Video Gallery

A cross-attention DiT conditioned by a UMT5-XXL text encoder and sampled with a Rectified-Flow Euler scheduler. The I2V variant adds a separate CLIP-image cross-attention branch per block.

A cross-attention DiT conditioned by a UMT5-XXL text encoder and sampled with a Rectified-Flow Euler scheduler. The I2V variant adds a separate CLIP-image cross-attention branch per block.

A cross-attention DiT conditioned by a UMT5-XXL text encoder and sampled with a Rectified-Flow Euler scheduler. The I2V variant adds a separate CLIP-image cross-attention branch per block.

Wan's DiT rebuilt as a two-expert Mixture-of-Experts routed by timestep, with per-token (not per-sample) modulation. TI2V-5B is instead one dense model, sharing a single checkpoint across T2V and I2V.

Wan's DiT rebuilt as a two-expert Mixture-of-Experts routed by timestep, with per-token (not per-sample) modulation. TI2V-5B is instead one dense model, sharing a single checkpoint across T2V and I2V.

A DiT with AdaLN-LoRA modulation, conditioned by a 7B Qwen2.5-VL ("Reason1") text tower with every decoder layer concatenated, and sampled with a flow-matching UniPC multistep solver.

A DiT with AdaLN-LoRA modulation, conditioned by a 7B Qwen2.5-VL ("Reason1") text tower with every decoder layer concatenated, and sampled with a flow-matching UniPC multistep solver.

An omnimodal Mixture-of-Transformers — a full weight set per modality — with dual causal / full-attention pathways, interleaved 3D mRoPE, no AdaLN, and a Karras-sigma UniPC sampler.

An omnimodal Mixture-of-Transformers — a full weight set per modality — with dual causal / full-attention pathways, interleaved 3D mRoPE, no AdaLN, and a Karras-sigma UniPC sampler.

A from-scratch LTX port: a pixel-unshuffle-patchify VAE with a noise-conditioned decoder, a plain T5-XXL text encoder, and a Rectified-Flow Euler sampler on a linear-quadratic sigma schedule.

A from-scratch LTX port: a pixel-unshuffle-patchify VAE with a noise-conditioned decoder, a plain T5-XXL text encoder, and a Rectified-Flow Euler sampler on a linear-quadratic sigma schedule.

A from-scratch LTX port: a pixel-unshuffle-patchify VAE with a noise-conditioned decoder, a plain T5-XXL text encoder, and a Rectified-Flow Euler sampler on a linear-quadratic sigma schedule.

A 22B LTX DiT with cross-attention AdaLN and per-head gated attention, an 8-layer embeddings connector, a Gemma-4 12B text encoder, and an ancestral (SDE) Euler sampler.

A 22B LTX DiT with cross-attention AdaLN and per-head gated attention, an 8-layer embeddings connector, a Gemma-4 12B text encoder, and an ancestral (SDE) Euler sampler.

A 22B LTX DiT (cross-attention AdaLN, per-head gated attention, Gemma-4 text encoder), here decoded with the slower transformer / neighborhood-attention diffusion VAE rather than the conv decoder.

A 22B LTX DiT (cross-attention AdaLN, per-head gated attention, Gemma-4 text encoder), here decoded with the slower transformer / neighborhood-attention diffusion VAE rather than the conv decoder.

8.3B MMDiT that conditions by concatenating Qwen2.5-VL and byT5 text tokens directly onto the video sequence — no cross-attention. SigLIP adds the reference frame for I2V. Rectified-Flow Euler.

8.3B MMDiT that conditions by concatenating Qwen2.5-VL and byT5 text tokens directly onto the video sequence — no cross-attention. SigLIP adds the reference frame for I2V. Rectified-Flow Euler.

13B dual-stream + single-stream MMDiT with joint full-sequence self-attention, conditioned by a Llama-3-8B tower plus a pooled CLIP-L vector folded into AdaLN. Flow-matching Euler; token_replace for I2V.

A DiT running one joint self-attention over [text; video] with no cross-attention, per-sample AdaLN, 3D RoPE (fixed 3D sincos on 2B), a causal 3D-conv VAE, and a T5-v1.1-XXL text encoder. v-prediction with zero-terminal-SNR.

A DiT running one joint self-attention over [text; video] with no cross-attention, per-sample AdaLN, 3D RoPE (fixed 3D sincos on 2B), a causal 3D-conv VAE, and a T5-v1.1-XXL text encoder. v-prediction with zero-terminal-SNR.

CogVideoX with temporal patchification (patch_size_t=2), a "slice"-grid 3D RoPE, and a native 1360×768 / 81-frame recipe run under DeepSpeed-Ulysses sequence parallelism.