Kuzushi Labs · Applied research and engineering

Engineering creative generation pipelines

XGen's advertising and marketing workflows: recovered ComfyUI components, Nano Banana images, Wan 2.1 video, LoRA provenance, Runpod workers and creative evaluation.

Arunank Sharan · June 2026 · 6 articles · 5,343 words

A technical retrospective and reference implementation. Historical graph references remain separate from the new Nano Banana → Wan 2.1 recipe. Offline examples are checked; GPU performance and historical Pod allocations are not inferred.

Parts 1–2 establish the workflow and approved keyframe. Parts 3–4 develop the ComfyUI graph and model/LoRA record. Parts 5–6 cover hardware, operating costs, quality and the implementation handoff.

1. PART 01 · 5 min read

   From a marketing brief to a reproducible creative pipeline

   What 240 embedded ComfyUI graphs reveal, and how a creative brief becomes an approved image and video.

2. PART 02 · 5 min read

   Image generation: Nano Banana, local models and structured prompts

   An exact image request, structured creative prompts, local model alternatives and the approved-keyframe boundary.

3. PART 03 · 5 min read

   Wan 2.1 image-to-video: an inspectable ComfyUI graph

   The complete 13-node API graph, image conditioning, sampling parameters, ordered frames and runtime validation.

4. PART 04 · 5 min read

   LoRAs, model identity and the reproducibility record

   Recovered model and LoRA filenames, compatibility, execution evidence, adapter ablations and artifact identity.

5. PART 05 · 5 min read

   Runpod workers: GPU memory, CPU work and job economics

   Explicit 24, 48 and 80 GB GPU profiles, host resources, persistent storage, retries and cost per accepted clip.

6. PART 06 · 5 min read

   Evaluate the creative asset, then release the pipeline

   Motion prompt JSON, campaign-specific quality checks, finishing and the exact implementation handoff sequence.