RKNN LCM Client Nodes for ComfyUI

This repository provides ComfyUI custom nodes that allow ComfyUI workflows to call an external LCM Stable Diffusion server running on Rockchip NPUs (RK3588 / RKNPU2) over HTTP.

These nodes are clients only. They do not run inference locally inside ComfyUI.

---

What is the LCM Server?

The LCM server is a standalone FastAPI service that runs:

• Stable Diffusion 1.5 – Latent Consistency Model (LCM)
• RKNN2-accelerated inference on Rockchip NPUs
• Optional server-side Super-Resolution (SR) as a post-process

The server lives here:

👉 LCM Server Repository

Key characteristics of the LCM server

• Runs on RK3588 / RK356x devices
• Uses RKNN2 (not CUDA, not CPU inference)
• Optimized for low-step LCM generation (2–8 steps typical)
• Supports:
  • /generate → text-to-image
  • /superres → image super-resolution
• Manages:
  • Worker isolation (one RKNN runtime per worker)
  • Request queuing / backpressure
  • Optional HAProxy / multi-node routing

This ComfyUI plugin simply calls that service.

---

What This ComfyUI Plugin Does

This plugin adds HTTP client nodes to ComfyUI that:

• Send prompts or images to the LCM server
• Receive finished images back
• Convert results into ComfyUI IMAGE tensors
• Fit cleanly into standard ComfyUI workflows

What it does not do

• ❌ No local Stable Diffusion inference
• ❌ No CUDA / GPU usage
• ❌ No model loading inside ComfyUI
• ❌ No sampler / latent manipulation

Think of these nodes as remote render nodes.

---

Provided Nodes

1. RKNN LCM Generate (HTTP)

Calls the server's /generate endpoint.

Purpose

• Generate images using LCM-SD on an external Rockchip device.

Inputs

• API base URL(s) (single or multiple, round-robin)
• Prompt (multiline)
• Image size (e.g. 512x512)
• Steps
• CFG
• Seed (-1 = random)
• Super-Resolution magnitude slider:
  • 0 = off
  • 1–4 = number of SR passes (server may clamp)

Outputs

• IMAGE (ComfyUI tensor)
• Info string (backend, seed, SR status, timing)

---

2. RKNN SuperRes Upload (HTTP)

Calls the server's /superres endpoint.

Purpose

• Super-resolve an existing ComfyUI image using the server's RKNN SR model.

Inputs

• API base URL(s)
• IMAGE input
• SR magnitude:
  • 0 = bypass
  • 1–4 = number of SR passes
• Timeout

Outputs

• IMAGE (upscaled)
• Info string (passes, scale, backend)

---

Backend Selection & Scaling

The nodes accept multiple API bases, for example: