Vertex Veo Insert Content

This is the node that makes Adforge feel like an ad toolkit instead of just a wrapper around a video API. Vertex Veo Insert Content takes an existing video and a mask, and replaces whatever's inside the mask with new content - a product dropped into a shelf shot, a logo composited over footage, a new prop where an empty space was. It's video inpainting in the good sense: the masked region gets regenerated from your prompt, and everything outside the mask stays put.

If that sounds like the classic inpainting trade-off, it is, and the KB puts the case for masks well: instruction-based editing models drift on the bits you didn't ask about, while a mask leaves the rest of the frame alone. That "exact placement, boundary you decide" property is precisely what you want when the unmasked footage is expensive-to-reshoot b-roll. Veo 3.1 is the version that got serious about this granular editing.

How it works

The node builds a VideoGenerationMask in INSERT mode and passes it in the generation config. Your video comes from the video input (a VIDEO type, read to bytes and uploaded) or from video_gcs_uri - you need one. Your mask comes from the mask_image input or mask_image_gcs_uri, and you need one of those too. Then it's the same pipeline as the rest of the pack: async generate_videos call, polled every five seconds, Veo writes the edited clip to GCS, and the node downloads it back. Outputs are the standard videos (VIDEO list) and video_path_list (local paths).

The inputs that matter

• prompt - required, and it's a description of the new content, not the video. "A matte black espresso machine on the counter," not "edit my kitchen video."
• mask_image (or mask_image_gcs_uri) - the area of interest, an IMAGE. White or colored region on the area to replace; Veo treats it as the region to regenerate.
• video (or video_gcs_uri) - the footage to edit. This pack's own LoadVideoGCS node is a natural upstream here: bucket footage → local path → into this node's video input.
• video_mime_type / mask_mime_type - video/mp4 and image/png by default; match reality.
• duration_seconds - 1 to 10 here, the widest range in the pack, since you're editing an existing clip rather than inventing one from scratch.
• generate_audio - defaults to off for editing; you generally want the original clip's sound intact.

The rest of the config - model, aspect_ratio, resolution, fps, seed, number_of_videos, enhance_prompt, person_generation, negative_prompt - is shared across the Veo nodes. Defaults are fine to start.

Installing and authenticating

Search adforge in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Creditas-labs/ComfyUI-Adforge
cd ComfyUI-Adforge
pip install -r requirements.txt   # or: uv sync

Restart; Python 3.12+. Auth is the same Vertex setup as every generation node: copy .env.example to custom_nodes/comfyui_adforge/.env with GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_STORAGE_BUCKET, GOOGLE_CLOUD_LOCATION, API_KEY (or export them), and a bucket that actually exists - Veo writes output there first. GOOGLE_APPLICATION_CREDENTIALS with a service account works instead of the key.

Common issues

• Mask too tight. Veo needs room to blend. A mask that hugs the object exactly gives you seams; a slightly generous mask gives you a believable composite.
• Mask/video mismatch. If the mask doesn't line up with the frame - different resolution, rotated, the video got resized upstream - you'll get garbage in the masked region and it's not the model's fault. Keep the mask on the same canvas as the source frames.
• Cost. It's a paid API call per edit, and iteration is the trap: fine-tuning what goes in the box means paying for the box each time. Batch your prompt ideas, then spend.