LemonSlice CWM-1 turns one photo into a live avatar

You've built a face you like. It talks, the lips line up, the voice fits. But the moment you watch it next to a real person on a video call, something is missing: the hands stay still, the shoulders never move, and the background looks like a photo. On September 23, 2026, LemonSlice released Character World Model-1 (CWM-1), an interactive avatar model that turns one image into a live character that moves its whole body while you talk to it.

• CWM-1 generates every frame of the character live during the conversation: face, body, hands and the scene around them. One image in, no fine-tuning or waiting.
• A built-in emotion engine picks actions and emotions as the conversation goes, from smiling, laughing, waving and nodding to touching hair, adjusting a collar, texting or taking a phone call.
• The background is generated too, so it isn't frozen: LemonSlice shows flowing water, people passing by, and shadows and reflections that follow the character. The company says a stream runs for 24+ hours without visible drift.
• Free to try in the LemonSlice web app with custom characters. API access is for Ultra and Enterprise customers, and CWM-1 Lite followed on October 2, 2026 at prices as low as $0.048 per minute.

For most of the last few years, a real-time AI avatar meant a talking head: a face that moves its lips in front of a still background. LemonSlice's own launch page makes the same point, that most of what makes a conversation feel real happens outside the mouth: shrugging, pointing, leaning in, fidgeting with your hair. CWM-1 is built to generate that part. It's a causal video diffusion transformer, which means it only looks at what has already happened and keeps generating the next frames as new audio comes in. LemonSlice says that if you interrupt the avatar with a question, it starts generating response frames in 471 milliseconds.

We covered whole-body motion before with Tavus Phoenix-4.5 in "AI avatars finally move their whole body". The difference with CWM-1 is how much of it you can steer. The same model works on a photo of a person, an anime character or a picture of a dog, and developers can either let the emotion engine choose actions on its own or trigger a specific action, such as a wave or crossed arms, at a specific moment. For a creator, that means the character you designed can greet a viewer, laugh at the right line and touch its hair without you animating any of it.

How to try live whole-body avatars now that CWM-1 is out

The following images were generated using Nano Banana 2:

1. Start from a waist-up photo with the hands in frame. CWM-1 animates hands, shoulders and posture, so give it a picture that shows them. A tight face crop leaves the model little body to work with.
2. Give the character something to touch. The action list includes touching hair, adjusting a collar, texting and taking a phone call. Loose hair, a visible collar or a phone on the table gives those actions something to land on.
3. Choose a background that can move. The scene is generated live, so a café with people behind, a rainy window or a workshop full of objects gives the model more to do than a plain wall.
4. If you build on the API, start automatic, then take control. In the docs, setting action_engine to "natural" lets LemonSlice pick actions from the conversation. You can then trigger your own; CWM-1 plays one action at a time, and actions such as blow_kiss and deep_breath work best while the avatar is silent.

Worked example — the café host. A waist-up portrait with both hands visible, hair loose and a sweater collar in frame, so a hair touch or a collar adjustment has something real to move.

Prompt (Nano Banana 2, 1:1):
Square editorial portrait, bright and well exposed. A Middle Eastern woman in her 20s with dark wavy hair, wearing a mustard yellow knit sweater, in a plant-filled café, facing the camera from the waist up, mid-gesture: one hand lifted to touch her hair, the other adjusting the collar of her sweater, relaxed warm smile. Green plants and soft daylight behind her, shallow depth of field. Natural skin texture, realistic hands with five fingers. No text, no logos.

Worked example — the workshop storyteller. A different kind of character: an older man mid-wave next to a workbench full of tools, with sawdust in the light, so the body and the scene both have room to move.

Prompt (Nano Banana 2, 1:1):
Square editorial portrait, bright and well exposed. A cheerful white man in his 70s wearing a tweed flat cap and a canvas work apron over a checked shirt, standing in a sunlit woodworking workshop, laughing with one hand raised in a wave and the other resting on a wooden workbench with hand tools. Sawdust in the light, shelves of timber behind him, shallow depth of field. Natural wrinkled skin, realistic hands with five fingers. No text, no logos, no signs.

The honest take: everything above is what LemonSlice says and shows in its own demos. Independent coverage from OSBBD on September 24, 2026 repeated the launch claims and noted it had not tested their consistency, and we haven't run a 24-hour stream either. The free web demo is the quickest way to judge it with your own photo before you plan anything around the API. Receipts: LemonSlice CWM-1 launch page, CWM-1 emotion engine docs, CWM-1 Lite announcement, OSBBD coverage.

The bottom line for AI avatar creators

CWM-1 moves real-time avatars from a moving face to a character with a body, hands and a scene that reacts while you talk to it. You don't need to train anything: one good image is the input. That makes the image itself the part worth getting right, with a clear face, visible hands, something to touch and a setting that can come alive.

If you don't have that image yet, start with your avatar.

FAQs about LemonSlice CWM-1

What is LemonSlice CWM-1?

Character World Model-1 is a real-time interactive avatar model released on September 23, 2026. It generates the character's face, body, hands and background live during a conversation, starting from one image.

Is CWM-1 free?

You can try it for free in the LemonSlice web app with custom characters. API access is available to LemonSlice Ultra and Enterprise customers.

What can the avatar do with its body?

The emotion engine includes actions such as smile, laugh, wave, nod, blow a kiss, heart hands, dance, lean forward, cross arms, touch hair, touch collar, text and take a phone call. It plays one action at a time.

Does it only work with photos of real people?

No. LemonSlice says the same model animates humans, anime characters and animals in any style, without character-specific training.

What is CWM-1 Lite?

A lower-cost version announced on October 2, 2026, with the same controllability and prices as low as $0.048 per minute, according to LemonSlice.