{"id":12613,"date":"2026-11-17T11:01:26","date_gmt":"2026-11-17T02:01:26","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=12613"},"modified":"2026-11-17T11:10:34","modified_gmt":"2026-11-17T02:10:34","slug":"worl-helios-rl-trained-lora-adapter","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/worl-helios-rl-trained-lora-adapter\/","title":{"rendered":"WORL-Helios Video Generation Model: File List"},"content":{"rendered":"<p><em>Sample outputs are available on the <a href=\"https:\/\/huggingface.co\/Efficient-Large-Model\/WORL-Helios\">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\/Efficient-Large-Model\/WORL-Helios\">Efficient-Large-Model\/WORL-Helios<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-efficient-large-model-en\/\">Efficient Large Model (NVIDIA and MIT): models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-20<\/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>Efficient-Large-Model has released &#8220;WORL-Helios&#8221;, a reinforcement learning-trained LoRA adapter targeting the video generation model &#8220;Helios-Distilled&#8221;. This model functions as an adapter that accepts text prompts and video history information as inputs to perform autoregressive continuous video generation (Text-to-Video).<\/p>\n<p>This repository is provided as a standalone PEFT-format LoRA adapter rather than a complete video generation pipeline. Therefore, actually generating videos requires a base Transformer model along with a text encoder, tokenizer, VAE, scheduler, and a corresponding Helios execution environment separately.<\/p>\n<h2>Specifications<\/h2>\n<p>Specifications listed in the public documentation are as follows:<\/p>\n<ul>\n<li>Adapter parameters: 419,430,400 (640 adapter tensors across 40 Transformer blocks)<\/li>\n<li>Base model: <code>BestWishYsh\/Helios-Distilled<\/code> (Transformer class: <code>HeliosTransformer3DModel<\/code>, root base model: <code>Wan-AI\/Wan2.1-T2V-14B-Diffusers<\/code>)<\/li>\n<li>Adapter precision: FP32<\/li>\n<li>LoRA configuration: rank 128 \/ alpha 128<\/li>\n<li>Attention targets: <code>to_q<\/code>, <code>to_k<\/code>, <code>to_v<\/code>, <code>to_out.0<\/code><\/li>\n<li>Training methodology and settings: WORL training using VideoAlign&#8217;s VQ (Visual Quality), MQ (Motion Quality), and TA (Text Alignment) rewards, interactive prompt mode, 16 FPS<\/li>\n<li>Output and frame specifications (common to base model): Autoregressive generation in units of 33 frames per chunk. Recommended frame counts are multiples of 33 (values that are not multiples are rounded up). Supports video output at 16 FPS and 24 FPS<\/li>\n<li>License terms: Apache-2.0 (an open-source license permitting commercial use, modification, and redistribution, subject to compliance with copyright notices and patent terms)<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>According to the model card by the publisher Efficient-Large-Model, no new generation benchmark scores or hardware compatibility matrices have been established for the release of this adapter, and no objective comparative evaluation figures have been published.<\/p>\n<p>As reference information, the inference speed and frame count estimates published in the model card for the base model &#8220;Helios-Distilled&#8221; are as follows:<\/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>num_frames<\/th>\n<th>Adjusted Frames<\/th>\n<th>24 FPS<\/th>\n<th>16 FPS<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>1449<\/td>\n<td>1452 (33\u00d744)<\/td>\n<td>~60s (1min)<\/td>\n<td>~90s (1min 30s)<\/td>\n<\/tr>\n<tr>\n<td>720<\/td>\n<td>726 (33\u00d722)<\/td>\n<td>~30s<\/td>\n<td>~45s<\/td>\n<\/tr>\n<tr>\n<td>240<\/td>\n<td>264 (33\u00d78)<\/td>\n<td>~11s<\/td>\n<td>~16s<\/td>\n<\/tr>\n<tr>\n<td>129<\/td>\n<td>132 (33\u00d74)<\/td>\n<td>~5.5s<\/td>\n<td>~8s<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to the base model&#8217;s published documentation, Helios-Distilled is a distillation model employing x0 prediction and a custom scheduler (<code>HeliosDMDScheduler<\/code>), and is reported to achieve an end-to-end inference speed of 19.5 FPS on a single H100 GPU. As the table above shows, it is characterized by high efficiency, capable of generating 1452 frames (approximately 1 minute at 24 FPS) in about 60 seconds. On the other hand, being a distillation model, a quality tradeoff may exist when compared to &#8220;Helios-Base&#8221;, which prioritizes maximum quality.