Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
InfrastructureInsight · On RadarHabr AI→ original

Imagecaster trains a neural network on your face from 10–100 photos and tries on clothes at home

Imagecaster is a program for generating images on a home GPU, and its author has taught it to draw a specific person. The neural network is fine-tuned with a small LoRA add-on trained on 10–100 photos, and the new frame shows that person in the required scene and clothing. The Klein model transfers clothing from a marketplace photo, while the face is swapped in a separate step so the likeness does not drift. A failed frame is therefore simply discarded rather than paid for, as it would be in the cloud.

AI-processed from Habr AI; edited by Hamidun News
Imagecaster trains a neural network on your face from 10–100 photos and tries on clothes at home
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The stylist now works inside the app itself, not only through MCP
Key metric #2Training runs on ai-toolkit, and generation runs on stable-diffusion.cpp
Defects remainthree hands, four toes on one foot, a missing jacket

Generating images on a home GPU is no longer surprising to anyone. It is harder to get output that is not a "similar girl from a magazine cover" but a specific person: yourself, your wife, or a friend, in the right clothes and scene. The author of the open-source Imagecaster app has added exactly that. The program now fine-tunes a neural network on 10–100 of your photos (using a small LoRA add-on), then swaps the face for an exact likeness, and transfers clothing from a marketplace or smartphone photo onto the subject with a separate model. Everything is computed locally, with no cloud subscriptions and no payment per generation.

How the chain works

LoRA is a small set of additional weights on top of a base model. Think of a universal studio where you bring a "personal attachment" carrying the traits of one person. Each base model needs its own LoRA: one trained for one model will not work on another. Unlike a simple face swap on a finished photo, here a new frame is created first with the required scene, pose, and angle, and the person is already built into it.

Training is run by a server based on ai-toolkit, and the images are rendered by stable-diffusion.cpp. In the examples, the author uses the fast Z-Image model: 8 generation steps, the Euler sampler, CFG 1, and frames mostly at 1024×1536 pixels. There is no binding to a single model; you can choose from those the app supports.

The stages are separate, which makes it easier to find errors:

  • LoRA creates the person in a new scene;
  • Klein, a separate model, transfers clothing from the garment photo;
  • a face swap at the very end restores the likeness.

The last step is needed because, according to the author, the LoRA alone sometimes produces only a "slightly" similar person and can change the face from frame to frame. If there is no face in the frame, this stage is skipped.

Three ways to train a LoRA

The first is to point to a folder with photos of one person. You need 10 to 100 snapshots in JPEG, PNG, or WebP. Variety matters more than quantity: ten different frames with different angles, lighting, and clothing are worth more than ten nearly identical ones.

The second is to build the set gradually through the "My Images" section. Group photos can be added there too: the app finds faces, compares them with the selected person, and crops out the right one. If the match is ambiguous, it will ask you to indicate the face manually, and if the face is small or blurred, it will warn you. You can start with ten photos and quick training, check the result, and later press "Fine-tune": the old snapshots remain, and the new ones are used more often. There is also retraining from scratch and rollback to a previous version.

The third is to hand the launch to an AI agent via MCP. This is an open protocol through which Claude, Codex, and similar clients control external programs. It is enough to write: "Train a LoRA for Ada on the photos from this folder using Z-Image." The agent gets the list of models, creates a job, and, if you wish, postpones the run so it does not occupy the GPU right away.

Stylist, privacy, and defects

The stylist previously worked only through MCP; now it is built into the app. In it you choose the person, model, style, clothing (optional), hairstyle, makeup, background, and poses, and then a shoot plan is drafted. Through MCP, the agent can suggest looks on its own and send frames to the queue.

The author dealt with privacy separately: the original photos do not get into the history of a "smart" cloud AI. They are described by a local vision model, and MCP sends outward only text and result numbers.

In the test, the agent proposed three series: a studio set with a graphite suit, an urban evening with a burgundy trench coat, and a garden with a sage-green dress. In the studio series, the model left out the jacket. The agent noticed this itself, clarified the clothing description, and re-rendered the frames, and the jacket appeared.

The author also honestly showed failed frames. In one, the model made the heroine heavier and did not show the face; in another, there were four toes on one foot; in a third, there were three hands. The fictional Ada comes out too "perfect," but on real people, in the author's view, the result is more interesting and the recognizability is good.

The economics are simple. Shooting is never free: you need hardware, electricity, and time. But a failed attempt does not deduct paid credits, as it would with cloud services that charge per generation. You can make dozens of variants and discard the bad ones.

The practical takeaway for developers: the combination of "LoRA + separate face swap + separate clothing transfer" is more reliable than one model for every case. And an agent that looks at the result and corrects its own prompt is already capable of fixing its own mistakes. The app and the server can run on one computer or be spread across a home network; the source code and builds are published.

Story

Key Questions Answered

How the chain works?

LoRA is a small set of additional weights on top of a base model. Think of a universal studio where you bring a "personal attachment" carrying the traits of one person. Each base model needs its own LoRA: one trained for one model will not work on another. Unlike a simple face swap on a finished photo, here a new frame is created first with the required scene, pose, and angle, and the person is already built into it.

What are the key details on Three ways to train a LoRA?

The first is to point to a folder with photos of one person. You need 10 to 100 snapshots in JPEG, PNG, or WebP. Variety matters more than quantity: ten different frames with different angles, lighting, and clothing are worth more than ten nearly identical ones. The second is to build the set gradually through the "My Images" section. Group photos can be added there too: the app finds faces, compares them with the selected person, and crops out the right one.

What are the key details on Stylist, privacy, and defects?

The stylist previously worked only through MCP; now it is built into the app. In it you choose the person, model, style, clothing (optional), hairstyle, makeup, background, and poses, and then a shoot plan is drafted. Through MCP, the agent can suggest looks on its own and send frames to the queue. The author dealt with privacy separately: the original photos do not get into the history of a "smart" cloud AI. They are described by a local vision model, and MCP sends outward only text and result numbers.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…