Release of Version 3 of the Photorealism Model

I am sharing the third test iteration of my photorealism model. This update introduces full manual labeling, giving you control over image modification through instructions.

Key improvements in version 3

Detailed control of instructions: All images in the dataset of 678 pairs have been manually labeled. This allows you to request specific changes or the preservation of specific features directly in the instructions.

If the model insists on modifying things that weren't requested, reduce its intensity a bit or try to provide more detail in the prompt.

Training steps: 1500

• Trigger Word: transform to photorealism

For example:

- transform to photorealism, keep background unchanged

- transform to photorealism, keep the fur skin

- transform to photorealism, reduce head size

Sampler / Scheduler:

- Euler / Simple

- Euler / Linear_Quadratic <- Strong modification

I'd love to see the tests you've done using the model. Feel free to post them.

========================================

What’s New in v2

Dataset Expansion: Increased to 674 high-quality professional photos with broader contextual variety to enhance overall model comprehension.

Lighting & Effects: Added diverse lighting conditions and volumetric light effects for better depth and atmospheric accuracy.

New contexts added:

- Vector art

- Modern anime

- Digital painting

- Manga

- Cartoon

- Oil paint

- 8-bit pixel art

- Simple draw sketch

Everything is geared toward conversion specifically to photorealism while maintaining consistency, though it can also yield good results in other styles.

It has not been tested or designed for the 6-step Viggle model, therefore I am not sure if it will have good results if used in conjunction with it.

Training Steps: Extended from 250 to 1500 steps.

• Trigger Word: transform to photorealism

Sampler / Scheduler:

- Euler / Simple

- Euler / Linear_Quadratic <- Strong modification

==================================================

First Test: Photorealism Model

Sharing my first practical test in training a model focused on transforming images into photorealism.

Technical Specifications

• Software: Ostris (ai-toolkit)

• Dataset: 172 high-quality images (2K resolution), with all "before and after" pairs manually aligned.

• Trigger Word: transform to photorealism

• Train Steps: 250 (currently; planning to extend training to 2,500–5,000 steps).

Methodology

To ensure that the final render quality is grounded in real-world details rather than synthetic approximations, I used a reverse workflow: real photographs were used as the AFTER state, which were then converted into vector art using the Qwen Image 2.1 model to serve as the BEFORE state. This guarantees that the learning baseline is anchored in genuine realism.

Recommended Settings

• Sampler / Scheduler: Euler / Simple

• CFG: 4

• Steps: 25–40

Important Note regarding ComfyUI:

The official text encoder currently has incorrect standard multiples. To achieve zero pixel drift and maintain maximum fidelity, it is recommended to use:

• Multiples of 28 for the VL

• Multiples of 16 for the VAE

Since this is my initial test with a relatively low step count, I appreciate your understanding if some outputs are not yet fully refined. Feedback is always welcome!