Video Gen Pipeline — Decisions

2026-04-11 · 5 decisions to confirm before execution

Review each decision below. Recommendations are pre-selected based on the architecture analysis. Override any that don't match your priorities, add notes, then submit.

0 of 5 decisions confirmed

Decision 1

Model Download Priority

Which model to download and test first? This determines the validation path for the full stack (download → ComfyUI load → test render).

Wan 2.2 5B first Recommended

Smallest model (10GB). Fastest to validate full stack end-to-end. LTX 2.3 second for Blackwell testing.

LTX Video 2.3 first

Best Blackwell optimization. Tests NVFP4/FP8 path immediately but larger download (~8GB UNet + encoders).

Decision 2

Overnight Schedule Window

When to switch from voice pipeline to video rendering. Voice calls for current tenants (GDS, CRU, North Group, noboxAI) rarely come after 9pm.

10pm – 7am (9 hours) Recommended

Longest batch window. Covers most rendering workloads overnight.

11pm – 6am (7 hours)

Preserves more voice pipeline availability. Better if tenant call volume grows past 10pm.

Custom schedule

Specify a different window in the notes below.

Decision 3

Output Delivery Method

How generated media files (videos, images) are delivered. Can start simple and add complexity later.

Caddy file share (LAN) Recommended

Add a /media/ route to existing Caddy reverse proxy. Zero new infrastructure. LAN access first.

FTP to client delivery server

Push outputs to external server for client access. Adds FTP config and network dependency.

CDN integration

Upload to CDN for public/client access. More complex but scales. Consider after volume justifies it.

Decision 4

Workflow Complexity

Whether to build simple single-model workflows first or jump directly to multi-model pipelines (e.g., FLUX.2 keyframe → Wan 14B animation).

Single-model workflows first Recommended

Prove the stack works before adding complexity. Multi-model variants layer on top of proven templates.

Multi-model pipelines from day one

Build image-to-video chains immediately. More powerful but larger debugging surface.

Decision 5

Blackwell CUDA Fallback

The official ComfyUI Docker image may not include sm_120 kernels for Blackwell GPU. Same issue that forced Kokoro TTS to CPU image.

Accept PyTorch nightly fallback Recommended

If official image fails, build from PyTorch nightly base with CUDA 13.1. Adds ~10min to first build (one-time cost).

Wait for official image update

Block until upstream ships Blackwell support. Timeline unpredictable.

Additional Notes or Context