ViFeEdit — Public project presence¶

Development line: project:vifeedit · thread public-project-presence
Last event: 2026-03-18 · 1 dated since 2026-03-18 · Researched: 2026-09-05 · confidence: medium

What it is¶

ViFeEdit adapts Wan and similar text-to-video Diffusion Transformers for video generation and editing without video training data. - 3D attention reparameterization: separates spatial independence. - Dual-path pipeline: routes separate timestep embeddings across paired 2D source and target images.

Training takes 100–250 paired images and fewer than 20 epochs in practice. We treat it as a task-specific tuning workflow rather than a turnkey video editor.

Development line¶

• 2026-03-18 — ViFeEdit public project resources were linked. On 2026-03-18, we linked the development line to a public GitHub repository and a Hugging Face link. The links identify the project resources. They do not establish a release, version, capability change, or publication status.

What changed¶

2026-03-16: the ViFeEdit preprint introduced video-free tuning for video Diffusion Transformers using only 2D image pairs.
2026-03-17: the authors released training and inference code.
2026-03-18: the project became available through the linked repository. The repository adds the practical limits omitted by the event links: 100–250 paired images per task, training limited to the first 20 epochs, and an inference plus optional post-processing path.

How to use this¶

As of 2026-03-18, we use the linked GitHub repository as the primary project source. We treat the Hugging Face link as a related resource until research confirms its role.

1. Clone the repository and install it in an isolated Python environment with pip install -e .. Reuse a DiffSynth Conda environment only if its dependencies are compatible. — https://github.com/Lexie-YU/ViFeEdit
2. Create paired source and target images plus a metadata_vife.csv that maps each target prompt and target image to the source image used as the inference-video reference. — https://github.com/Lexie-YU/ViFeEdit
3. Set data and output paths in train_vife.sh, run the script, and pick a checkpoint from the first 20 epochs instead of running the configured 500 epochs. — https://github.com/Lexie-YU/ViFeEdit
4. Configure the provided inference example and run python inference.py. Use postprocess.py only when additional output alignment is needed. — https://github.com/Lexie-YU/ViFeEdit

Best practices¶

• Start with a narrowly defined edit and 100–250 paired images. Keep source and target pairing and metadata names consistent, because training uses the source image as the later source-video reference. — https://github.com/Lexie-YU/ViFeEdit
• Treat the first 20 epochs as the practical training window. Evaluate checkpoints there before extending a run. — https://github.com/Lexie-YU/ViFeEdit
• Pin or verify transformers==4.55.0. If editable installation fails on pkg_resources, use the repository’s documented --no-build-isolation flag rather than changing dependencies. — https://github.com/Lexie-YU/ViFeEdit

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The linked hf.ru URL returned an error during verification, so we do not treat it as an official model-weight release.
• The repository documents code, data preparation, and inference, but does not identify a downloadable ViFeEdit checkpoint or a supported production deployment target.
• No post-2026-03-18 first-party release note exists in the sources used here.

Sources¶

Agent brief¶

• Subject: project:vifeedit, thread public-project-presence, 1 dated events 2026-03-18 → 2026-03-18.
• Practical note: As of 2026-03-18, use the linked GitHub repository as the identified project source. Treat the Hugging Face link as a related resource until research establishes its exact role.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.