XGEN-JING: Egocentric Interactive Model on MiniMax H3

XGEN Labs published XGEN-JING, an egocentric interactive experience model built on MiniMax H3. Where a normal video model generates a fixed clip, JING takes actions, reference images and an observation history and generates first-person video and audio: you move, interact with objects and hold conversations.

The official XGEN-JING cover from the model repository.

What it does

The model is trained for three behaviours, all generated jointly as video plus audio:

• Camera control. Move through everyday places and imagined worlds with keyboard input, exploring the same starting point by walking it differently.
• Interaction and dialogue. Direct object interactions and character conversations with text instead of motion prompts.
• Reference conditioning. Compose an experience from character, object and scene images taken from a story, so the same set of references can be explored under different actions.

JING Flash is built on MiniMax H3's Ref2VA path plus FlashGen, the 4-step H3 LoRA, which is what makes the four-step bidirectional inference possible. The text encoder, tokenizer, processor, video and audio VAEs and schedulers all come from diffusers-format MiniMax H3 rather than being part of the release, so only the jing_flash_v1 transformer is new.

Release status

This is a preview-class release, and the release plan is explicit about it:

The distinction matters for what "interactive" means here. A four-step bidirectional model generates the whole clip with the action sequence known up front; the causal model, which would respond step by step as you press keys, is not out yet. XGEN Labs has published a gallery and a demo video alongside the weights.

Also included in the repository is a prompt skills document, which generates validated inference cases from stories and reference images.

Availability

The official XGEN-JING demo video.

• Weights: XGENlabs/XGEN-JING (jing_flash_v1, plus cover and demo assets)
• Code: XGEN-Labs/XGEN-JING
• Project page: xgenlabs.ai/research.html
• Gallery: xgenlabs.ai/gallery.html

There is no ComfyUI support: the inference path is diffusers plus an SGLang runtime pinned to a specific diffusion revision, and the validated demo configuration spreads the stack across six H100 GPUs (one for the text encoder, one for the video and audio VAEs, four for the DiT with sequence parallelism) with FlashAttention-4 as the default attention backend. Until a causal model and lighter inference path arrive, this is a look at where H3-based interactive models are heading rather than something runnable on a local ComfyUI box.

For the earlier interactive world model on the same base, see H3-World.