The Problem
Producing ad creatives is a coordination tax: briefing designers, waiting on concepts, revising over email, then adapting stills into video for every placement — days of back-and-forth for assets that are usually variations on one product shot.
The Solution
A fully agentic creative pipeline driven from a single Telegram chat. A product photo and one-line brief trigger GPT-4o vision profiling, then an OpenAI agent with structured output parsing drafts schema-validated image/video prompts, caption, and aspect ratio as a seasoned creative director would. Nothing generates until a human approves in-chat — rejections route feedback to a dedicated revision agent. On approval, kie.ai's gpt4o-image places the real product into the AI-designed scene, that approved still seeds Veo 3 image-to-video for visual continuity, and both deliverables land back in the originating conversation.
How the workflow runs
01 · Intake
User sends a product photo with a caption brief; n8n fetches the file via Telegram's API.
02 · Analyse
chatgpt-4o-latest produces a detailed product-and-brand description while explicitly ignoring the background.
03 · Direct
The creative agent returns strict JSON — image_prompt, video_prompt, caption, creative_summary, aspect_ratio, video_model — after a Think-tool self-check.
04 · Gate
The summary ships to Telegram as an approval card; 'No' routes feedback to the revision agent and loops the concept.
05 · Render
Approved prompts drive gpt4o-image to place the actual product inside the designed scenario at the chosen aspect ratio.
06 · Animate
The generated still seeds Veo 3 image-to-video, so the motion ad inherits the exact look of the approved frame; both assets return to chat.
AI layer
- Vision product profiling
- GPT-4o grounds every downstream prompt in what the product actually looks like, not assumptions.
- Schema-validated creativity
- Structured Output Parser guarantees machine-usable JSON from a freeform creative task.
- Self-reviewing agent
- The Think tool makes the agent double-check its concept before spending generation credits.
Automation layer
- Chat-native operations
- Intake, approval, revision, and delivery all happen in one Telegram thread — zero dashboards.
- Bounded revision loop
- Concepts iterate in cheap prompt-space through the revision agent before any paid generation runs.
- Poll-based generation handoff
- Wait/poll patterns around kie.ai endpoints turn async image/video jobs into a linear, reliable workflow.
- Continuity by construction
- Feeding the approved still directly into Veo 3 eliminates brand drift between image and video ads.
Technical challenges
Generated ads often drift from the real product — models invent packaging, logos, and colours.
The pipeline anchors generation to the actual photo twice: vision analysis grounds the prompts, and the original product is composited into the generated scene.
Autonomous generation wastes credits on off-target concepts nobody asked for.
An explicit human gate approves or revises the concept in chat before any image or video spend occurs.
Image and video APIs are asynchronous with long processing times, breaking naive workflow chains.
Wait-and-poll sub-flows around each kie.ai job keep the n8n execution reliable regardless of queue times.
Outcome
- Product photo to finished image + video ad in one chat conversation, minutes instead of days.
- Every generated asset traces back to an explicitly human-approved concept — no wasted renders.
- Image and video stay visually consistent by construction, ready for multi-format campaigns.
Lessons learned
- “Human-in-the-loop gates belong before expensive steps, not after — approve concepts, not invoices.”
- “Structured outputs turn 'creative' LLM tasks into dependable automation inputs.”
- “Reusing one artifact across stages (photo → analysis → prompt → still → video) keeps whole pipelines coherent.”