Echo-Team/Echo-WM-Flash-Diffusers

Echo-WM Flash — Diffusers Format

This repository provides the Diffusers-format version of Echo-WM Flash Preview. The weights are converted from the official Echo-Team/Echo-WM checkpoint, with the model architecture and autoregressive inference workflow adapted to the Diffusers Modular Pipeline API.

Echo-WM Flash is a four-step causal audiovisual world model. Given an initial image, a text prompt, and a camera-action sequence, it jointly generates video and synchronized audio with a bounded sink-plus-FIFO KV cache.

Diffusers components

• EchoWMTransformer3DModel
• EchoWMFlashModularPipeline
• EchoWMFlashBlocks

Installation

pip install -U diffusers transformers accelerate

Usage

The following example uses the official wm_causal_cases/0079 input, prompt, and camera trajectory.

import torch
from diffusers import ComponentsManager, ModularPipeline
from diffusers.utils import load_image

components_manager = ComponentsManager()
components_manager.enable_auto_cpu_offload(device="cuda")
pipe = ModularPipeline.from_pretrained(
    "Echo-Team/Echo-WM-Flash-Diffusers", components_manager=components_manager
)
pipe.load_components(dtype=torch.bfloat16)

image = load_image(
    "https://raw.githubusercontent.com/jd-opensource/JoyAI-Echo/main/echo_wm/examples/wm_causal_cases/0079/input.jpg"
)
prompt = (
    "An enchanted crystal cave with massive prismatic crystal formations in purple, teal, and pink. "
    "Bioluminescent fungi glow on the cave floor and walls. Floating light motes drift through the air. "
    "The crystals refract light into rainbow spectra. To the right, a large crystalline cave monster with "
    "glowing purple eyes lurks behind tall crystal clusters. Deep cavern atmosphere with ethereal luminescence. "
    "Further to the right beyond the monster, a subterranean crystal pool glows with turquoise light, fed by a "
    "thin waterfall dripping from a stalactite cluster. The cave opens into a wider chamber with an ancient stone "
    "altar covered in glowing runes. First-person viewer. First-person view with the right hand holding a twisted "
    "wooden magic wand topped with a bright blue-white crystal orb that radiates light. The wand rotates together "
    "with the viewer's perspective when turning."
)

result = pipe(
    image=image,
    prompt=prompt,
    action="l-96,l-96,l-96,l-96",
    num_frames=241,
    generator=torch.Generator(device="cuda").manual_seed(42),
    output=["videos", "audio"],
)

video = result["videos"][0]
audio = result["audio"][0]

Flash uses four distilled denoising steps and requires num_frames = 1 + 24 * n. The default 241-frame output uses 31 latent video frames. Unlike Base, Flash is guidance-distilled and does not accept negative_prompt.

Camera controls

Each action segment uses the form <keys>-<duration>. W/S move forward or backward, A/D strafe, I/K pitch, J/L yaw, and none holds the camera still. Segments are joined with commas, for example w-60,a-60,w-60,d-60.

Original project

• Original checkpoints
• Project page
• Source code
• Paper

Citation

@article{zhang2026echowm,
  title         = {EchoWM: Open and Enterable Omnimodal World Models},
  author        = {Zhang, Songchun and Li, Yaowei and Zhuang, Junhao and Jin, Weiyang and Wang, Haoyu and Lu, Xin and Sun, Yilang and Zhang, Shiyi and Li, Haoran and Ma, Xiaoxiao and Li, Yuming and Liu, Yijun and Su, Yaofeng and Ma, Yanwen and Wu, Haoyu and Su, Zihan and Ma, Yue and Zhang, Lvmin and Huang, Haoyang and Xue, Zeyue and Rao, Anyi and Duan, Nan},
  journal       = {arXiv preprint arXiv:2608.23189},
  year          = {2026},
  eprint        = {2608.23189},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

License

Academic research and non-commercial use only. This model follows the LTX-2 Community License.