Text-to-image material generation, AI image editing for texture application, first-frame / last-frame video transitions, controlled reveal sequences, AI audio generation, all with Open Source AI models, and ComfyUI-based pipelines.

VALENTAINS 2026

VALENTAINS is a self-initiated concept project exploring how generative AI can be used to create symbolic, material-driven storytelling through structured pipelines.

The project begins with the creation of a heart shape in 2D, developed in Affinity as both a white and red heart on a black background. The white heart was then translated into a clean white 3D base form, used as the structural foundation for material exploration. The red 2D heart was also converted into a red 3D version, serving as the emotional core of the sequence.

From there, we generated an internal material library using Open Source text-to-image models and our very own prompt-system for materials, inside a ComfyUI pipeline. Materials generated included:

• Cracked asphalt

• Concrete

• Soil and earth

• Ice and snow

• White ceramic

• White canvas paper

• Marble

• Rough stone

• Red stone

• Brown leather

• Egg shell

• Light stone

These materials were generated using the text-to-image Open Source AI model Qwen Image 2512, forming a reusable texture system.

We then used another Open Source AI model for image-editing called Qwen Image Edit 2511. We applied these generated textures directly onto the 3D white heart, preserving the exact geometry while transforming the surface material. The result was a series of identical heart forms, each with a distinct material identity.

Now it was time to generate videos and for that we used WAN 2.2, another Open Source AI video model for image-to-video by using a first-fram-last-frame workflow. Read more about that further down. Next step in the workflow is upscaling the video (or as we call it in GENERAITR Post-Generation). For that we used yet another Open Source AI model called SeedVR2.

And finally, it was time to add some sound effects and music. For that, we used the final Open Source model, this time an Audio model for text-to-audio called Stable Audio.

It was then all put together and edited in Final Cut Pro, both as a slightly longer video and as short social media videos of each heart’s heartbeat.

Phase 2, Revealing the Red Heart

• Cracking open

• Breaking apart

• Crumbling

• Shattering

• Melting

• Burning

• Peeling away 

• Unfolding

• Cutting in half

• Core breaking through

Not all movements was used in the final videos.

Open Source & Infrastructure

VALENTAINS is built entirely using Open Source AI models:

• Qwen Image 2512 for material generation

• Qwen Image Edit 2511 for texture application

• WAN 2.2 for motion and video generation

• Stable Audio for audio generation (heartbeat)

• Ace-Step 1.5 for music generation
• SeedVR for upscaling

All pipelines were built and executed in ComfyUI, using structured node-based workflows for repeatability and control.

All generation was performed on EU-based servers, ensuring transparency, data control, and alignment with regulatory requirements.

In Summary

VALENTAINS explores:

• Structured material generation as a reusable texture library

• Image-to-image texture application with preserved geometry

• First-frame / last-frame and prompts for controlled motion

• Symbolic material transformation

• Audio and music generation to support

• Open Source AI pipelines executed in ComfyUI

• Emotion-driven storytelling through structured generative systems

The project demonstrates how generative AI can move beyond isolated images and into controlled, repeatable audiovisual storytelling systems.