Redrob Studio

English | 한국어

A Next.js studio for image and audio tools, backed by ComfyUI on your own local GPU. Nothing is sent to a hosted inference API: the web app talks only to a ComfyUI process you control.

There is no chat surface and no video generation stack — home is a tools catalog; each tool opens its own workspace.

Built by Janghoon Lee (이장훈) for Redrob. Licensed under the Apache License 2.0.

Repository: redrob-labs/redrob-studio

Requirements

Layer	Minimum
Node.js	20+ (see package.json engines)
Package manager	yarn
OS (app)	Any OS that can run Next.js
OS (GPU host)	Ubuntu recommended; NVIDIA GPU + recent CUDA drivers
Disk (GPU host)	Tens of GB free for ComfyUI custom nodes and model weights
VRAM - Create only	~12 GB NVIDIA VRAM recommended (redrob-image UNET + CLIP + VAE)
VRAM - full stack	~24 GB+ recommended (Create + Qwen TTS/ASR/VL + ACE-Step music)

The Next.js app itself is lightweight. Generation runs on ComfyUI; without a GPU host you can still open the UI, but tools will not produce output.

Features

Modality	Tools
Image	Create (Redrob Image), Upscale (Lanczos), Compress, Convert, Vision
Audio	Speak, Clone, Design, Music (Redrob Tune / ACE-Step), Transcribe, Extract audio (from a video file), Convert

Tools are registered in src/workflows/manifest.json and backed by ComfyUI API workflow templates in src/workflows/, resolved server-side by id.

The UI ships in English and Korean. Switch with the EN / 한국어 toggle in the header; the choice is stored in localStorage. Messages live in src/lib/i18n/messages.js. Home (/) is the tools catalog; each tool opens its own workspace (for example /image/create).

Only OSS-safe weights are included: every shipped workflow depends on Apache-2.0, MIT, or BSD licensed models, or on no learned weights at all. The full inventory is in eval/THIRD_PARTY_WEIGHTS.md.

Architecture

Browser  --HTTP/SSE-->  Next.js (/api/comfy, /api/generate)
                             |
                             |  COMFYUI_URL (server-only)
                             v
                        ComfyUI backend

The browser never talks to ComfyUI directly. COMFYUI_URL is server-only and is never exposed as a NEXT_PUBLIC_* variable. Progress is relayed to the client as SSE from a server-side ComfyUI WebSocket subscription, and workflow graphs are assembled on the server from allowlisted templates (workflowId plus params only). See SECURITY.md for the deployment invariants.

Quick start

yarn install
cp .env.example .env.local   # set COMFYUI_URL
yarn dev

# .env.local
COMFYUI_URL=http://127.0.0.1:8188

COMFYUI_URL defaults to http://127.0.0.1:8188 when unset. Restart the dev server after changing .env.local.

GPU host setup

On an Ubuntu machine with an NVIDIA GPU, the bootstrap script installs ComfyUI, the required custom nodes, and the permissive weights. It does no cloud deploy and handles no secrets.

export COMFY_HOME=~/ComfyUI
# optional: export HF_TOKEN=hf_...   # only if a Hugging Face repo needs auth
bash scripts/setup-comfyui-local.sh                    # full stack
# bash scripts/setup-comfyui-local.sh --profile create # Create models only

Item	Path
ComfyUI	$COMFY_HOME (default ~/ComfyUI)
venv	$COMFY_HOME/venv
Models	$COMFY_HOME/models (override with COMFY_MODELS)

Start ComfyUI on loopback, then point Next.js at it:

cd "$COMFY_HOME"
./venv/bin/python main.py --listen 127.0.0.1 --port 8188

Keep ComfyUI on the host network only. You can open http://127.0.0.1:8188 on the GPU host to inspect graphs manually by dragging workflow JSON from src/workflows/ onto the canvas.

Create model

Image Create runs a single UNET, redrob-labs/redrob-image (task: redrob-image).

File under $COMFY_HOME/models	Role	Source
diffusion_models/redrob-image.safetensors	Merged UNET	redrob-labs/redrob-image
text_encoders/qwen_3_4b_fp8_mixed.safetensors	CLIP	Comfy-Org/z_image_turbo
vae/ae.safetensors	VAE	Comfy-Org/HiDream-I1_ComfyUI

Sampler defaults: 8 steps, CFG 1, res_multistep / sgm_uniform.

Troubleshooting

COMFYUI_URL is not set, or the proxy returns 502

Set COMFYUI_URL in .env.local and restart the dev server. The app reaches ComfyUI only through /api/comfy and /api/generate.

Cannot connect to port 8188

Confirm ComfyUI is listening on the URL in COMFYUI_URL:

ss -tlnp | grep 8188
curl -sS "$COMFYUI_URL/system_stats" | head

Missing node or model

Install the custom node pack on the GPU host and restart ComfyUI, or re-run scripts/setup-comfyui-local.sh. Model filenames and folders must match the workflow JSON exactly.

Inspecting ComfyUI errors

cd "$COMFY_HOME" && source venv/bin/activate
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
python main.py --listen 127.0.0.1 --port 8188

Repository layout

Path	Purpose
src/workflows/manifest.json	Studio task registry
src/workflows/*.json	ComfyUI API workflow templates
src/lib/workflows.js	Manifest loader
src/lib/comfyui.js	Prompt building, upload, progress
src/app/api/comfy/**	Server proxy and SSE progress relay
src/app/api/generate	Allowlisted generation entry
src/hooks/useToolRunner.js	Studio run state and options

License and attribution

Copyright 2026 Redrob. Built by Janghoon Lee (이장훈).

Original code in this repository is licensed under the Apache License 2.0. See LICENSE for the full text and NOTICE for attribution.

Third-party model weights are downloaded separately, are not redistributed here, and remain under their upstream licenses. See eval/THIRD_PARTY_WEIGHTS.md.

Security policy and reporting: SECURITY.md.