🎨 Hrio Design｜设计师模板面板

Hrio_Design_Template_Node (🎨 Hrio Design｜设计师模板面板) is the flagship of the Hrio Design pack, and it's genuinely the cleverest node here. Instead of making you write three prompts, it ships a built-in template panel: pick one of twelve design briefs, and the node fires off three concurrent cloud generations - three proposals you can compare side by side. That's the whole job of this pack compressed into one node: a designer's proposal machine for graphic-design (平面) and interior-design (室内) pitches.

There's a real web editor behind this - the pack registers a /hrio-design/editor route in ComfyUI with theme skins, designer-type and output-strategy options, creativity levels, and a prompt editor. So the node is the ComfyUI end of a small product, not just a sampler wrapper. Keep that in mind: a lot of its behavior (the mode dropdown, the retry loop, the caching) is built to be driven from that UI.

And to be clear about the mechanism, since the input names invite confusion: mode is a prompt template, model is the actual generator. The tooltip says it outright - the frontend syncs mode and never touches model. Underneath, it's the same private Gemini-format backend as the rest of the pack. No local weights, no GPU, API key required, pay per call, your prompts and reference images leave your machine.

How it works

Three concurrent requests. The pack uploads any reference images once and reuses them across all three calls (the tooltips say the same batch of images is shared by the front/side/back requests), fires variant A/B/C together, and collects the results. There's an auto-retry loop: auto_retry_until_success (on by default) retries a view until it succeeds or you hit the cap, and the views are cached, so a partial failure doesn't nuke the good ones.

The inputs that matter

• mode - the template dropdown with 12 choices: six graphic (brand KV, event/promo poster, social/小红书 cover, web hero/banner, PPT/proposal cover, brand extension) and six interior (space-style concept, mood render, material moodboard, soft-furnishing plan, wall/floor material swap, lighting-mood adjustment). Pick a brief, get three takes.
• model - banano2 (default), banano-pro, veo3.1, or gemini3.1-pro. This is the actual image model; changing it changes style, not the template.
• api_key - blank falls back to config.ini's [banana] api_key.
• image_size - default 4K here (note the difference from the pack's other nodes); all three images use the same size. aspect_ratio - default 16:9, shared across the three views.
• generate_scope - regenerate all three, or only the front/side/back. Pick "only regenerate the front" and the other two reuse the last successful cached results instead of re-billing you. This is the node's best cost-control feature.
• auto_retry_until_success, max_retry_per_view (default 8, tooltip recommends 5–12), retry_interval_sec (default 1.5) - the retry loop. Crank max_retry_per_view way up and the node will run for a very long time; the tooltip warns you.

The optional mode_actual and cache_key are frontend-sync internals you should leave alone. image_1–image_10 are shared reference images. And automation_payload is how the pack's UI drives batch runs - there's a whole automation subsystem here that scans input folders (files like 001.png or project subfolders 001/front.png) and groups jobs by number.

Outputs

• variant_a_image, variant_b_image, variant_c_image - three IMAGE outputs, one per proposal. Wire all three into a preview/save stack to compare.
• images - an IMAGE output (likely the three stacked) for downstream processing in one stream.
• info - a STRING with run metadata. mp4url - a STRING URL, present because the template panel can also produce a video of the result set.

Installing it

In ComfyUI Manager search Hrio-Design, or:

cd ComfyUI/custom_nodes
git clone https://github.com/moki-flank/Hrio-Design

Restart ComfyUI. Dependencies are only requests, numpy, pillow; there are no model downloads because everything is server-side.

Where people get burned

Two things. First, this is a metered cloud node running three concurrent paid generations per run, plus retries on top - a bad run with retries is a stack of bills, which is exactly why generate_scope exists. Second, remember the mode/model split: people hunt for a missing "style" field in mode when the real lever is model. And the same pack-wide caveats apply - verify_ssl = false in config.ini, a key you can only get from the author, and effectively zero community footprint on Reddit if something goes wrong.