip install -r comfyui-ai-hive/requirements.txt

Only requests and Pillow, no model downloads. Set AI_HIVE_API_KEY (or ~/.ai-hive/config.json) before launching ComfyUI.

Gotchas

Same family as the whole pack: paid and metered (video is the expensive direction - watch your job length), data leaves the machine to a third-party server, and a timeout doesn't cancel the job - it keeps running and billing, so pull results later with AI-HIVE 查询任务 instead of re-queueing. And since the built-in guidance keeps the model from inventing claims, it's on you to keep the facts in the prompt accurate. Closed-model quality, metered price, no local compute - the trade is the same as always, just aimed at making the ad.