ComfyUI FaceSwap — RunPod Serverless

Character generation with optional face swap, deployed as a RunPod serverless endpoint.

Endpoint

• ID: bblp777ptfep17
• API: https://api.runpod.ai/v2/bblp777ptfep17
• Build: Automatic on push to main

Request Format

{
  "input": {
    "prompt": "a confident woman in a business suit, professional lighting",
    "negative_prompt": "bad quality, blurry, deformed",
    "width": 832,
    "height": 1216,
    "steps": 35,
    "cfg": 2.0,
    "seed": 42,
    "image_url": "https://example.com/face.png",
    "face_description": "",
    "output": {
      "include_base64": true,
      "save_to_volume": false,
      "volume_path": "outputs"
    }
  }
}

Field	Required	Default	Description
prompt	yes	—	Character description
negative_prompt	no	""	What to avoid
width	no	832	Image width
height	no	1216	Image height
steps	no	35	Sampling steps
cfg	no	2.0	CFG scale
seed	no	random	Reproducible seed
image_url	no	—	Face reference URL (enables face swap)
face_description	no	""	Which face to pick from reference
output.include_base64	no	true	Return base64 PNG
output.save_to_volume	no	false	Save to network volume
output.volume_path	no	"outputs"	Folder on /runpod-volume/

Modes

• Text only: Omit image_url — generates character from prompt using CyberRealistic XL
• Face swap: Provide image_url — generates character, then applies face from reference using IPAdapter + InstantID
• Face picking: When face_description is provided alongside image_url, uses Qwen2.5-VL-3B vision model to pick the matching face from multi-face scenes instead of defaulting to the largest face

Models

Model	Purpose	Size
CyberRealistic XL v7	Primary SDXL generation	~6.5 GB
Juggernaut XI	InstantID face refinement	~6.5 GB
CLIP-ViT-H-14	Vision encoding for IPAdapter	~3.9 GB
IPAdapter FaceID Plus v2 SDXL	Face identity transfer	~0.8 GB
IPAdapter Plus Face SDXL	Face style transfer	~0.8 GB
InstantID ControlNet + Adapter	Face structure preservation	~1.7 GB
SAM ViT-B	Face segmentation masks	~0.4 GB
InsightFace (antelopev2 + buffalo_l)	Face detection/recognition	~0.3 GB
YOLOv8m Face	Face bounding box detection	~0.05 GB
Qwen2.5-VL-3B Q4_K_M (GGUF)	Vision-based face picking	~2.3 GB VRAM

All models are baked into the Docker image at build time for fast cold starts. Total VRAM at startup: ~21.7 GB. Peak during face swap: ~29.5 GB on a 32 GB GPU (~3 GB headroom).

Benchmark

Prerequisites

pip install requests
export RUNPOD_API_KEY=your_key

Quick Start

# 20 requests: text + face swap, portrait + square, seed for reproducibility
python benchmark.py \
  --count 20 \
  --modes text,face \
  --resolutions portrait square \
  --face-url https://files.catbox.moe/az73pf.png \
  --seed 42

Benchmark Configs

Full benchmark (text + face swap, all resolutions)

python benchmark.py \
  --count 20 \
  --modes text,face \
  --resolutions portrait square landscape \
  --face-url https://files.catbox.moe/az73pf.png \
  --seed 42 \
  --concurrency 3

Text-only performance test

python benchmark.py \
  --count 10 \
  --modes text \
  --resolutions portrait square \
  --seed 100

Face swap quality validation

python benchmark.py \
  --count 10 \
  --modes face \
  --resolutions portrait \
  --face-url https://files.catbox.moe/az73pf.png \
  --seed 200

High-res stress test

python benchmark.py \
  --count 10 \
  --modes text,face \
  --resolutions portrait landscape square \
  --face-url https://files.catbox.moe/az73pf.png \
  --concurrency 5 \
  --seed 300

CLI Options

Flag	Default	Description
--count	10	Total requests
--modes	text	Comma-separated: text, face
--resolutions	portrait	Space-separated presets (see below)
--face-url	—	Face image URL (required for face mode)
--face-description	—	Which face to pick from reference
--steps	35	Sampling steps
--cfg	2.0	CFG scale
--seed	random	Base seed (incremented per request)
--concurrency	3	Max parallel requests
--output-dir	benchmark_results	Where to save images + report
--timeout	600	Per-job timeout (seconds)

Resolution Presets

Name	Dimensions
portrait	832 x 1216
landscape	1216 x 832
square	1024 x 1024
small-portrait	640 x 960
small-landscape	960 x 640

Output

Images are saved to benchmark_results/<timestamp>/ with filenames like:

003_faceswap_1024x1024_steps35_cfg2.0_seed45.png

A report.json with full timing data is saved alongside the images.

The first request runs sequentially to capture cold start time. Remaining requests run concurrently. The summary shows per-resolution/mode breakdowns.