frankxai video-engine-routing — when you outgrow your default engine
The default stack (see video-gen) assumes a subscription engine. This skill is the
escalation path: what to do when credits run out, volume grows, or the pipeline needs
control the subscription engine cannot give.
The decision tree
- Under ~50 clips/month, mixed styles, no custom pipeline → stay on the subscription engine (Higgsfield or equivalent). Sunk cost, routing across many models, zero ops.
- Bursty overflow or budget clips at volume → pay-per-use API (fal.ai). No subscription, no credit expiry, cheapest per-clip for budget models.
- Custom pipeline: LoRA character consistency at batch scale, ControlNet/IPAdapter chains, motion transfer, community models → cloud ComfyUI (RunComfy machines or Comfy Cloud). You are paying for control, not per-clip price.
- Local GPU rendering → only with a recent NVIDIA card with 16 GB+ VRAM. On an iGPU machine (Intel Arc/AMD APU, shared memory): never. Modern video models blow through shared-memory budgets, and quantized variants degrade video coherence. Cloud is faster and cheaper than wasted local hours.
Pricing reference (mid-2026, one 5-second 1080p clip)
| Route | Cost per 5s clip | Notes |
|---|---|---|
| Higgsfield subscription | ~$0.30–0.50 in credits | tiers $15–129/mo; credits don't roll over |
| fal.ai — Seedance 2.0 Fast | ~$0.11 | cheapest respectable social clip |
| fal.ai — Kling | ~$0.35 | pay-per-second, no minimums |
| fal.ai — Veo 3.1 Fast | ~$0.50 (720p) | 4K tier ~3× |
| RunComfy machines | $0.99–9.59/hr + Pro $20/mo | custom ComfyUI workflows; cold-start overhead makes single clips uneconomic |
| Comfy Cloud | $20/mo ≈ 180 GPU-min | official ComfyUI cloud, credit-based |
| ComfyDeploy | — | defunct; stack was open-sourced, managed service gone |
Recompute before deciding: prices move quarterly. The structure (subscription vs per-use vs machine-rental) is what stays stable.
Graduation triggers
- To fal.ai: monthly credit top-ups exceed ~$20 two months running, or you need one premium model your subscription lacks.
- To cloud ComfyUI: you need the same character across 10+ videos via LoRA, or a ControlNet/IPAdapter/motion-control chain no API engine exposes, or 100+ clips/month of one repeatable workflow where machine-hours beat per-clip pricing.
- Back down: if a rented workflow runs less than monthly, cancel the machine and take the per-use price. Idle infrastructure is the most expensive engine.
What not to do
- Don't install vendor API-wrapper skill packs for engines you don't pay for — every one of them is lock-in to that vendor's endpoint with no portable workflow definition.
- Don't run modern video models on integrated GPUs (see rule 4 above).
- Don't hold two overlapping subscriptions to hedge; pick one default and use per-use APIs as the hedge.
Credits
The single-dispatcher/model-leaf routing pattern used by video-gen is adapted from the
router design in agentspace-so's RunComfy skill pack — pattern absorbed, dependency skipped.
Disclosure: I am a Higgsfield Ambassador; Higgsfield links on my site may be affiliate links. This skill's routing logic does not depend on that relationship — the decision tree sends you to fal.ai or ComfyUI whenever those genuinely win.