BindWeave: Subject-Consistent Video Generation
BindWeave is a unified subject-consistent video generation framework by ByteDance Research. Built on MLLM-DiT architecture for single and multi-subject prompts (ICLR 2026).
BindWeave
Wan2.1Subject-ConsistentMLLM-DiTICLR 2026Unified subject-consistent video generation framework by ByteDance Research. MLLM-DiT architecture handles single and multi-subject prompts with precise entity grounding and cross-modal integration. Accepted at ICLR 2026.
| Developer | ByteDance Research |
| Release Date | 2025-10 |
| Architecture | MLLM-DiT (Multimodal LLM + Diffusion Transformer) |
| Parameters | 14B fp16 |
| Capabilities | Single and multi-subject consistent video generation |
| Venue | ICLR 2026 |
Overview
BindWeave is a unified subject-consistent video generation framework developed by ByteDance Research. It addresses the challenge of generating videos that maintain consistent subjects across frames, especially when handling multiple distinct subjects in a single scene.
The framework couples a pretrained multimodal large language model (MLLM) with a diffusion transformer (DiT) architecture. The MLLM provides precise entity grounding and cross-modal understanding, while the DiT handles high-quality video generation. This enables BindWeave to accurately render subjects described in complex prompts with their specified visual attributes.
ComfyUI Integration
BindWeave 14B fp16 and fp8 scaled variants are available in ComfyUI via the Kijai WanVideoWrapper custom nodes. The model weights are hosted on Hugging Face under both the official ByteDance Research repository and the Kijai WanVideo_comfy collection.
Resources
| Resource | Link |
|---|---|
| Paper | arXiv:2510.00438 |
| GitHub | github.com/bytedance/BindWeave |
| HuggingFace (Original) | BytedanceResearch/BindWeave |
| HuggingFace (Kijai) | Kijai/WanVideo_comfy |
Guides and workflows related to this model series.