Consistent characters
Keep the same face, outfit, and style across every generation.
Getting the same character in shot after shot is the difference between a gallery and a film. Ehudai gives you three levels of control.
Level 1: Reference images
Every reference-capable image model accepts your images alongside the prompt. Generate or upload one strong character image, then attach it as a reference to every generation of that character — paste it, drag it from history, or select it on a moodboard canvas.
For the strongest identity carry-over, use edit-mode models (Seedream Edit, GPT Image 2, Qwen Edit): they treat the reference as the subject to keep, not just inspiration.
Level 2: Character anchoring in stories
In AI Studio, build your cast on the story's Moodboard (the + on the character rail). When generating images for a shot, attach the character's moodboard image as an anchor — every shot generation of that character then starts from the same face.
In Auteur this is systematic: each character gets an identity plate rendered once at high quality, then referenced into every costume board, storyboard, and clip that character appears in.
Level 3: Train a custom LoRA
For a look no reference image can pin down — a specific character across wildly different scenes, or a signature style — train a custom model from the Moodboard.
Preparing training data
Training uses before/after image pairs in a ZIP:
training.zip
├── shot1_start.jpg ← input image
├── shot1_end.jpg ← what it should become
├── shot1.txt ← edit instruction for this pair
├── shot2_start.jpg
├── shot2_start2.jpg ← optional extra reference for the pair
├── shot2_end.jpg
├── shot2.txt
└── ...- Every pair needs a
_startand_endimage with matching names. - Give each pair its edit instruction in its own
.txtfile, or fill in the dialog's Default caption to cover every pair that has no file. A pair with neither falls back to a generic caption, so caption every pair one way or the other. - Extra references for a pair are named
_start2,_start3, and so on.
Training
Click Train LoRA in the Moodboard toolbar to open the Train Model dialog (the same dialog on the standalone Moodboard and on a story's Moodboard), then fill in:
| Field | What to enter |
|---|---|
| Model | Qwen Image Edit Plus — LoRA trainer (the one option today) |
| Training Name | A name you'll recognise later |
| Reference Image Count | How many _start images each pair has — 1 to 3 |
| Default caption | Optional. The edit instruction for any pair without a .txt file |
| Training Data (ZIP) | Your ZIP file |
Start the training (80 credits). The number of training steps is fixed and can't be changed, so the quality of your pairs is what you control: consistent, well-captioned pairs matter most. Training runs in the background — you can close the dialog, and the new LoRA appears in your trained list when it finishes.
Using it
On a Moodboard, either the standalone tool or your AI Studio story's board, click Select LoRA and pick up to 3 trained models under Select Trained LoRAs. Selected LoRAs appear as chips under Selected LoRAs and steer every generation on that board toward your trained look. Combine with a reference image of the character for maximum consistency.
On the standalone Moodboard, LoRAs run on Qwen Image Edit Plus LoRA. Select at least one canvas image as the base for it to edit; the model picker stays locked to that model, and shows its price, until you remove the LoRAs. See Moodboard.