AI influencer ComfyUI workflow

What is ComfyUI?

ComfyUI is open-source software that runs AI models through a visual graph of blocks, called nodes.

It runs on Windows, Linux and macOS, and a desktop app installs it on Windows and macOS. The code is GPL-3.0 (ComfyUI on GitHub), so the software costs nothing. Your costs are the graphics card and your time.

What ComfyUI gives you is control. Every step of the picture is a block you can see and swap, and on your own card nothing leaves your machine. That matters when you make hundreds of posts of one character.

It also means nobody holds your hand. There is no character builder and no “keep this face” button, only the blocks below.

The workflow blocks, explained in plain words

Each block is one job. Read them left to right, the way the wires run.

This beginner tutorial installs ComfyUI, loads a first model and builds a text-to-image workflow.

How to Use ComfyUI (Step-by-Step Tutorial)Kevin Stratvert

Checkpoint loader: the base model

The checkpoint is the main image model, the file that knows how to draw. Everything else adjusts its output.

FLUX.2 [klein] 4B from Black Forest Labs is the safe start for a commercial account. Its model card puts it at about 13 GB of graphics memory, on an RTX 3090, 4070 or above, under Apache 2.0 (FLUX.2 klein 4B). The larger 9B version sits under the FLUX Non-Commercial License, so it stays out of anything that earns (FLUX.2 klein).

ComfyUI also lists Qwen Image and older FLUX models among the image models it supports.

LoRA loader: your character

A LoRA is a small trained file that holds one character’s face. The loader block sits between the checkpoint and the prompt, with a strength slider for how hard it pushes.

A LoRA holds the face without a reference photo in every prompt, and it takes a separate training run. Our guide to training an AI influencer LoRA covers the dataset and the cost.

Face identity node: one photo, one face

A face identity node copies a face from a single reference picture, with no training. Two open options are PuLID and InstantID.

InstantID works with SDXL, an older model family, from one image. PuLID has an SDXL version and a FLUX.1 [dev] one. The community ComfyUI node for PuLID, PuLID_ComfyUI, uses the SDXL model and has been in “maintenance only” mode since 14 April 2025.

Both nodes need an SDXL or FLUX.1 [dev] checkpoint, so they do not plug into a FLUX.2 [klein] workflow. On klein, the LoRA is the face route.

This masterclass builds a consistent photoreal character from one input image on open models (published October 2025).

Create HYPERREALISTIC Consistent AI Characters - FREE & LOCAL! [Full ComfyUI Masterclass 2025]Mickmumpitz

That clause is a reason to train a LoRA instead, since the LoRA route has no InsightFace step.

Upscaler: the final sharpening

An upscaler makes the finished image bigger and crisper for a feed or a print. Real-ESRGAN is the common open one: its x4plus model makes an image 4 times wider and taller, under a BSD-3-Clause licence.

Image to video: making the character move

The video block takes your best still and animates it. Wan 2.2 from Alibaba is the open choice, under Apache 2.0, and ComfyUI has supported it since 28 July 2025 (Wan2.2).

Its small model, TI2V-5B, runs on one consumer card with 24 GB. The Wan team says it makes “a 5-second 720P video in under 9 minutes” on a consumer GPU; this page prices that on an RTX 4090, a 24 GB card. The larger I2V-A14B model needs 80 GB on a single card.

Starting from the still keeps the face your LoRA or identity node made, instead of letting the video model invent one.

How to set it up, step by step

These steps follow ComfyUI’s own README and template library.

1. Install ComfyUI. On Windows or macOS the desktop app is the route the README calls “The easiest way to get started”. Linux uses the manual install.

2. Open a template. comfy.org/workflows lists 662 or more ready-made workflows by model and task, with image, image to video and motion control among the tags. Pick one built for your base model, so the blocks are already wired.

3. Download the models it asks for. The checkpoint, any LoRA and the upscaler each go into their own folder under ComfyUI/models/. PuLID’s model goes in ComfyUI/models/pulid/, for example.

4. Add your character. Insert a LoRA loader after the checkpoint and put the trigger word in the prompt. A face identity node fed with one reference photo works only on an SDXL or FLUX.1 [dev] checkpoint. It does not plug into a FLUX.2 [klein] workflow.

5. Generate a batch and keep the best stills. Change the setting and clothes in the prompt while the face block holds the person.

6. Upscale the keepers with Real-ESRGAN.

7. Animate one still with Wan 2.2 image to video. For dance or gesture clips, the template library also lists a “Wan Animate 2: Motion Transfer” workflow, which copies movement from a reference video.

For the prompt side, our consistent AI character guide covers how to describe the same person scene after scene.

What does it cost on a rented GPU?

Under a cent an image, and cents a clip, if you count GPU time only.

Runpod rents an RTX 5090 at $0.69 an hour on Community Cloud and $1.19 on Secure Cloud. An RTX 4090 runs $0.34 and $0.89.

Those rates come from Runpod’s pricing page, updated on 8 October 2026.

The table leaves out the time the machine runs before the first image. A rented pod bills from the moment it starts, so downloading a 13 GB model and trying a workflow run on the same rate.

One hour of setup on the Community Cloud 5090 costs as much as about 3,000 images.

Comfy Cloud is the hosted middle ground. It runs ComfyUI for you on Blackwell RTX 6000 Pro cards with 96 GB.

Its Creator plan at $35 a month is the first that imports your own models, LoRAs included.

Standard and Creator cap each workflow run at 30 minutes (Comfy Cloud pricing).

When does a ready-made app save you time?

When you need a character this week and do not want to manage models. ComfyUI costs less per picture. A ready-made app costs less of your time.

Higgsfield is the clearest example for this reader. Its AI Influencer Studio builds a character from 19 menu settings, with no prompt. Each run returns a portrait and a full-body shot for 1.12 credits, which Higgsfield prices at about $0.06.

The same studio animates that character from a 3 to 30 second reference clip, using up to 40 images a video, at 1080p (AI Influencer). That is the job the Wan 2.2 and motion transfer blocks do in ComfyUI.

Soul ID replaces the LoRA. You upload 20 or more photos and training takes about 3 to 5 minutes, by Higgsfield’s Soul ID post. The identity stays inside Higgsfield.

Plans start at $19 a month (25 October 2026). Our Higgsfield AI Influencer walkthrough shows each screen.

The trade is plain. ComfyUI gives you the files and open licences, at near-zero cost per image once it works. Higgsfield trains a character in 3 to 5 minutes with Soul ID, paid in credits.

The AI influencer hub and our best AI influencer generators list set out the other app routes.

Next on Higgsfield: keeping an AI character consistent →