Training Image Editing Models Without Human Labels via Video Deltas

haremlifegame · reddit · 2026-08-08

A community discussion reveals a highly inspiring training method that can be used to fine-tune MiniMax H3 or build video model derivatives.

The core idea is to generate massive training data without human instructions:

Take a video dataset and use a vision-language model (like Qwen) to analyze the video, specifically describing what changed between the first and last frames.
Train the model to take the first frame and the 'description of changes' as input to generate the last frame.
By inverting the order of operations, the model learns to execute instructions on 'what ought to change,' effectively becoming a powerful image editor.

This technique is an excellent approach for developing lighting LoRAs, image editing models, or fine-tuning video models.

More from Multimodal