工具｜API 多模型并发生图引擎

Every other Tikpan image node asks you to pick one model and cross your fingers. The Parallel Image Engine is the node you reach for when you can't decide - it fires the same prompt at a list of models and a list of relay hosts at once, then hands you whatever came back. The use cases in the node's own description are A/B testing, batch image generation, and failover redundancy, and it genuinely does all three without making you build a spaghetti graph of duplicate generator nodes.

It's part of the Tikpan Pro pack, the official ComfyUI plugin for the Tikpan (攀升AI) API aggregator - a paid cloud service, prepaid tokens, key from tikpan.com. That means this node doesn't run a single weight locally. Every generation leaves your machine and costs per call (external-api-nodes.md). Keep that in mind before you set it to "run everything."

How it works

The mechanism is simple but well-built. You give it two multiline lists:

• Models - one model ID per line (defaults to grok-imagine-image and grok-imagine-image-pro). Lines starting with # are ignored.
• Relay_Hosts - one relay endpoint per line, defaulting to https://tikpan.com.

The node builds a job for every model × relay host combination, then runs them through a thread pool capped by Max_Concurrency (1–12). Your Strategy decides what "running" means:

• failover - tries jobs in order, stops at the first success. Default, and the cheapest. Best for daily commercial use.
• parallel_all - fires every job, returns every image. Great for batch/A-B, and exactly as expensive as it sounds.
• race_first_success - fires all of them and keeps whichever wins. Fastest wall-clock time, most wasted spend.

Inputs and outputs that matter

For a beginner, four fields do the real work: API_Key, Prompt, Models, and Strategy. The rest are tuning. Images_Per_Model (1–4) sets how many images each model generates per relay; Size is one of 1024x1024, 1792x1024, 1024x1792, applied uniformly. In optional land, Response_Format (url is recommended) and Timeout_Seconds (default 240) are worth knowing, and Skip_Error (default on) means a bad model or relay gets skipped instead of killing the whole batch - exactly what you want in a bulk job.

Outputs are Images (a single IMAGE batch - wire it into a Preview Image or Save Image node), Stage_Log (a human-readable STRING of what happened), and Result_JSON (the full per-job records if you want to inspect results programmatically).

Installing

This is the whole pack's install story, same for every Tikpan node: search "Tikpan" in ComfyUI Manager and install, or

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

then restart ComfyUI. Grab a sk- key from tikpan.com and paste it in. Nothing heavy installs - the pack rides on requests/PIL/torch, which ComfyUI already ships.

Common issues

• Rate limiting. Crank Max_Concurrency to 12 and the relays will notice. The tooltip says it plainly: bigger is faster but more likely to trip a 429. Start at 3.
• Black image or hard error when everything fails. If every model and relay fails and Skip_Error is off, the node raises. With it on you get a black placeholder and a log explaining why. In batch flows, keep it on.
• Double-check the cost before parallel_all. This is the one node in the pack where you can spend on four models at once without noticing. failover is the financially sane default.
• Security posture. It's an API node holding your key and phoning home by design - the same shape of node that's burned this ecosystem before. Read what a fresh install does; don't paste keys into nodes you don't trust (external-api-nodes.md § Security).