Evaluate the creative asset, then release the pipeline

A generative-advertising pipeline succeeds when it produces an asset the campaign can use. Model quality, rendering speed and low GPU cost are useful contributors, but none is the final acceptance criterion.

The public XGen reference keeps the approved image, motion prompt, graph and worker profile separate. That makes it possible to diagnose a rejection: did the image establish the wrong product, did the motion distort it, or did finishing damage an otherwise usable clip?

Use different prompts for image and motion

An image prompt describes subject, composition, lighting, material and required empty space. A motion prompt describes what changes over time and what must remain stable.

The reference motion asks for a slow dolly toward a stationary bottle, stable lighting and one continuous shot. Repeating a long image-style prompt is less useful than explicitly defining camera movement and invariants.

{
  "shot_id": "aster-bottle-001",
  "input_image": "approved-keyframe.png",
  "motion": "very slow dolly forward",
  "preserve": ["bottle geometry", "cap", "color", "position"],
  "avoid": ["scene cuts", "flicker", "extra objects", "generated lettering"]
}

This is the creative specification. The full video prompt JSON also records positive/negative conditioning text, model family, seed, frame count, dimensions, sampling parameters and delivery settings.

Evaluate before and after motion

Image review checks product silhouette, reference fidelity, lighting, composition and unwanted content. Only approved images proceed to motion. Video review then checks temporal consistency, object deformation, camera stability, flicker and the beginning/end of the shot.

Inspect frames across the sequence, not only the thumbnail. A clip can start correctly and develop a second cap or warped label halfway through. Optical flow and perceptual metrics can flag unusual changes, but a reviewer still needs to assess whether the actual product remains correct.

A general video benchmark can inform the evaluation dimensions, while campaign-specific acceptance requires its own examples and reviewers. The goal is to use automated signals to narrow review, not to treat one aggregate score as proof of brand fidelity.

A small but useful acceptance set

Build an authored set that varies the product’s hard cases: reflective surfaces, thin handles, repeated patterns, transparent material and close-up geometry. Vary camera movement separately from those appearance conditions.

For each case, store the input reference, required invariants, permitted motion and rejection rules. Run the same cases across candidate checkpoints and configurations. Keep the seed set fixed for paired comparisons, then broaden it before release.

Separate hard failures from preferences. A changed product shape is a hard rejection when exact identity matters. A slightly warmer background may be a stylistic preference. Combining both into a single average can allow serious errors to be offset by attractive lighting.

The finishing stage is deterministic where possible

Exact campaign text, prices, logos and call-to-action elements can be composited after generation. This reduces the burden on a stochastic model and makes reviewable edits easier.

Encoding should preserve the intended dimensions, frame order, duration and color treatment. Check the output with a media probe rather than trusting a successful process exit. Audio must be aligned to the encoded timeline, and any interpolation step should be recorded separately from the base generation rate.

The retained project includes audio-related and interpolation-labeled sample collections. Their presence supports a record of those experiments; it does not establish that every video used the same finishing pipeline.

Budget a release comparison

For a candidate pipeline, report accepted outputs, rejected outputs by reason, retries, model time, total wall time and reviewer effort. Also record cold starts and failed attempts, since they contribute to cost even when no asset reaches the reviewer.

Useful measures include acceptance rate, cost per accepted asset, time to first acceptable variation and the fraction requiring manual correction. A faster sampler can lose commercially if it increases product distortions or review effort.

Do not infer campaign return on ad spend from visual quality. Marketing performance requires a controlled deployment and an appropriate attribution window. The engineering release decision establishes asset quality and operating behavior; campaign effectiveness is another experiment.

Exact handoff sequence

1. Review the creative brief and authorized reference assets.
2. Send the image request through the selected provider, keeping credentials outside stored JSON.
3. Save and approve a keyframe, then prepare the exact video dimensions.
4. Download the chosen model artifacts from their publishers and record hashes and use review.
5. Provision one candidate GPU worker, upload the approved image, update its returned path in the graph and prompt specification, and validate the installed node schema against the graph.
6. Run a budgeted smoke generation, saving timing, memory and output evidence.
7. Encode only that attempt’s ordered frames, then validate the resulting file.
8. Evaluate the acceptance set and retain rejected outputs with reasons.
9. Freeze the accepted model/workflow/runtime bundle and its rollback target.
10. Admit limited campaign work with spend limits, ownership and review intact.

The reference kit includes the JSON files, offline validator and operating instructions for these steps. It contains neither model weights nor original client media.

What has and has not been verified

The public graph’s node references, required fields, dimensions, frame count and consistency with the supplied prompt specification are checked offline. The kit also includes a read-only schema check against a running ComfyUI instance.

No GPU generation or paid provider request was performed while preparing this publication. Hardware headroom, visual quality and production timing therefore remain explicit smoke-test and evaluation gates.

That boundary keeps the handoff useful: a developer receives concrete files and acceptance steps, while the portfolio accurately distinguishes recovered experimentation from the new reference pipeline.