NOTE: This LoRA is experimental and may not work fully as intended.

In many cases, the model will prefer to clone the previous frame over restoring the current frame, preventing the multi-frame temporal referencing from working as intended. It presently works best when served only the current frame, without any prior frame data. Depending on content, it often still presents good temporal stability. It is generally successful at reducing VAE and diffusion artifacts characteristic of Minimax H3 such as temporal smearing, however it is sometimes overly aggressive and may remove fine particles unless prompted not to. Despite it's issues, it often results in a perceptual increase in detail at the same resolution and may still be useful.

Qwen-Image-2.1 β€” MiniMax-H3 Temporal Video Restoration LoRA

A multi-image edit LoRA for Qwen-Image-2.1 trained to remove motion smearing and temporal degradation artifacts from MiniMax-H3 video frames while attempting to preserve composition, lighting, color, and frame-to-frame consistency. Trained at 1344x768.

How It Works (2-Reference Conditioning)

This was LoRA is intended to use Qwen-Image-2.1's multi-image conditioning to perform autoregressive frame-by-frame restoration:

  • <image1>: Clean or previously restored reference frame (t - 1).
  • <image2>: Current degraded frame (t) acting as the structural canvas.

For frame 0, supply either a clean keyframe or frame 0 itself as <image1>, then feed each restored output back into <image1> for the next frame (t + 1).

Trigger / Conditioning Prompt

Remove the smearing artifacts from <image2> as the canvas, keeping its exact composition, poses, lighting, colors, and rendering medium intact. Transfer the native appearance, motion characteristics, and temporal consistency from <image1>.
Downloads last month
1
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Alex240z/Qwen2.1H3Restore

Adapter
(174)
this model