The most common flaw in AI video series is the character changing from one scene to the next: hair color shifts, clothes lose a detail, the face looks like someone else. Video models have no “memory” of a character between generations. What works is controlling the starting point of every scene, and that is where a rented GPU complements GenBrasil.
The idea in one line
Train the character on a GPU as a LoRA, generate each scene's first frames with it and use each image as the first frame on GenBrasil. The video model animates something that already has the right face.
When it is worth it
- A series or channel with a recurring character (mascot, virtual host, story protagonist).
- A campaign with several videos that must look like the same brand.
- When the text prompt, even a detailed one, can't hold the look. For a one-scene character you need none of this: use a fixed description block and a first-frame image.
Step by step
1. Build the character image set
15 to 30 images, varied in angle, expression and background, but with the same character. If the character is invented, you can generate those images with an open image model and select the best. If it is a real person, only proceed with written consent (see responsible use).
2. Train the LoRA on a rented GPU
A LoRA is a small adjustment that “teaches” an image model to reproduce that character. Training one for a character usually fits on a 24 GB GPU in a few dozen minutes to a few hours, depending on the base model and the number of steps. The LoRA guide explains the concept and the ComfyUI template brings up the environment. Check the base model's license before using the result commercially.
3. Generate the scenes' first frames
In ComfyUI, load the LoRA and generate one image per scene: the character on the beach, in the office, in the rain. An image costs a fraction of a video and iterates in seconds, so this is the place to get composition, clothing and expression right.
4. Animate each scene on GenBrasil
Upload the approved image as the first frame and write only the motion: what the character does and what the camera does. Don't redescribe the face, it is already in the image. Use the model ladder so you don't spend on the expensive model too early.
5. Review and assemble
Check each clip side by side: did the face hold throughout the scene? If the model “melted” the character mid-motion, simplify the action or switch models. Then join the scenes in an editor or with ffmpeg.
What the GPU side costs
It is an hourly machine cost, in reais, only during training and image generation. Prices per GPU live in the console catalog and change, so we don't pin them here. The decision rule is simple: if the character appears in more than two or three scenes, a LoRA usually pays for itself in attempts you stop discarding on video.
Limits
A LoRA guarantees the look of the starting image, not of the whole video. Long or very abrupt motion can still drift the face. Short scenes and simple action hold better. And no workflow replaces human review of each clip.
For the video side, open GenBrasil: the Free plan needs no card and shows the cost before generating.
Frequently asked questions
Do I need to train a LoRA for a consistent character?
Not for a single scene. For a recurring character across scenes, a LoRA plus using the image as first frame gives the most stable result.
How many images are needed to train a character LoRA?
Usually 15 to 30 well-varied images are enough, but it depends on the base model and the level of detail you want.
Does the LoRA solve consistency for the whole video?
It solves the look of the first frame. During motion the video model can still drift. Short scenes and simple action reduce that risk.
Make your video on GenBrasil
Eight AI video models in one account, with the cost shown before you generate and payment in reais by Pix or card. The interface is in Portuguese. The Free plan needs no card.
Go to GenBrasil ↗Rather build your own workflow? See the hourly GPU templates.
Read next: how to extend an AI video · who makes each video model