ComfyUI Field Notes: Local vs Cloud vs Rented Pod
What I tried
I've been running ComfyUI on three different setups depending on what I'm making and how much I care about cost that day. The hardware at home is an RTX 4070 in a Legion laptop — accessible anywhere via Tailscale, fronted by a lightweight proxy at comfy.havoclabs.org. For anything that outgrows 12 GB VRAM, I spin up a rented GPU pod. And for video generation specifically, I've been running test prompts through Kling's API to get a baseline.
This note is about what the comparison actually looked like — not a benchmark, just a real working session with real friction.
Setup friction
Local is effectively zero friction once it's running. ComfyUI is installed, models are on disk, and the Tailscale tunnel keeps it accessible from anywhere as long as the laptop is powered on and connected. The catch is that last part. If the lid is closed, the laptop's suspended. If I'm traveling with it, it's not on the desk. The setup is great when it works; it's invisible when it doesn't.
Rented pod takes about 8–12 minutes to go from zero to usable on a cold pod. Provision the pod, wait for the Docker image to pull, wait for the ComfyUI template to start, then download whichever models aren't already cached in the pod's persistent volume. After the first session those models persist, so subsequent cold starts drop to 3–4 minutes. Not bad, but not nothing.
Kling API is instant. Paste your prompt, get a clip. No GPU management, no model versions to track. The tradeoff is you're working inside their prompt system with their model — no workflow customization, no ControlNet, no LoRA.
Speed and quality
For image generation with SDXL, the RTX 4070 runs a 1024×1024 generation in about 8–12 seconds at 20 steps. Usable. FLUX.1-schnell at the same resolution runs in 18–25 seconds — slower but the outputs are meaningfully better on detailed prompts. Both fit in 12 GB VRAM with moderate batch sizes.
The rented A5000 pod (24 GB VRAM) runs the same FLUX.1-schnell workflow about 40% faster, and can handle FLUX.1-dev without the memory compromises I make locally. For video, LTX-Video 2.3 barely fits on the 4070 at reduced resolution — the A5000 runs it at full resolution without drama.
Kling 1.6 produces smooth, coherent video that local LTX can't match yet on motion quality. For simple prompts it's very good. For anything with specific composition or character consistency, it's a coin flip.
What it cost
| Setup | Cost model | Practical cost |
|---|---|---|
| Local RTX 4070 | Hardware sunk cost + ~$0.10/hr electricity | ~$0 per session |
| Rented A5000 pod (24 GB) | ~$0.44/hr billed while pod is running | ~$1–2 for a typical 2–3 hr session |
| Kling 1.6 API | Credits per second of video generated | ~$0.10–0.18 per 5-second clip |
The local setup wins on cost by a wide margin for sustained work. The catch is idle time on the rented pod — I burned $0.80 one afternoon forgetting to stop a pod while I was doing something else. Stop the pod when you're done. It doesn't stop itself.
What broke
Local: The laptop suspended mid-generation once because it hit a thermal threshold. ComfyUI doesn't recover from that gracefully — the queue was gone on resume. Not a big deal for images, annoying for a 3-minute video generation. I now run longer jobs with a power plan that disables sleep.
Rented pod: The persistent volume wasn't mounted correctly on one pod restart — I lost my custom node installations and had to reinstall them. The pod's persistent-volume setup is worth reading before you assume your environment persists. It mostly does, until it doesn't.
Kling: Prompt restrictions are real. Anything that reads as conflict, weapons, or ambiguous character interactions gets flagged. I had a few rejections on prompts that seemed innocuous to me. No explanation given. You learn to rephrase.
Would I repeat it?
Local: Yes for images, yes for SDXL and FLUX.1-schnell workflows, and yes as the always-on option for quick iterations. No for video at full quality — 12 GB VRAM is the floor for LTX, not the ceiling.
Rented pod: Yes for video generation and for FLUX.1-dev. The cost is low enough for sporadic use. The setup friction is acceptable. The idle billing requires discipline. I'll keep using it for anything that needs more than 12 GB VRAM.
Kling: Yes for fast prototype clips and client-facing demos where I want polished motion without workflow overhead. No for production volume — the per-clip cost adds up fast and you're at the mercy of their content policy.
Local for daily image work. The rented pod for video and large-model runs. Cloud APIs for quick demos and anything that needs to be polished without a lot of setup.
The three don't compete — they cover different slots in the same workflow. The mistake is trying to force one into all three roles.