Character Studio: train a character and use it
How to get a character that looks like your photos: which photos to pick, how training works on your own graphics card or on LU Cloud, and the steps after training that decide whether the likeness holds. Written for version 3.0.6 from the app's source.
What it does, and what it does not promise
Character Studio trains a LoRA: a small add-on file for an image model that has learned one person or character from your photos. After training, a trigger word in the prompt brings that character into the picture. The lane is in Create, tab "Character", with two parts: "Train new" and "Use character".
A finished training run does not guarantee a perfect likeness. How close the result gets depends on the photos, on using the model the character was trained for, on the LoRA strength, and on a prompt that describes the character's defining features. This page goes through each of those, because each one is a common reason for a disappointing first result.
One more thing to know: the chat model does not know how the built in trainer works. If you ask it to build a "training workflow", it may invent one. You never need a workflow or a node for this; everything happens in the Character tab.
Local or Cloud, and the Mac
| On your own graphics card (Local mode) | On LU Cloud (Cloud mode) | |
|---|---|---|
| Windows and Linux | Yes, with an NVIDIA card with about 12 GB of memory. AMD cards can train on Linux only, and that path is not proven. | Yes, no graphics card needed |
| Mac | No. On the Mac the Character tab shows as a Cloud tool in Local mode. | Yes |
| Photos | 4 to 100 | 4 to 30 |
| One time setup | About 3 GB for the trainer, about 19 GB of base files | None |
| Picking the model afterwards | You pick a Z Image model yourself | Automatic |
The Cloud/Local switch is at the top right of the app. LU Cloud needs an account with credits; the prices are on the pricing page.
Pick the photos
The stage says it in one line: "One person or character, varied angles and lighting works best". In practice:
- One subject only. No group photos, no second person in the frame. Crop them out first.
- Variety. Different angles (front, side, three quarters), different light, different backgrounds and expressions. Ten good, varied photos beat thirty that look the same.
- Clear faces. Sharp, not tiny in the frame, not covered by sunglasses or hands in most of them.
- Files. PNG, JPG or WebP. Drop them on the stage ("Drop 4 to 30 photos of your character here" in Cloud mode) or use the add button. The counter shows how many you have and says "need at least 4" until you reach four.
If you saved a character from a video earlier, "or start from a saved character" on the stage adds its photos in one click.
Train on LU Cloud
Turn on Cloud at the top right, open Create, then "Character", then "Train new".
Add 4 to 30 photos. Type a trigger word into the field "Trigger word, e.g. davechar". Use one made up word that is not a normal English word, so the model cannot confuse it with something it already knows: "davechar" works, "dave" or "man" do not.
Press "Create". "Training runs in the cloud and lands the character on your shelf." When it is done, the character appears under "Use character" with the name of the model family it was trained for. The download button next to it saves the trained file to your computer.
Train on your own graphics card
In Local mode, open Create, "Character", "Train new". The panel reads "Trains fully on your GPU. One time setup, about 3 GB." Press "Set up trainer". It downloads the trainer, makes its own Python environment and installs PyTorch into it; the status line names each step. The trainer needs Python 3.10, 3.11 or 3.12; on Windows the app installs Python 3.12 if it finds none. The first setup also offers a field for the install folder, if your system drive is small.
Next the panel reads "Z Image training base files are missing (about 19 GB, one time)." Press "Download base files". They go to your ComfyUI models folder ("Goes to your configured model folder (Settings, ComfyUI).").
Add 4 to 100 photos and a trigger word, as for Cloud above. Pick how long it trains: "Quick" (400 steps), "Standard" (1200, the default) or "Thorough" (2400). Start with Standard. Quick is for a test run; Thorough can help when Standard is not close enough, but takes about twice as long.
Press "Create". "Runs on your GPU and takes a while. The local chat model pauses for the run. The character lands in your local LoRAs." The status line counts the steps. Do not start image or video jobs during the run: the card is busy. At the end the status reads "Character ready" with the file name and the trigger word.
Use the character
Open Create, "Character", "Use character", and click your character. The line under it tells you what to do: on Cloud "Put "(trigger)" in your prompt to summon the character.", locally "Put "(trigger)" in your prompt. Works best with a Z Image base model." Picking a local character switches its LoRA on and any other character's LoRA off.
This is the step most people miss. The local trainer trains against Z Image. Picking the character does not change the image model, so if a different model (for example a Qwen or Flux model) was selected before, it stays selected and the character will not look right. Open the model picker in the prompt box and choose a Z Image model. On Cloud this happens by itself: the character runs on the model it was trained for.
Put the trigger word in the prompt, then describe the character's defining features in plain words: hair colour and length, eye colour, skin tone, age, build, glasses, beard, a typical piece of clothing. The trigger word carries the face; the description keeps the rest from drifting. Example: photo of davechar, short curly dark hair, green eyes, light stubble, wearing a grey hoodie, sitting in a cafe, soft window light.
When the likeness is off
Go through these in order. Change one thing at a time and keep the seed fixed (Advanced settings, Output, "Seed (−1 = random)"), so you can see what each change did.
- Right model? Locally: is a Z Image model selected? This alone fixes most bad first results.
- Trigger word in the prompt? Spelled exactly as you trained it.
- Features described? Add the defining features to the prompt, as above.
- Strength (local). Under Advanced settings, Expert, "LoRA stack", each active LoRA has a "Strength" slider. If the character barely shows, raise it a little; if pictures look burnt, stiff or all the same, lower it. Change it in small steps. Keep other LoRAs off while you test.
- The photos. If nothing helps, the photo set is usually the cause: too few angles, too similar, or a second face in some of them. Train again with a better set, or try "Thorough".
When it goes wrong
"This card has (N) GB of memory and the local training recipe needs (N) GB…" The card is too small for local training. Use Character Studio in Cloud mode, which trains the same character without this limit.
"Training a character needs a PyTorch build that can drive your AMD card…" Local training needs an NVIDIA card, or Linux with ROCm. The rest of the app keeps working on the AMD card, and nothing was downloaded. Train in Cloud mode instead.
The trainer setup broke or a run fails at the start. The run repairs its own environment in most cases and names the cause. If it keeps failing, "Reinstall trainer" under the train controls sets it up again; then send the log file from Settings, General, Troubleshoot when you ask for help.
"No local characters yet. Train one first." under "Use character", although a run finished. The list reads the LoRAs that ComfyUI sees. Make sure ComfyUI is running, then open the tab again.
The character looks right in some pictures and not in others. That is normal for a LoRA. Fix the seed once you have a good one, and keep the description in the prompt.
I am on a Mac and the Character tab says it runs on LU Cloud. Local training is not available on the Mac. Switch to Cloud mode to train and use characters there.
More on the Create lanes: Create: images, video, audio and characters. Back to the handbook overview.