Skip to content

Feature request: MLX port for CogVideoX (text-to-video / image-to-video) #42

Description

@dahai80

Summary

Requesting an MLX port / pipeline for CogVideoX (e.g. THUDM/CogVideoX-5b, THUDM/CogVideoX-2b) so it can run natively on Apple Silicon alongside the existing ltx_2 and wan_2 pipelines.

Why

mlx-video currently ships only ltx_2 and wan_2 (confirmed under mlx_video/models/). CogVideoX is a popular open text-to-video / image-to-video diffusion model, but there is no native MLX path for it. Without one, downstream Mac inference servers (we run fusion-mlx on top of mlx-video) cannot offer CogVideoX and must either fall back to CPU/PyTorch or decline the request.

What's needed

An MLX implementation of the CogVideoX pipeline:

  • 3D VAE (encoder/decoder) in MLX
  • DiT transformer in MLX
  • The CogVideoX noise scheduler (CogVideoXDDPM-style) in MLX
  • A generate_video(...) entrypoint mirroring the ltx_2/wan_2 signatures (prompt, num_frames, height, width, num_inference_steps, cfg_scale, seed, output_path, tiling)

Context

In our downstream server we currently expose CogVideoX as an explicitly unimplemented backend that auto-detects cogvideo/cog_video model paths and fails loudly pointing here. We would love to replace the NotImplementedError with a real delegator to mlx-video once a port lands.

Happy to help test a WIP port on M-series hardware if anyone picks this up.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions