AlayaLab/HelloWorld

TL;DR — HelloWorld adds a "social interaction" channel to camera-controllable video world models: press F and the on-screen character turns to the viewer, waves, nods, or mouths a greeting, while camera trajectory and scene reconstruction stay intact. The interesting part is how: the model is finetuned entirely on data it (well, an off-the-shelf T2V model) synthesized itself, and the timing of the interaction is controlled at inference with a training-free temporal attention mask — no interaction-timing labels, no mocap, no real dataset at all. Code, training recipe, and LoRA weights are out; the benchmark isn't yet.

Why this is worth a look

Video world models (Genie-style, camera-conditioned diffusion, etc.) have converged on two control axes: camera pose and scene content. Characters in these worlds are scenery — they don't notice you. HelloWorld proposes a third axis: viewer-directed social response, treated as a first-class conditioning signal alongside camera trajectory. That's a small but genuinely new framing, and it's the kind of primitive you'd need for anything resembling an inhabited interactive world (games, embodied-agent sims, telepresence). It also generalizes beyond humans in the demos — crows, mascots, skeleton props, anime figurines all respond — suggesting the model learned "attend to viewer" as an abstract behavior rather than a human-animation prior.

How it works

Two ideas, one at training time and one at inference:

Self-distillation on synthetic data. There's no dataset of "person reacts to camera at time t" videos, so they make one: LTX-2.3 text-to-video generates clips containing both social interactions and camera motion, and the base model is LoRA-finetuned on this synthetic corpus. The claimed benefit is that the model learns camera-pose conditioning (via warped-video conditions, judging by the example layout) without the finetuning washing out its interaction abilities. The whole training pipeline is prompt-driven — train/prompts.tsv plus a config — so retargeting to your own characters or scenes is an edit-and-rerun affair with no data collection.

Training-free temporal control. The F-press has to localize the response in time. Rather than train a timing-conditioned model, they apply a temporal cross-attention mask at inference: the interaction prompt tokens only attend to (or are attended by) the frames inside the press window. Before and after, the character behaves as background. This is a clean trick — analogous in spirit to spatial attention masking for regional prompting in image diffusion, transposed to time — and it means any timing behavior is adjustable without retraining.

What's actually released

• Inference code (inference/, pinned env), with a one-line script and batch reproduction of the bundled examples.
• Training code (train/) that regenerates the synthetic data and reproduces helloworld_lora_v1 end to end — notable, since most video-model releases omit the data pipeline entirely.
• Weights: LoRA at oyly/HelloWorld_V1 on Hugging Face.
• Benchmark: HelloWorldBench (400 samples, with ActAcc / TimeAcc / GazeDev for action correctness, timing, and gaze direction) is not yet released — the last unchecked box on the schedule.

The evidence in the README is qualitative: seven warp-condition vs. generation pairs and a 1080p demo video. No quantitative numbers appear in the README — no benchmark scores, no baselines, no FVD. The paper (arXiv:2608.05070) presumably has them, but you can't verify claims from the repo alone until the benchmark ships. One credibility point in its favor: ScrappyLabs independently reproduced the method and released their own LoRA, and the authors say results align with the paper — rare for a two-week-old repo.

Caveats

• The base video model isn't named in the README (the LoRA has to attach to something — check inference/README before assuming your GPU can run it). Hardware requirements and license are also unstated.
• It's a LoRA on a diffusion video model, so this is offline clip generation with a scheduled F-press, not a real-time interactive world — the "button press" is a conditioning signal, not a live input loop.
• Interaction vocabulary looks shallow: turn, wave, nod, short greeting. Multi-turn or contingent social behavior is out of scope.
• 48 stars, benchmark pending, and self-reported alignment with an unofficial reproduction — promising but early.

Try it

git clone https://github.com/AlayaLab/HelloWorld
cd HelloWorld/inference
# env setup: pinned requirements in env/, weights from HF oyly/HelloWorld_V1
bash run_helloworld.sh                       # single clip
INPUT=examples_batch.json bash run_batch.sh  # reproduce bundled examples

Demo video: https://youtu.be/j4scl5Y7gXo · Paper: https://arxiv.org/abs/2608.05070