HunyuanVideo 1.5 Technical Report

TL;DR

HunyuanVideo 1.5 is Tencent’s open-source 8.3B-parameter DiT video generation model targeting consumer-grade GPUs (14GB VRAM, RTX-class). The headline architectural contribution is SSTA (Selective and Sliding Tile Attention) — a sparse attention mechanism that dynamically prunes redundant spatiotemporal KV blocks, claimed to give a 1.87× end-to-end speedup over FlashAttention-3 for 10-second 720p synthesis. The release includes T2V and I2V variants at 480p and 720p, CFG-distilled and step-distilled checkpoints (8/12-step I2V), an explicit 480p→720p and 720p→1080p super-resolution network, and the training code (Muon optimizer, FSDP + sequence parallel). Bilingual (Chinese-English) prompt understanding is supported via glyph-aware ByT5 text encoding alongside a large multimodal model.

Key claims

The DiT backbone is 8.3B parameters paired with a 3D causal VAE that compresses 16× spatially and 4× temporally, yielding a model deployable on consumer GPUs with as little as 14 GB VRAM with offloading enabled [§System Requirements; GitHub README].
SSTA — Selective and Sliding Tile Attention — combines a static local-window prior with dynamic global block selection, pruning redundant spatiotemporal KV blocks for 1.87× end-to-end speedup vs FlashAttention-3 on 10-second 720p synthesis [§Key Features; abstract; Fig. 1].
The model is trained with the Muon optimizer (momentum + Newton-Schulz orthogonalization), and the released training code (train.py) is built on Muon end-to-end — including for LoRA fine-tuning [§Training; GitHub README].
Bilingual (Chinese-English) prompt understanding is handled by combining a large multimodal model for semantic comprehension with a dedicated ByT5 glyph encoder for in-video text rendering accuracy [§Enhanced Multimodal Understanding; abstract].
A separate few-step super-resolution network is released for 480→720 (6 steps) and 720→1080 (8 steps) upscaling, framed as a distinct stage rather than an end-to-end higher-resolution generator [§Optimal Inference Configurations; §Video Super-Resolution Enhancement].
A unified DiT architecture is jointly trained for text-to-image (T2I), text-to-video (T2V), and image-to-video (I2V) tasks under a single backbone [§Method; multi-task training paragraph].
The 480p I2V step-distilled checkpoint generates videos in 8–12 steps and runs end-to-end on a single RTX 4090 in ≈75 seconds (≈75% wall-clock reduction vs the non-distilled model) [GitHub News, Dec 5 2025].
Release covers 13 model variants (480p/720p × T2V/I2V × {base, CFG-distill, step-distill, sparse-CFG-distill}) plus 720p-sr and 1080p-sr super-resolution checkpoints; sparse-attention variants (720P-T2V-sparse-cfg-distill) require the flex-block-attn backend and only the 720p models are equipped with sparse attention [§Model Cards; §Optimal Inference Configurations].

Method

The pipeline is a two-stage framework: an 8.3B DiT trained as a multi-task (T2I + T2V + I2V) generator over a 16×-spatial / 4×-temporal 3D causal VAE latent, followed by a separate few-step super-resolution network for 1080p upscaling. The DiT itself replaces full spatiotemporal attention with SSTA, which keeps a static local sliding tile (locality prior) while dynamically selecting which non-local tiles to attend to per query — a hybrid of NSA-style block sparsity and sliding-window attention. Text conditioning is dual-branch: a large multimodal model produces semantic-level bilingual embeddings, and a ByT5 byte-level encoder produces glyph-aware embeddings used to drive in-video text rendering. Training uses the Muon optimizer with FSDP + sequence parallel + gradient checkpointing; LoRA fine-tuning is explicitly supported and the README requires Muon for any continued training. Inference at 480p I2V is further accelerated via a step-distilled student (8–12 steps), CFG distillation (≈2× speedup), SageAttention, and feature caching (DeepCache / TeaCache / TaylorCache).

Results

1.87× end-to-end inference speedup on 10-second 720p video generation vs FlashAttention-3, attributed to SSTA [§Key Features].
Step-distilled 480p I2V: ≈75% end-to-end wall-clock reduction on RTX 4090, generating a clip in ~75 seconds [GitHub News].
Memory footprint: 14 GB VRAM minimum with offloading; the Wan2GP community port runs at 6 GB VRAM [§System Requirements; §Community Contributions].
1080p output achieved via the dedicated SR network (8-step 720→1080) rather than direct 1080p DiT inference [§Optimal Inference Configurations].

The technical report PDF is paywalled behind an arxiv 429 at filing time; quantitative VBench / per-attribute numbers from the paper are not enumerated above. The numbers in this page come from the README, abstract, and arxiv abstract listing — claims, not directly verified in the PDF.

Why it’s interesting

HunyuanVideo 1.5 is the first major open video-foundation release to ship structured sparse attention as a production component, not a research probe. It’s the wiki’s third concrete instance — alongside Radial Attention: O(n log n) Sparse Attention with Energy Decay for Long Video Generation (energy-decay O(n log n)) and VideoNSA: Native Sparse Attention Scales Video Understanding (NSA-style block sparsity for video understanding) — of structured video sparsity moving from paper to inference path. It’s also the second open video model after Lucy Edit 1.1 Dev: Open-Weight Instruction-Guided Video Editing on Wan2.2 5B and the FastWan / Wan2.2 series (FastWan: Generating a 5-Second Video in 5 Seconds via Sparse Distillation) to treat distillation + sparse attention + a dedicated SR stage as one release rather than separate research artifacts. Two things to track: (a) whether SSTA’s static-tile-plus-dynamic-selection combination beats radial attention and NSA-for-video at matched compute, and (b) whether the explicit 480p/720p → 1080p SR stage becomes the standard factorization for open video releases vs end-to-end 1080p DiTs.