<\/p>\n<p>Regarding quality and limitations, the model card explicitly notes limitations to be aware of. It states that during long-duration video generation, temporal artifacts, breakdowns in character or subject identity (identity drift), geometric inconsistencies, and discrepancies with the prompt tend to accumulate. It is also explained that users should note the generated video is synthetic data and not a physically validated simulation.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>WORL-Helios is a LoRA adapter developed primarily for research purposes in video generation and reinforcement learning (RL) within world models.<\/p>\n<p>This model aims to adapt the video Transformer of the base model &#8220;Helios-Distilled&#8221; to perform autoregressive video generation. It is suited for tasks that accept conditioning information, such as preceding video frames, in addition to text prompt instructions, to continuously generate the subsequent footage.<\/p>\n<p>Optimization employs the &#8220;WORL (World Model Reinforcement Learning)&#8221; method, utilizing three types of rewards based on VideoAlign (VQ for visual quality, MQ for motion quality, and TA for text alignment). Furthermore, since training was conducted assuming an interactive prompt mode and a 16 FPS setting, it is suited for video generation research emphasizing real-time performance and interactivity, as well as experiments with world models that model interactions with an environment.<\/p>\n<p>The base model Helios-Distilled features a high-efficiency video generation architecture, supporting Image-to-Video and Video-to-Video in addition to Text-to-Video. WORL-Helios is geared toward researchers and developers who want to leverage this high-efficiency inference pipeline while tuning the behavior of autoregressive generation through reinforcement learning. However, note that verification of generation results is required since long-duration generation may produce identity drift of subjects or geometric inconsistencies.<\/p>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong> \u2014 14.3B parameters (taken from the base model BestWishYsh\/Helios-Distilled)<\/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>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>More than 154GB of VRAM (multi-GPU or CPU offload required)<\/td>\n<td>F32<\/td>\n<td>128.3GB<\/td>\n<td>153.9GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. The listed files are distributed in F32 (32-bit). Loading in a lower precision (BF16 or quantized) needs less memory, so treat this as the size of the files as distributed, not a minimum. This site has not measured the lower-precision figures. This release is an adapter (LoRA etc.); the table shows what the base model <a href=\"https:\/\/huggingface.co\/BestWishYsh\/Helios-Distilled\">BestWishYsh\/Helios-Distilled<\/a> needs. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><!-- \/lmw:hardware --><\/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:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/Efficient-Large-Model\/WORL-Helios\/tree\/main\">Efficient-Large-Model\/WORL-Helios<\/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>adapter_model.safetensors<\/code><\/td>\n<td>1.68GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>The weight files for WORL-Helios are available in the Hugging Face repository <code>Efficient-Large-Model\/WORL-Helios<\/code>. Because it is not a gated model, downloads can be performed without prior application for use or license agreement procedures.<\/p>\n<p>The distribution format is PEFT-compatible safetensors, and the repository includes adapter weights (<code>adapter_model.safetensors<\/code>), configuration files (<code>adapter_config.json<\/code>), and a manifest recording checksums and source revisions (<code>RELEASE_MANIFEST.json<\/code>). According to the publisher&#8217;s explanation, the weights and adapter configuration themselves are identical in content to the preceding repository <code>qyoo\/helios-worl<\/code> (revision <code>c17176f56773<\/code>).