Anima-2.9B_LiquidMix / Anima_LiquidMix

🎈 What is Anima_LiquidMix?

Anima_Liquid is an experimental merge model designed to modify rendering tendencies while preserving Anima’s original prompt understanding as much as possible.

It was created by integrating custom-made LoRAs into multiple checkpoints in order to test how much texture, rendering characteristics, and visual behavior can be controlled through model merging.

Because the basic internal structure of Anima is preserved after merging, the model can also be used as a base checkpoint for LoRA training.

This makes it useful for purposes such as:

• Testing LoRA behavior

• Training experiments

• Comparing rendering behavior between different merged checkpoints

🧬 Anima 2.9B / 40-Layer Version

As a new experimental branch, a version based on Anima 2.9B has been created, expanding the main DiT from the original 28 blocks to 40 blocks.

In this version, the existing Anima_LiquidMix weights were expanded into the 40-layer structure while attempting to preserve the original LiquidMix rendering characteristics as much as possible.

The 12 additional DiT blocks introduced in Anima 2.9B are placed at the following positions:

2, 5, 8, 11, 14, 17, 21, 24, 27, 30, 33, 36

This 2.9B checkpoint can also be used as a base model for future LoRA training.

🐲 Changes from Preview Base

• Modified contrast behavior

• Added more variation in color and lighting

• Enhanced texture rendering

• Modified glow and effect rendering

• Changed rendering tendencies for fantasy / furry / creature-themed characters

• Added experimental support for the 40-layer Anima 2.9B architecture

• Improved support for newer character knowledge derived from Anima 2.9B

🧪 Purpose

This model was not created to reproduce any specific art style.

Instead, it is an experimental model intended to observe and test how model merging and LoRA integration affect rendering behavior.

The Anima 2.9B version also explores how rendering changes when a checkpoint originally developed around the 28-layer Anima architecture is expanded into a 40-layer DiT structure.

Because of this, generated results may vary significantly depending on:

• Prompt

• Sampler

• Seed

• LoRA

• Checkpoint version

• Other generation settings

🍩 Recommended Use

• Comparing multiple Anima-based checkpoints

• Comparing 28-layer and 40-layer Anima models

• Testing Anima 2.9B merge behavior

• Testing LoRA-integrated models

• LoRA training experiments

• Generating fantasy / furry / insect / monster characters

• Testing glow, effects, and high-contrast rendering

• Testing newer characters learned by Anima 2.9B-derived checkpoints

🦠 Recommended Settings

• Sampler: ER SDE / Euler a / DPM++

• Steps: 20–30

• CFG: 4–6

• Resolution: 1024–1536

The Anima 2.9B version may behave somewhat differently from the original 28-layer models.

Depending on the prompt and subject matter, adjusting the sampler or CFG may improve results.

🐽 Requirements

For the text encoder, qwen_3_06b_base is recommended.

For the VAE, QwenimageVAE_liquid1127 is recommended for more stable rendering, color balance, and highlight reproduction.

When using the 40-layer Anima 2.9B version, your ComfyUI setup must also be capable of correctly detecting and loading the expanded 40-block DiT architecture.

In addition, older LoRAs made for the original 28-layer Anima architecture may not be fully compatible with the 40-layer model without modification.

For legacy 28-layer LoRAs, a compatibility LoRA loader can be used to remap the original block indices to their corresponding positions in the 40-layer model.

📛 Notes

• This is not an official model

• This is an experimental merge created for testing purposes

• The Anima 2.9B version is also experimental

• The 40-layer version has a different internal structure from the original 28-layer Anima models

• Legacy LoRAs may behave differently on the 40-layer model

• Unexpected outputs or unstable behavior may occur

• Compatibility may vary depending on the ComfyUI version and model loader implementation

🥚 Feedback

Feedback and example generations are welcome.

In particular, comparison results between:

• The original Anima_LiquidMix

• The 40-layer Anima 2.9B version

• Legacy 28-layer LoRAs used with the 40-layer version

would be especially helpful for further testing and development.