4 free upgrades for local AI video released this week

The last four days were unusually good for anyone running video models on their own machine. Four separate tools shipped between September 28 and September 29, each free, each doing a job that used to need a script or a paid API call. Here is what to try first.

ComfyUI v0.38.0 (Comfy-Org)

The release landed on September 29 and the headline feature is a Draft mode for Seedance 2.5 that renders a fast 480p preview, then uses that preview's draft_task_id to drive the 1080p final. You iterate cheaply, pick the take, and pay for one full-resolution render instead of four.

If you work with existing footage, the more useful addition is ID-V2V: identity-preserving video-to-video on Wan 2.1 and VACE-based models, built to stop a face drifting off its source across a clip. The release also adds the w6a8 quantization format, drops the torchaudio dependency, and deletes the Sora video nodes, which have been dead since OpenAI shut that API on September 24. Pull the repository or run your install's updater and restart.

Try it

• Tool: ComfyUI
• Maker: Comfy-Org
• Version: v0.38.0
• Released: September 29, 2026
• Cost: free
• Get it: https://github.com/comfy-org/ComfyUI/releases/tag/v0.38.0

Vlo v0.3.0 (PxTicks)

Vlo is a free, local, open-source timeline editor, and v0.3.0, tagged on September 28, is the release that makes it interesting to a ComfyUI user. Its new live ComfyUI bridge shares assets with a running ComfyUI instance and captures media back into the timeline, so a workflow can run against a selection instead of against a file you exported first.

The same release adds an extension SDK, adjustment clips, colour grading, composite clips and subtimelines, audio stem splitting through Sam-audio, and bundled MiniMax workflows. It is a young project, and the composite-clip feature is the kind of thing that eats an evening, so start with the bridge before you rebuild your edit around it.

Try it

• Tool: Vlo
• Maker: PxTicks
• Version: v0.3.0
• Released: September 28, 2026
• Cost: free, open source
• Get it: https://github.com/PxTicks/vlo/releases

LanPaint 2.2.0 (scraed)

LanPaint is a masked-sampler inpainting tool for ComfyUI, and 2.2.0, tagged on September 28, adds Qwen-Image 2.1 to both the text-to-image and the image-edit checkpoints, plus a fix for a MiniMax H3 bug.

The part worth knowing is the alpha channel. Qwen-Image 2.1 outputs transparent PNGs, and because the sampler can now work on the alpha channel as well as the colour, a repaired region can be given a new outline rather than just new pixels inside the old one. If you are cutting a character out of a frame, that is the difference between a patch and an edit.

Try it

• Tool: LanPaint
• Maker: scraed
• Version: 2.2.0
• Released: September 28, 2026
• Cost: free, open source
• Get it: https://github.com/scraed/LanPaint/releases

AnyAngle (lilylilith)

AnyAngle is the odd one out: a LoRA for Qwen-Image 2.1 that changes the camera angle of an existing picture, listed by the community's ComfyUI Wiki on September 28. Prompt-only angle changes tend to land on a few fixed azimuths and elevations and drift into a different look when you push past them, so this splits the job in two.

You reconstruct the scene as a Gaussian splat or a 3D model, move the camera in Blender and render a coarse image of the new view, then feed that render and the original into Qwen-Image 2.1 with the LoRA and the prompt "Change the camera angle from image2 to image1." The card calls for LoRA strength 1.0 and CFG 3.0 with 20 steps or more; the weight is a rank-24 adapter at about 120MB in bf16. It fails where the reconstruction does: a coarse or spatially wrong splat puts objects in the wrong place and can deform faces.

Try it

• Tool: AnyAngle
• Maker: lilylilith
• Version: rank-24 LoRA for Qwen-Image 2.1
• Released: September 28, 2026
• Cost: free
• Get it: https://huggingface.co/lilylilith/QI_2.1_AnyAngle