Tencent releases interactive video AI with a single-RTX-4090 configuration

Tencent has released Hy-Nomas, a video-generating model that lets users change instructions and camera movement as a scene unfolds. The Apache 2.0 release includes downloadable weights, code and a web demo. Tencent also provides a configuration designed to run video generation on one NVIDIA GeForce RTX 4090, with specific memory-saving measures.

A scene that accepts another instruction

Hy-Nomas builds its output one video segment at a time. Users can introduce objects, describe actions, change a scene’s appearance or style, and supply camera trajectories for viewpoint and movement. Successive instructions guide later segments, rather than serving only as the initial description for a finished clip.

The model combines current instructions with stored context from earlier generation. That context lives in a key/value cache, a record of information the model can reuse when producing the next segment. Its standard configuration keeps the full history. Each new segment can draw on past segments without needing any future footage to exist first.

Underneath is a diffusion transformer, which generates each segment by progressively removing noise from a compressed video representation. A Qwen2.5-VL-7B-Instruct language-model component encodes the text, while separate camera-conditioning modules supply movement information. A fine-tuned PixelUMM decoder then turns the compressed output into video frames.

Tencent’s distilled inference configuration uses eight denoising steps and specifies 720 × 1280 output at 24 frames per second. Those figures describe the generation settings and the resulting video’s frame rate; they should not be read as a measured real-time generation speed on the RTX 4090.

What changes to fit one graphics card

The RTX 4090 path uses lower-precision storage: INT8 model weights and cached history, plus an FP8 video decoder. It also offloads work to the CPU. The main model’s linear computations remain in BF16, so an INT8 checkpoint does not mean every calculation uses eight-bit arithmetic.

Memory handling changes too. Instead of exposing the full history to GPU attention, the single-GPU configuration selects a bounded set of historical frames from the compressed representation and keeps an archive on the CPU. Its default budget is 18 historical latent frames. That is a context budget, not a limit of 18 output frames or the length of the current segment.

Developers must also match the downloaded assets to their deployment path. Tencent’s model card distinguishes three practical choices:

• Standard offline generation requires the distilled model, a matching PixelUMM decoder and the Qwen text encoder.
• Single-GPU generation requires the compatible INT8 model and FP8 decoder pair.
• Optional cinematic and robotics adapters are available, but the INT8 loader cannot load a separate LoRA adapter during inference.

Setup instructions and a test suite

Tencent recommends its pre-packaged Docker image, which bundles Python, PyTorch and the required CUDA attention kernels. The documented host setup requires Linux, a compatible NVIDIA driver, Docker and NVIDIA Container Toolkit. Offline generation can read prepared prompt files directly; the web demo instead uses Nomas-PE to turn user instructions into structured generation prompts.

Alongside the model, Tencent is releasing Hy-Nomas-Bench: four benchmarks containing 280 test cases. They cover multi-turn text interaction, camera navigation, interleaved text and camera control, and spatial memory. The suite supplies structured prompts, interaction sequences, keyboard timelines and reference camera trajectories to evaluate control and consistency as generation continues.