Tencent publishes Hy-Nomas in Diffusers format: 4-step distilled video world model at 704 × 1280, 24 FPS

Tencent has published the distilled Hy-Nomas model in BF16 Diffusers format on Hugging Face, loadable through HyNomasModularPipeline. Hy-Nomas is a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control. The example configuration uses 4 denoising steps per segment and outputs 704 × 1280 video at 24 FPS.

How it works

Hy-Nomas generates a world one video segment at a time. Between segments, users can introduce objects, change styles and scenes, describe actions, or control the camera. It combines a segment-autoregressive diffusion transformer (DiT), a Qwen text encoder, camera-conditioning modules and a PixelUMM video decoder. A persistent key/value (KV) cache carries context between segments.

Diffusion runs in a compressed video latent space, and autoregression runs across segments. Attention is causal across segments, so future segments are not needed to generate the current one.

Architecture details

• Backbone: dense multimodal diffusion transformer (MMDiT) with 54 dual-stream blocks and 0 single-stream blocks.
• Dimensions: hidden size 2,048; 16 attention heads with 128 dimensions per head; feed-forward expansion 4× (intermediate size 8,192); 32 latent input and output channels.
• Position encoding: 3D RoPE with temporal/height/width dimensions [16, 56, 56] and theta 256.
• Prediction and sampling: flow velocity, with Euler ODE sampling.
• Text encoder: the language-model component of Qwen/Qwen2.5-VL-7B-Instruct. Its 3,584-dimensional hidden states pass through a two-layer token refiner and are projected into the DiT's 2,048-dimensional space.
• Camera conditioning: normalized 4 × 4 camera extrinsic matrices are grouped over four video frames and flattened into 64-dimensional inputs. Video tokens query the projected camera features through cross-attention in each double-stream block.
• Latents: DCAE-V f32t4c32, meaning 32 channels, 32× spatial compression in each dimension and 4× temporal compression.
• Decoder: a fine-tuned PixelUMM decoder, a transformer with 36 layers, a 4,096-dimensional hidden size, 32 query heads and 8 KV heads. It processes RGB video in 4 × 16 × 16 tubelets, using one decoding step with temporal windows of 10 latent frames and an overlap of 5 latent frames.

Hy-Nomas-Bench

Tencent also releases Hy-Nomas-Bench, a suite of four benchmarks with 280 test cases. They cover multi-turn free-form text interaction, camera navigation, interleaved text and camera control, and spatial memory. The suite evaluates interactive controllability and consistency during continuous world generation, and provides structured prompts, interaction sequences, keyboard timelines and reference camera trajectories. Scores for Hy-Nomas on the benchmark are not given on this page.

What is in the Diffusers directory

The repository holds the BF16 Diffusers conversion of the distilled model, serialized with HyNomasModularPipeline.save_pretrained(). It includes the transformer (HyNomasTransformer3DModel), the PixelUMM decoder, the Qwen2_5_VLTextModel text encoder, the Qwen tokenizer and a FlowMatchEulerDiscreteScheduler configuration.

The pipeline requires a custom Diffusers branch, installed with pip install -e "diffusers[hy-nomas] @ git+https://github.com/jiaxiangc/diffusers.git@hy-nomas". Native checkpoints, quantized variants, adapters and Nomas-PE assets live in the original model repository, not in this directory.

Nomas-PE is the separate prompt-enhancement service used by the original web demo to turn user instructions into structured generation prompts. The Diffusers pipeline does not run it. It accepts plain text or prepared structured prompts directly.

Context window and pricing do not apply to this weights release, and no API pricing is listed.

What this means

The Diffusers release puts the distilled Hy-Nomas model into a standard open-source inference workflow. The pipeline retains the attention KV cache across segments, and continued generation can reuse the state returned by an earlier call. Developers can therefore extend a single world over successive turns rather than regenerating from scratch.

The main differences from the web demo are practical. Users of this directory must supply their own structured prompts, since Nomas-PE is not included, and must use the custom Diffusers branch. Anyone needing native checkpoints, quantized variants or adapters has to go to the original repository.

Hy-Nomas-Bench gives researchers a shared test set for interactive controllability, camera navigation and spatial memory, which are the properties a persistent, multi-turn world generator has to maintain.