Qwen Image 2.1 — ComfyUI Serverless Worker

A production-ready Runpod Serverless worker running Qwen Image 2.1 (official Comfy-Org INT8 ConvRot) on ComfyUI — text-to-image and image editing with wildcard prompt expansion, an uncensored Qwen3-VL prompt enhancer, and Pony/Illustrious checkpoint workflows.

Built for ForgeHub, a desktop + self-hosted frontend that drives these workflows — but the endpoint speaks plain Runpod API, so anything can call it.

Get the app

Deployed the endpoint? Drive it with ForgeHub — workflows, wildcard browsing, job queue and outputs included. Paste your API key and it detects the endpoint by itself.

Workflows included

Workflow	Description
QwenImage21-T2I-Wildcards.json	Text-to-image, wildcard expansion + QwenVL prompt enhancer, detail LoRA
QwenImageEdit21-Wildcards.json	Instruction-based image editing (1–2 reference images)
PimpMyPony-HiResFix-FaceDet.json	Pony/Illustrious txt2img — wildcards + TagForge, hires fix e face detailer in gruppi bypassabili

Requirements

• GPU: 48GB class (ADA_48_PRO pool — e.g. RTX 6000 Ada / L40S)
• Network volume: ≥60GB attached to the endpoint at /runpod-volume. On first boot every worker auto-downloads missing models from models-manifest.txt (~45GB — the first cold start takes a while; subsequent workers reuse the volume). Disable with VOLUME_AUTOPOPULATE=false if you manage the volume yourself.
• CUDA: 12.9+ hosts (image is CUDA 13.0)

Models (auto-provisioned)

File	Source
qwen_image_2.1_int8_convrot.safetensors	Comfy-Org/Qwen-Image-2.1
qwen3-vl-8b-heretic-1.3.0-int8convrot.safetensors	craftingmod/Qwen3-VL-8B-Heretic-INT8
qwen_image_2.1_vae_bf16.safetensors	Comfy-Org/Qwen-Image-2.1
elusarcas-qwen2-1-detailer-v1.safetensors	reverentelusarca/elusarcas-qwen-2.1-detail-enhancer-lora
ilustmix_v111.safetensors, pimpmypony_pmpInCaseEnhanced.safetensors	huchukato/garage
Qwen3.5-9B-…-Q6_K.gguf + mmproj-BF16.gguf	DavidAU/Qwen3.5-9B-The-Defiant-Fable-…-GGUF

Job input

{
  "input": {
    "workflow": "QwenImage21-T2I-Wildcards.json",
    "prompt": "a __pmp/prmpt/lctns__ street scene, cinematic light",
    "image_size": "1344x768",
    "seed": 12345,
    "images": ["https://… or base64"]
  }
}

• workflow — file in /opt/workflows (see table above), or pass "prompt_graph": {…} with a full ComfyUI API-format graph (this is what ForgeHub does)
• prompt — patched into every prompt widget and WildcardProcessor node
• images — URLs or base64, mapped to LoadImage nodes in id order
• image_size / resolution, seed, params ({"node:widget": value} overrides), vl_preset, enhance
• "action": "health" — returns {"status": "ok", "workflows": […], "volume_mounted": bool} without running a generation

Custom nodes baked in

ComfyUI-QwenVL-Mod · ComfyUI-TagForge (wildcards) · ComfyUI-Pixaroma · ComfyUI-KJNodes · rgthree-comfy

Companion stack

• 🎬 Video sibling: huchukato/runpod-minimax-h3 (MiniMax H3 Turbo T2VA/I2VA/FL2VA/R2VA)