Video summary

How to Create an AI Influencer from scratch Without Training a LoRA

Main summary

Key takeaways

Technology

Summary of the subtitles (tech workflow + product features)

The video explains how to create an AI influencer from scratch without training a LoRA. Instead of training, it focuses on building a stable, consistent synthetic face identity from references (e.g., Pinterest). Then, it generates a clean creator dataset that can be reused for later generation and/or fine-tuning on-demand.

Core idea: “Direction first,” not training

The creator argues that most people mistakenly think the hard part is training a LoRA, when the real challenge is creating an identity that doesn’t fall apart after multiple generations.

Key emphasis:

  • Consistent identity across images (with special focus on SFW-like consistency, “save for work”)
  • A dataset that preserves the same character across:
    • poses
    • outfits
    • lighting
    • styles

Workflow overview (ComfyUI + custom nodes)

Uses ComfyUI with custom nodes from the Matrix Lab pack.

1) Reference selection via Pinterest (visual direction)

  • Don’t save every attractive image—this creates unclear signals.
  • Pick 3–5 images that share the same overall creator “type” (example: “blonde model selfie/portrait”).
  • The goal is not copying one specific person, but capturing a consistent look direction.

2) Face merge to create a new synthetic creator face

  • The references are fed into a face merge workflow in ComfyUI.
  • Generation models mentioned:
    • OpenAI “image 2” (chosen in this step for more realistic face/details)
    • Nano Banana 2 and Nano Banana Pro (also mentioned as useful)
  • A custom “face merge prompt” guides the workflow toward a unique creator face, rather than generating random AI “girls.”

3) Generation parameters for quality (dataset stability)

The face merge node is run with:

  • An API key entry (provider name mentioned as “waist speed” in subtitles; unclear exact service)
  • 2–4 generations initially (to compare results and manage token usage)
  • Resolution: 4K (called one of the most important settings to reduce later character drift)
  • Aspect ratio: ~3:4 (“3 to four”) Chosen to focus on the face while still allowing upper-body composition later.

4) Identity stability test (swap workflow)

Next is an identity swap test:

  • The newly created face is transferred onto a target image (example images from Pinterest).
  • The video stresses permission/copyright awareness: use only photos you’re allowed to use.
  • Evaluation checks include:
    • Does the face still match the base identity?
    • Are eyes and mouth stable?
    • Is skin realistic?
    • Does it look like one consistent character rather than a random face swap?

5) Repeat testing to build the dataset

  • Repeat face creation + swap testing 3 to 10 times
  • Purpose: collect enough strong, identity-consistent images.

6) Dataset generation workflow (multi-branch)

A “real data set workflow” runs with three branches:

  • Nano Banana 2 / Nano Banana Pro (top branch)
  • Image 2 (middle branch)
  • Cream 4.5 (bottom branch), noted for more “uncensored flexibility” for generating not-SFW datasets

Other details:

  • The workflow supports six image slots per run.
  • If you have more images (e.g., 10), only the best ones are selected.
  • You can also run with fewer images (e.g., 4).

7) Prompt stack for efficient dataset creation

A prompt stack data node is used:

  • Instead of rewriting prompts, you advance through them sequentially (e.g., “next prompt”).
  • The generation goal is not random aesthetics; it is:
    • maintaining the same creator identity
    • across varied scenarios (pose/outfit/lighting/style)

8) Why the dataset matters (LoRA-free but LoRA-ready)

The creator claims beginners often skip the dataset step and jump directly to:

  • LoRA training, or
  • random prompting

A clean dataset becomes an asset because it can be used to:

  • train LoRA “on almost every model,” including (as mentioned):
    • Set image turbo
    • Flux 2
    • Quen 2.2
    • (and others)
  • or generate content directly with API models for SFW—and potentially more advanced NSFW—content using the same identity base

Core promise:

  • avoids “random images where the face looks different every two images”
  • replaces it with controlled identity

Reviews / guides / tutorials mentioned

The video is positioned as an explicit workflow tutorial/guide for:

  • Building a synthetic AI influencer face without training a LoRA
  • Using ComfyUI workflows (Matrix Lab pack) for:
    • face merge from reference images
    • identity swap testing
    • dataset generation with prompt stacks

Free resources mentioned:

  • a prompt library (referenced later)
  • full Matrix Lab workflows and custom nodes via the Matrix Lab community (in description)
  • prompts from the video placed in a free digital community

Main speakers / sources

  • Speaker: the video author/instructor (referred to as singular “I” voice; no name provided in subtitles)
  • Tooling / sources used:
    • ComfyUI
    • Matrix Lab pack (custom nodes/workflows)
    • OpenAI “Image 2”
    • Nano Banana 2 / Nano Banana Pro
    • Cream 4.5
    • Pinterest (reference source)
    • Mentions Matrix Lab community + free digital community + resources distributed via the description

Original video