Video & Image Generation Pipeline — Architecture Briefing

32 GB

VRAM (GDDR7 ECC)

Generation Models

~70 GB

Model Footprint

9 hr

Nightly Batch Window

2 TB

NVMe Storage

Infrastructure System Architecture

Profile-based GPU multiplexing on a single 32GB card. One workload at a time — voice AI during business hours, media generation off-peak.

GPU Profile Switching — Voice / Video / Image Modes

Storage Layout

Drive	Mount	Role	Capacity
Samsung 9100 PRO #0	`/`	OS, Docker images, compose config	1 TB
Samsung 9100 PRO #1	`/data`	Models, outputs, workflows, HF cache	2 TB

Critical Fix Applied: Volume Paths

ComfyUI volumes were using relative ./data/ paths (1TB OS drive). Changed to absolute /data/comfyui/ (2TB NVMe) matching the HuggingFace cache pattern. Without this fix, model downloads would fill the OS drive.

Models Generation Stack

Five models covering video speed, video quality, and image generation. All open-weight, self-hosted, no API fees.

VRAM Budget per Model (FP8 Quantization)

#	Model	Params	VRAM	Role
1	Wan 2.2 5B	5B	8-12 GB	Daily workhorse — quick B-roll, tests
2	LTX Video 2.3	22B	12-16 GB	Speed lane — 5-8x faster, audio-inclusive
3	FLUX.2 dev	12B	12-16 GB	Image gen — keyframes, thumbnails, stills
4	Wan 2.2 14B	14B	16-24 GB	Quality lane — cinematic clips, B-roll
5	HunyuanVideo 1.5	8.3B	~22 GB	Cinematic alt — overnight runs

Workflow Strategy: Prototype fast, refine for quality

LTX 2.3 for rapid iteration (5-8x faster than Wan). Switch to Wan 2.2 14B when a scene needs cinematic quality. FLUX.2 for all still image work (keyframes, thumbnails, reference images). This mirrors the validated "prototype fast with LTX, refine with Wan" pattern from the open-source community.

Timeline Implementation Pipeline

Six phases over 8 days. Phases 4-5 run in parallel.

Implementation Timeline — 8 Days to Production

P1: Models

→

P2: ComfyUI Build

→

P3: Workflows

→

P4+5: Automation + Image

→

P6: Monitoring

Phase Transition Gate

○All 5 models downloaded and loadable in ComfyUI
○At least 3 workflow templates produce valid output
○One complete overnight batch cycle succeeds (switch → render → restore)
○Voice pipeline confirmed healthy after restore

Operations Overnight Batch Automation

Cron-driven profile switching with healthcheck validation and sequential batch processing.

Overnight Batch Cycle — 10pm to 7am

Automation Scripts

Script	Purpose
`switch-to-video.sh`	Teardown all profiles, build ComfyUI, healthcheck wait (up to 10min), optional workflow submission
`switch-to-image.sh`	Start ComfyUI + shared services (no voice pipeline)
`submit-workflow.sh`	POST workflow JSON to ComfyUI `/prompt` API with `jq` parameter injection
`batch-render.sh`	Process queue directory sequentially, poll completion, rename `.done`/`.failed`
`cleanup-outputs.sh`	30-day scratch retention, disk usage report per category
`gpu-monitor.sh`	nvidia-smi polling with thermal alert (default 85°C threshold)

Cron Schedule

Time	Action
10:00 PM	Switch to video mode (teardown voice, build ComfyUI, healthcheck)
10:15 PM	Batch render (process workflow queue)
7:00 AM	Restore default mode (voice pipeline)
8:00 AM Sun	Weekly output cleanup (scratch retention)

Action Needed Operator Decisions

Five decisions to confirm before execution begins. Recommendations provided for each.

D1: Model Download Priority

Recommended: Wan 5B first — Smallest model (10GB), fastest to validate the full stack end-to-end. LTX 2.3 second for Blackwell CUDA compatibility testing.

Alternative: LTX 2.3 first if Blackwell optimization is the primary concern. But Wan 5B gives a working pipeline in less time.

D2: Overnight Schedule Window

Recommended: 10pm–7am (9 hours) — Covers most batch workloads. Voice calls rarely come after 9pm for current tenants (GDS, CRU, North Group, noboxAI).

Alternative: 11pm-6am (7 hours) preserves more voice pipeline availability. Consider if tenant call volume grows past 10pm.

D3: Output Delivery Method

Recommended: Caddy file share (LAN) — Add a /media/ route to existing Caddy reverse proxy. Simplest first step, zero new infrastructure.

Alternatives: FTP to client delivery server or CDN integration. These can layer on top later once output volume justifies external access.

D4: Workflow Complexity

Recommended: Single-model workflows first — Prove the stack works with simple text-to-video before adding multi-model pipelines (FLUX.2 keyframe → Wan animation).

Multi-model variants layer on top of proven single-model templates. Reduces debugging surface during initial setup.

D5: Blackwell CUDA Fallback

Recommended: Accept PyTorch nightly build — If the official ComfyUI image lacks sm_120 kernels, build from PyTorch nightly base. Adds ~10min to first build, one-time cost.

Same class of issue that forced Kokoro TTS to CPU image. Blocking on upstream image updates is unpredictable.

Deferred Watch List

External dependencies to monitor. No action needed until activation triggers fire.

Item	Expected	Activation Trigger	Impact
Wan 2.7 (MoE 27B)	Mid-late Q2 2026	Open weights on HuggingFace (Apache 2.0)	Supersedes Wan 2.2 as quality lane. Voice cloning + instruction editing.
LTX 2.3 NVFP4	TBD	ComfyUI NVFP4 node + quantized weights	2.5x performance, 60% less VRAM on Blackwell.
n8n Integration	After CP-5	Automation scripts proven in production	Webhook-triggered renders, completion notifications.
Grafana Dashboard	After CP-3	ComfyUI running + nvidia exporter	GPU temp, VRAM, render queue, disk usage monitoring.
Multi-Model Pipelines	After CP-4	3+ single-model workflows proven	FLUX.2 keyframe → Wan animation, ControlNet variants.