{"id":13326,"date":"2026-11-17T11:02:16","date_gmt":"2026-11-17T02:02:16","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=13326"},"modified":"2026-11-17T11:11:28","modified_gmt":"2026-11-17T02:11:28","slug":"hy-nomas-diffusers-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/hy-nomas-diffusers-released\/","title":{"rendered":"Hy-Nomas-Diffusers Interactive Video Generation Model: File List"},"content":{"rendered":"<p><em>Sample outputs are available on the <a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas-Diffusers\">model card<\/a>.<\/em><\/p>\n<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas-Diffusers\">tencent\/Hy-Nomas-Diffusers<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-tencent-hy-nomas-en\/\">Hy-Nomas guide (2 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-tencent-en\/\">Tencent: models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-24<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>A Diffusers version of &#8220;Hy-Nomas,&#8221; a free-form interactive video generation model developed by Tencent, has been released. It functions as a &#8220;world model&#8221; that sequentially generates video segments by combining text instructions and camera controls, allowing users to add objects, change styles, switch scenes, describe actions, and manipulate camera work while continuing generation.<\/p>\n<p>This model is built on a segment-autoregressive Diffusion Transformer (DiT) combined with a Qwen text encoder, a camera conditioning module, and a PixelUMM video decoder. It features a persistent KV cache (Key\/Value Cache) that maintains context between segments, making it possible to expand the world while preserving consistency. The newly released version is a distilled model converted to BF16 precision Diffusers format, supporting fast inference with a small number of steps.<\/p>\n<h2>Specifications<\/h2>\n<p>The main specifications of this model based on the documentation are as follows:<\/p>\n<ul>\n<li><strong>Architecture<\/strong>: Segment-autoregressive Diffusion Transformer (DiT)<\/li>\n<li><strong>Transformer Blocks<\/strong>: 54 double-stream MMDiT blocks<\/li>\n<li><strong>Latent Space<\/strong>: DCAE-V <code>f32t4c32<\/code> (32 channels, spatial 32x and temporal 4x compression)<\/li>\n<li><strong>Text Encoder<\/strong>: Qwen2.5-VL-7B-Instruct (using language model components)<\/li>\n<li><strong>Video Decoder<\/strong>: Fine-tuned PixelUMM (Transformer-based)<\/li>\n<li><strong>Decoder Specifications<\/strong>: 36 layers, hidden size 4,096, 32 query heads, 8 KV heads<\/li>\n<li><strong>Output Resolution<\/strong>: 704 \u00d7 1280 (default setting)<\/li>\n<li><strong>Framerate<\/strong>: 24 FPS<\/li>\n<li><strong>Recommended Inference Steps<\/strong>: 4 steps for the distilled model (8 steps for the standard setting)<\/li>\n<li><strong>Sampler<\/strong>: Flow Match Euler Discrete Scheduler (Euler ODE sampling)<\/li>\n<li><strong>License<\/strong>: Apache-2.0<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>To evaluate the quality and controllability of this model, the development team has released a custom benchmark set called &#8220;Hy-Nomas-Bench.&#8221; Comprising 280 test cases, it measures four key capabilities: multi-turn free-form text dialogue, camera navigation, combined text-camera control, and spatial memory. This evaluates both the interactive controllability during continuous world generation and the consistency of the generated video.<\/p>\n<p>In terms of technical quality, the following features are noted in the model card:<\/p>\n<h3>Advanced Conditioning Mechanism<\/h3>\n<p>For text conditioning, it utilizes the 3,584-dimensional hidden states of Qwen2.5-VL-7B-Instruct. By projecting these into the 2,048-dimensional hidden space of the DiT via a 2-layer token refiner, it is reported to accurately reflect complex instructions that change from segment to segment. For camera control, it adopts &#8220;extrinsics_attn,&#8221; which processes 4&#215;4 camera extrinsic matrices aligned with the temporal compression of video latents, achieving precise viewpoint manipulation through cross-attention.<\/p>\n<h3>Autoregressive Consistency Maintenance<\/h3>\n<p>The generation process is performed autoregressively per video segment. Within each block, &#8220;segment-wise causal joint attention&#8221; functions, inheriting context from past segments through 3D RoPE positional encoding and the KV cache. Because this attention maintains causality across segments, it is capable of generating the current segment without needing future information, enabling efficient continuous generation.<\/p>\n<h3>High-Definition Video Decoding<\/h3>\n<p>Generated latent variables are restored to pixel space by a dedicated, fine-tuned PixelUMM decoder. This decoder employs a temporal window method that processes 10 latent frames with a 5-frame overlap, reconstructing high-quality RGB video in a single decoding step. This is said to yield efficient and visually superior video output independently of the DiT denoising process.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Hy-Nomas specializes in generating &#8220;video worlds&#8221; that dynamically change in response to user inputs. While traditional video generation models focus primarily on creating standalone clips, this model can continuously expand a consistent world through multi-turn dialogues.<\/p>\n<p>Specific strengths and anticipated use cases include:<\/p>\n<ul>\n<li><strong>Free-Form Interactive Generation<\/strong>: Users can introduce new objects into the generating world via text instructions, or switch the overall style and scenes midway. For example, it allows for step-by-step direction, such as starting with a forest landscape, introducing a fox upon instruction, and then making surrounding plants glow.<\/li>\n<li><strong>Precise Camera Work Control<\/strong>: Conditioning using 4&#215;4 camera extrinsic matrices enables intended camera controls, such as third-person perspectives or viewpoint movement along specific trajectories. This is suitable for cutscene prototyping in game development or creating navigation footage for virtual spaces.