Join my discord for access to Version 5: https://discord.gg/8wMr2mmaRF

Follow me on HuggingFace for all of my other releases that I cannot post here:

https://huggingface.co/obsxrver/

Update: I2Pee-4 Public Release

I2Pee-4 is the best freely available blink piss LoRA today. You can download and run it for free now.

No GPU? No Problem

Run ComfyUI on a cloud GPU using Vast.AI with this template https://cloud.vast.ai/?ref_id=208628&creator_id=208628&name=ComfyUI%20-%20Wan%202.2%20LoRA%20demo

WORKFLOW:

Download the workflow file https://huggingface.co/obsxrver/wan2.2-i2v-piss/resolve/main/config.json?download=true and drag into ComfyUI.

SageAttention:

The Vast.AI template already provisions the instance, downloading all necessary LoRA Files and installing SageAttention. However, if running locally, you will need to install SageAttention to run the workflow:

``` pip install sageattention ```

Technical Explanation

Blink introduces a novel training methodology for Image-to-Video (I2V) diffusion transformers, termed Identity Invariant Action Mapping (IIAM). This technique addresses the persistent challenge of identity drift and hallucination in generative video by decoupling subject identity from temporal action dynamics.

Traditional I2V training often results in the model creating a smooth, continuous video, Blink teaches the model to perform a hard-cut transition, switching to a completely new scene, while preserving the facial and identity characteristics of the input image. We achieve this by utilizing a specific data preprocessing strategy:

1. Temporal Splicing: Training data consists of a 0.3-second static frame (the "before" image of the subject. Derived from the video using Qwen-Image-Edit) prepended to a 4.7-second action sequence (the "after". Whatever you are training the LoRA to do). This "jumpcut" forces the model to attend to the initial static latents as the ground truth for identity.

2. Identity-Agnostic & Discontinuity-Forcing Captioning:

* Template: "a [man|woman], jumpcut, after the transition, [he|she] is (any change in appearance), [he|she] is (the action happening in the "after" sequence)"

By stripping any uniquely identifying descriptors (e.g., hair color, facial structure, clothing), using only generic identity references ("man", "woman", "he", "she"), the model is forced to map the learned motion vectors and physics simulations directly onto the visual topology provided in the initial frame. The model learns to bind the discontinuous transition - the "blink" - to the "jumpcut, after the transition" tokens.

Future Applications:

The IIAM/Blink framework is content-agnostic and can be applied to any set of action clips to create modular, "plug-and-play" motion LoRAs for generative video workflows.