🎈 What is Anima_LiquidMix?

Anima_LiquidMix is an experimental merge model designed to modify rendering characteristics while preserving Anima's original prompt understanding capabilities as much as possible.

It was created to explore how far texture, rendering behavior, and character knowledge can be modified through custom LoRAs and experimental model merging.

Because these experiments attempt to preserve the underlying internal structure of the base Anima architecture as much as possible, the resulting checkpoints can also be used experimentally as base checkpoints for LoRA training.

Possible use cases include:

• Testing LoRA behavior

• Training experiments

• Comparing rendering characteristics between different merged checkpoints

• Testing expanded Anima architectures

• Comparing prompt adherence across different generations of Anima models

🧬 Anima 2.9B / 40-Layer Version

In this experimental version, the existing Anima_LiquidMix weights are expanded into a 40-layer architecture while attempting to preserve the rendering characteristics of the original LiquidMix as much as possible.

The 12 additional DiT blocks are inserted at the following positions:

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

This version explores whether the rendering characteristics developed in the original LiquidMix can be preserved when adapted to the expanded 40-block architecture.

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

🫗 Anima 3.8B / 52-Layer Version

In this experimental version, LiquidMix is adapted to the 52-block Anima 3.8B architecture.

Relative to the 40-block structure used in the preceding LiquidMix experiments, the additional blocks used in this reconstruction are placed at the following positions:

3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47

The purpose of this version is to investigate whether the rendering characteristics developed in LiquidMix can be preserved when adapted to the larger Anima 3.8B architecture.

The underlying Anima 3.8B architecture and the Expanded Conditioning / Dual Encoder approach originate from the work of lylogummy on Anima 3.8B / Anima Expanded 3.8B.

The 3.8B version of LiquidMix can also be used experimentally as a base checkpoint for training native Anima 3.8B / 52-layer LoRAs.

🔀 Anima 3.8B LiquidMix10138 Dual

LiquidMix10138 Dual is an experimental LiquidMix reconstruction based on Anima3.8B_LiquidMix09035, adapted for use with the existing Anima 3.8B Dual Conditioning architecture.

It attempts to preserve some of the rendering characteristics of LiquidMix09035 in selected components while retaining the newer Gen2 and Connector / Adapter components of Anima 3.8B.

The internal configuration used in this experiment is as follows:

• STEM blocks: LiquidMix09035 weights

• Gen1 blocks: LiquidMix09035 weights

• Gen2 blocks: Anima 3.8B structure and weights

• Connector / Adapter components: Anima 3.8B structure and weights

The Dual Conditioning architecture, Qwen3.5 integration, and Expanded Conditioning approach used by this model are based on the original Anima 3.8B / Anima Expanded 3.8B work by lylogummy.

🧠 Qwen3.5 4B / Semantic Connector v2

Support for Qwen3.5 4B and the expanded Semantic Conditioning pathway originates from the original Anima 3.8B / Anima Expanded 3.8B architecture.

The Expanded Conditioning system can incorporate richer semantic information from natural-language prompts into image generation, in addition to conventional tag-based prompting.

This may be useful for instructions involving:

• Multiple characters with different attributes

• Different hair colors, outfits, or objects assigned to individual characters

• Spatial relationships such as left/right positioning

• Interactions between multiple characters

• More detailed composition instructions written in natural language

For example:

"A red-haired character is standing on the left, while a blue-haired character is standing on the right, and each character is holding a different object."

This is an example of the type of structured instruction for which the Expanded Semantic Conditioning architecture is intended.

Traditional Danbooru-style tags can still be used, allowing experimentation with tag-based prompts, natural-language prompts, or hybrid prompts combining both approaches.

For the design, implementation, and original documentation of the Qwen3.5 / Expanded Conditioning system, please refer to the original Anima 3.8B projects listed below.

🔀 Dual Conditioning

Qwen itself is not directly embedded in the LiquidMix10138 Dual checkpoint.

LiquidMix10138 Dual is intended to be compatible with the following two Conditioning pathways provided by the Anima 3.8B ecosystem:

• Qwen 0.6B: Anima Native Conditioning

• Qwen3.5 4B: Expanded Conditioning using the Anima 3.8B Connector / Adapter system

This makes it possible to experimentally compare the two Conditioning pathways without modifying the LiquidMix10138 image-generation checkpoint itself.

🐲 Changes / Experimental Goals

Depending on the LiquidMix version, the merge experiments may affect characteristics such as:

• Contrast behavior

• Color and lighting

• Texture rendering

• Glow and visual effects

• Rendering tendencies for fantasy / furry / creature characters

• Character knowledge inherited from the source Anima checkpoints

• Behavior with natural-language and hybrid prompts

• Differences in behavior between 28-layer / 40-layer / 52-layer architectures

🧪 Purpose

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

Anima_LiquidMix is an experimental project intended to observe and evaluate how model merging, LoRA integration, architectural adaptation, and different Conditioning systems affect image-generation behavior.