<\/li>\n<li><strong>Long and Consistent Video Production<\/strong>: Because generation is performed autoregressively per segment, utilizing the KV cache makes it possible to append video while maintaining the context of previous segments (such as object placement and appearance). The benchmark &#8220;Hy-Nomas-Bench&#8221; also evaluates spatial memory capabilities, and the model is designed to maintain world consistency even from relocated viewpoints.<\/li>\n<li><strong>Applications to Robotics and Simulation<\/strong>: Documentation for the original model (tencent\/Hy-Nomas) indicates the presence of LoRA adapters trained on robot arm datasets, keeping visualization of simulation environments involving physical behaviors in view as a potential use case.<\/li>\n<li><strong>Expression of Diverse Styles<\/strong>: It is capable of generating a wide range of visual styles according to instructions, from realistic cinematic footage to fantasy-style animation.<\/li>\n<\/ul>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas-Diffusers\/tree\/main\">tencent\/Hy-Nomas-Diffusers<\/a>, listed by this site from the Hugging Face API. Sizes are the actual file sizes.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>File<\/th>\n<th>Size<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>decoder\/diffusion_pytorch_model-*-of-00003.safetensors<\/code><\/td>\n<td>13.94GB (3 split files)<\/td>\n<\/tr>\n<tr>\n<td><code>text_encoder\/model-*-of-00003.safetensors<\/code><\/td>\n<td>14.14GB (3 split files)<\/td>\n<\/tr>\n<tr>\n<td><code>transformer\/diffusion_pytorch_model-*-of-00004.safetensors<\/code><\/td>\n<td>17.11GB (4 split files)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>The newly released &#8220;Hy-Nomas-Diffusers&#8221; is a distilled BF16 precision model available from the Hugging Face repository. It is distributed in Safetensors format and can be used via the Diffusers library.<\/p>\n<p>The steps for acquisition and usage are as follows:<\/p>\n<ul>\n<li><strong>Library Installation<\/strong>: To use this model, you need to install Diffusers from a specific branch supporting Hy-Nomas. Use the following command:<br \/>\n  <code>bash pip install -e \"diffusers[hy-nomas] @ git+https:\/\/github.com\/jiaxiangc\/diffusers.git@hy-nomas\"<\/code><\/li>\n<li><strong>Loading the Model<\/strong>: You can load components such as the transformer, decoder, and text encoder all at once using <code>HyNomasModularPipeline<\/code>. This pipeline accepts either plain text instructions or structured prompts.<\/li>\n<li><strong>Repository Structure<\/strong>: The <code>tencent\/Hy-Nomas-Diffusers<\/code> repository contains the transformer (<code>transformer\/<\/code>), video decoder (<code>decoder\/<\/code>), text encoder (<code>text_encoder\/<\/code>), tokenizer (<code>tokenizer\/<\/code>), and scheduler (<code>scheduler\/<\/code>).<\/li>\n<li><strong>Additional Components<\/strong>: While the original model repository (<a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas\">tencent\/Hy-Nomas<\/a>) also distributes INT8 quantized DiT versions, FP8 decoders, and LoRA adapters that enhance specific visual qualities, they are not included in this Diffusers repository and must be used selectively as needed.<\/li>\n<li><strong>Inference Optimization<\/strong>: To improve inference speed and memory efficiency, the use of kernels such as FlashAttention 2\/3 or MagiAttention is recommended. These can be installed separately and enabled on the pipeline side.<\/li>\n<\/ul>\n<p>Note that obtaining this model (<code>tencent\/Hy-Nomas-Diffusers<\/code>) does not require any special license agreement (gated setting) on Hugging Face.<\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as safetensors.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-releases-hy-nomas\/\">Hy-Nomas Video Generation Model: File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-qwen3-8-27b-dspine-b16\/\">Qwen3.8-27B-DSpine-b16 Draft Model for Speculative Decoding: 8GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-releases-wevisone-vl-vision-language-models\/\">Tencent Releases WeVisOne-VL Vision Language Models<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-releases-dspine-parallel-draft-model\/\">Qwen3-8B-DSpine-b16: 4GB+ VRAM<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Explore the same model family<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-tencent-hy-nomas-en\/\">Hy-Nomas family overview (2 articles)<\/a><\/li>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">Safetensors format guide and models<\/a><\/li>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-tencent-en\/\">Tencent: models, licenses and articles<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas-Diffusers\">https:\/\/huggingface.co\/tencent\/Hy-Nomas-Diffusers<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas\">https:\/\/huggingface.co\/tencent\/Hy-Nomas<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/datasets\/tencent\/Hy-Nomas-Bench\">https:\/\/huggingface.co\/datasets\/tencent\/Hy-Nomas-Bench<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Discover Hy-Nomas-Diffusers, Tencent&#8217;s open interactive video generation model acting as a world model with segment-wise autoregressive DiT.<\/p>\n","protected":false},"author":1,"featured_media":13325,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[747,1833,4060,4096,1172,1547,959],"class_list":["post-13326","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-diffusers-en","tag-dit-en","tag-hy-nomas-en","tag-hy-nomas-diffusers-en","tag-tencent-en","tag-verified","tag--en"],"lang":"en","translations":{"en":13326,"ja":13324},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13326","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=13326"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13326\/revisions"}],"predecessor-version":[{"id":17934,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13326\/revisions\/17934"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/13325"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=13326"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=13326"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=13326"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}