🎨 Enhanced Object Separation and Color Control
Overview
We proudly introduce Illustrious v3.5-v-parameterization (v3.5-v), showcasing significant advancements in understanding precise natural language prompts, such as "the red hair girl on the left, the blue hair girl on the right." The model now effectively demonstrates object separation and improved instruction-following capabilities, distinguishing itself from previous variants.
Knowledge Cutoff
The dataset for Illustrious v3.5-v-parameterization is updated comprehensively, covering knowledge up to November 2024. This expanded dataset incorporates diverse themes and content, ensuring broad applicability and current relevance.
Key Features
Resolution and Prompt Clarity
1. Supports generation resolutions from 256x256 to 2048x2048, enabling detailed and high-quality outputs.
2. Basic natural language prompt separation capabilities significantly improved, allowing clear positional and descriptive control.
Restored Color Controls
1. Color control tokens, previously unavailable in the intermediate version, are fully restored, enabling precise manipulation of visual attributes.
Zero Terminal SNR and v-parameterization Advantages
1. The model incorporates "Zero Terminal Signal-to-noise Ratio" (Zero Terminal SNR), enhancing its ability to disregard initial noise and faithfully interpret complex instructions.
2. The v-parameterization clearly demonstrates the ability to accurately separate and position described objects, addressing prior limitations attributed to denoising rather than CLIP capabilities.
Academic and Practical Significance
1. Proven through comparison with Illustrious v3.5-epsilon—trained with identical datasets and setups but different objectives—v3.5-v reliably positions objects and colors independent of low-frequency seed-determined features.
2. This variant showcases that the denoising objective was previously the critical limitation rather than CLIP’s capacity.
LoRA Compatibility and Training
1. Supports experimental LoRA training methods specific to v-parameterization.
2. Existing LoRA adapters from v0.1 remain fully compatible, even at higher resolutions such as 2048, with recommended usage.
Description
🎨 Enhanced Object Separation and Color Control
Overview
We proudly introduce Illustrious v3.5-v-parameterization (v3.5-v), showcasing significant advancements in understanding precise natural language prompts, such as "the red hair girl on the left, the blue hair girl on the right." The model now effectively demonstrates object separation and improved instruction-following capabilities, distinguishing itself from previous variants.
Knowledge Cutoff
The dataset for Illustrious v3.5-v-parameterization is updated comprehensively, covering knowledge up to November 2024. This expanded dataset incorporates diverse themes and content, ensuring broad applicability and current relevance.
Key Features
Resolution and Prompt Clarity
1. Supports generation resolutions from 256x256 to 2048x2048, enabling detailed and high-quality outputs.
2. Basic natural language prompt separation capabilities significantly improved, allowing clear positional and descriptive control.
Restored Color Controls
1. Color control tokens, previously unavailable in the intermediate version, are fully restored, enabling precise manipulation of visual attributes.
Zero Terminal SNR and v-parameterization Advantages
1. The model incorporates "Zero Terminal Signal-to-noise Ratio" (Zero Terminal SNR), enhancing its ability to disregard initial noise and faithfully interpret complex instructions.
2. The v-parameterization clearly demonstrates the ability to accurately separate and position described objects, addressing prior limitations attributed to denoising rather than CLIP capabilities.
Academic and Practical Significance
1. Proven through comparison with Illustrious v3.5-epsilon—trained with identical datasets and setups but different objectives—v3.5-v reliably positions objects and colors independent of low-frequency seed-determined features.
2. This variant showcases that the denoising objective was previously the critical limitation rather than CLIP’s capacity.
LoRA Compatibility and Training
1. Supports experimental LoRA training methods specific to v-parameterization.
2. Existing LoRA adapters from v0.1 remain fully compatible, even at higher resolutions such as 2048, with recommended usage.