<\/p>\n<p>To acquire the complete set of weights using the Hugging Face CLI, run the following command:<\/p>\n<pre><code class=\"language-bash\">hf download Efficient-Large-Model\/WORL-Helios --local-dir artifacts\/WORL-Helios\n<\/code><\/pre>\n<p>Because this repository provides a standalone LoRA adapter, running the video generation pipeline separately requires the execution environment of the official <a href=\"https:\/\/github.com\/PKU-YuanGroup\/Helios\">Helios<\/a> repository, the base model <code>BestWishYsh\/Helios-Distilled<\/code>, a text encoder, a tokenizer, a VAE, and a scheduler. In the original export configuration, <code>Wan-AI\/Wan2.1-T2V-14B-Diffusers<\/code> is specified for the VAE, and PEFT version 0.18.1 is recorded.<\/p>\n<p>After setting up the Helios and PEFT environments, an example implementation of a Python script that applies the adapter to the base Transformer model is as follows:<\/p>\n<pre><code class=\"language-python\">import torch\nfrom peft import PeftModel\nfrom helios.modules.transformer_helios import HeliosTransformer3DModel\n\nbase = HeliosTransformer3DModel.from_pretrained(\n    &quot;BestWishYsh\/Helios-Distilled&quot;,\n    subfolder=&quot;transformer&quot;,\n    torch_dtype=torch.bfloat16,\n    low_cpu_mem_usage=True,\n).eval()\n\nmodel = PeftModel.from_pretrained(\n    base,\n    &quot;Efficient-Large-Model\/WORL-Helios&quot;,\n    is_trainable=False,\n).eval()\n<\/code><\/pre>\n<p>When building the entire generation pipeline, it is recommended to pin the environment and run inference after confirming compatibility with a specific revision of the base model. Related releases such as <code>WORL-NVRM<\/code> and <code>WORL-Lingbot<\/code> are also pointed out.<\/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\/worl-lingbot-camera-conditioned-lora\/\">WORL-Lingbot Video Generation Model: File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/30\/longlive-plug-wan21-t2v-14b-few-step-2\/\">LongLive-Plug-Wan2.1-T2V-14B-few-step: Our Generated Video, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/30\/longlive-plug-wan2-1-t2v-14b-cfg\/\">LongLive-Plug-Wan2.1-T2V-14B-cfg: Our Generated Video, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/29\/longlive-plug-wan21-ti2v-5b-cfg-lora\/\">LongLive-Plug-Wan2.2-TI2V-5B-cfg: Our Generated Video, File List<\/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>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-efficient-large-model-en\/\">Efficient Large Model (NVIDIA and MIT): models, licenses and articles<\/a><\/li>\n<li><strong>Other models for the same task<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-by-task-en\/#task-video\">Other video generation models<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/Efficient-Large-Model\/WORL-Helios\">Efficient-Large-Model\/WORL-Helios<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/BestWishYsh\/Helios-Distilled\">BestWishYsh\/Helios-Distilled<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/PKU-YuanGroup\/Helios\">PKU-YuanGroup\/Helios (GitHub)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/qyoo\/helios-worl\/tree\/c17176f5677301f9e10abb8d0ff69733e60b5e6d\">qyoo\/helios-worl<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/Efficient-Large-Model\/WORL-NVRM\">Efficient-Large-Model\/WORL-NVRM<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/Efficient-Large-Model\/WORL-Lingbot\">Efficient-Large-Model\/WORL-Lingbot<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Explore WORL-Helios, a reinforcement learning-trained LoRA adapter for the Helios-Distilled video generation model by Efficient-Large-Model.<\/p>\n","protected":false},"author":1,"featured_media":12612,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[724,3773,111,3775,1547,3777,3779,959],"class_list":["post-12613","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-efficient-large-model-en","tag-helios-distilled-en","tag-lora-en","tag-peft-en","tag-verified","tag-wan2-1-t2v-diffusers-en","tag-worl-helios-en","tag--en"],"lang":"en","translations":{"en":12613,"ja":12611},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12613","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=12613"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12613\/revisions"}],"predecessor-version":[{"id":17865,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12613\/revisions\/17865"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/12612"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=12613"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=12613"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=12613"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}