Hy-Nomas for ComfyUI

tencent/Hy-Nomas (Tencent Hunyuan's interactive world model) converted for ComfyUI-Nomas, a native ComfyUI port. The stock Load Diffusion Model, Load CLIP and Load VAE nodes load these files.

File	Folder	Size	Notes
nomas_dit_distilled_int8_convrot.safetensors	models/diffusion_models	9.0 GB	the distilled DiT, int8 + ConvRot (0.86% mean weight error)
nomas_dit_distilled_bf16.safetensors	models/diffusion_models	17.1 GB	the distilled DiT in bf16 (base for LoRA merges)
nomas_pixelumm_vae_bf16.safetensors	models/vae	15.8 GB	the fine-tuned PixelUMM decoder plus the DCAE-V encoder, bf16

Text encoder: ComfyUI's own qwen_2.5_vl_7b.safetensors (Load CLIP, type qwen_image), from Comfy-Org/Qwen-Image_ComfyUI.

All files keep the released tensor names. The DiT files come from DiT/distilled/diffusion_pytorch_model_ema_0_995.safetensors (fp32 → bf16; int8 ConvRot on the double-block linears, with the camera branch, text refiner, embeddings and final layer kept in bf16). The VAE file comes from VAE/PixelUMM_FT/diffusion_pytorch_model.safetensors (fp32 → bf16).

Use

Generation runs at 1280×704 (the release's "720p"), segment by segment. Each segment attends to a KV cache of the earlier ones and can be steered by text, by a WASD/mouse camera move, or by both. On an RTX 5090 with the int8 DiT, a 3 s turn takes about 21 s, plus about 4 s per window to decode.

Licence

• The DiT files: Apache-2.0, as the Hy-Nomas release.
• nomas_pixelumm_vae_bf16.safetensors: the decoder is Tencent's fine-tune of NVIDIA PixelUMM, whose weights and derivatives are under the NVIDIA One-Way Noncommercial License (non-commercial research and evaluation only). The DCAE-V encoder in the same file comes from the Hy-Nomas release.