The Anima 2.9B version investigates how the characteristics of the original LiquidMix behave when adapted to a 40-layer DiT architecture.

The Anima 3.8B version extends this experiment further into the 52-block Anima 3.8B architecture.

LiquidMix10138 Dual goes one step further by investigating whether LiquidMix characteristics can be retained in selected STEM and Gen1 components while preserving the newer Gen2 and Connector-related components used by Anima 3.8B.

It is also possible to compare Native Qwen 0.6B Conditioning with Expanded Qwen3.5 4B Conditioning while using the same experimental LiquidMix checkpoint.

The Dual Encoder / Expanded Conditioning architecture used for this comparison originates from the Anima 3.8B / Anima Expanded 3.8B work by lylogummy.

🍩 Recommended Uses

• Comparing multiple Anima-based checkpoints

• Comparing 28-layer / 40-layer / 52-layer LiquidMix experimental models

• Comparing Qwen 0.6B and Qwen3.5 4B Conditioning

• LoRA training experiments

• Testing attribute binding between multiple characters

• Testing prompts involving spatial relationships and character interactions

• Generating fantasy / furry / insect / monster characters

• Testing glow effects, visual effects, and high-contrast rendering

• Comparing character knowledge inherited from different Anima checkpoints

🦠 Recommended Settings

General

Sampler: ER SDE / Euler a / DPM++
Steps: 20–30
CFG: 4–6
Resolution: 1024–1536

The 2.9B and 3.8B versions may behave differently from the original 28-layer LiquidMix.

🐽 Requirements

For standard / native Anima text encoding, qwen_3_06b_base is recommended.

For the VAE, QwenimageVAE_liquid1127 is recommended to obtain the rendering characteristics intended for LiquidMix.

Anima 2.9B

When using the 40-layer Anima 2.9B version, your ComfyUI environment must be able to correctly recognize and load the expanded 40-block DiT architecture.

LoRAs created for the original 28-layer Anima architecture are not necessarily directly compatible with the 40-layer model without modification.

When using legacy 28-layer LoRAs, an appropriate compatibility LoRA Loader / Block Remapping method may be required.

Anima 3.8B

When using the 3.8B version, your environment must support the 52-block Anima 3.8B architecture.

Native 3.8B LoRAs use the expanded architecture, so direct compatibility between 28L / 40L / 52L LoRAs should not be assumed.

The intended Conditioning configuration for LiquidMix10138 Dual is:

• Qwen 0.6B: Anima Native Conditioning

• Qwen3.5 4B: Anima 3.8B Expanded Conditioning

Qwen 0.6B and Qwen3.5 4B are external Text Encoders and are not embedded in the LiquidMix10138 checkpoint itself.

When using the Expanded Qwen3.5 pathway, an appropriate Anima 3.8B Expanded Conditioning / Connector environment and workflow are required.

For the requirements and implementation details of this Conditioning system, please refer to the original Anima 3.8B documentation.

Because the 3.8B architecture is larger than previous versions, it also requires more VRAM and computational resources, particularly during LoRA training.

📛 Notes / Important Information

• This is not an official Anima model.

• It is an experimental LiquidMix merge / adaptation model created for testing purposes.

• Both the Anima 2.9B and Anima 3.8B LiquidMix versions are experimental.

• The 40-layer and 52-layer versions have different internal structures from the original 28-layer model.

• Legacy LoRAs may behave differently on the expanded architectures.

• For training specifically for Anima 3.8B, Native 52L LoRAs are recommended.

• LiquidMix10138 Dual is intended for use with both the Native and Expanded Conditioning workflows of Anima 3.8B.

• The Dual Encoder / Expanded Conditioning architecture was not developed by Anima_LiquidMix.

• The Qwen models themselves are not embedded in the checkpoint.

• Qwen3.5 support requires an appropriate Anima 3.8B generation environment and workflow.

• Unexpected outputs or unstable behavior may occur.

• Compatibility may vary depending on the ComfyUI version, Custom Nodes, and Model Loader implementation.

🔗 Original Anima 3.8B / Expanded Conditioning

The Anima 3.8B architecture, Dual Encoder approach, Expanded Conditioning, and Qwen3.5 integration used by LiquidMix10138 Dual originate from the work of lylogummy on Anima 3.8B / Anima Expanded 3.8B.

For the original implementation, models, requirements, and documentation, please refer to the following projects:

Anima Expanded 3.8B
https://civarchive.com/models/2910974/anima-expanded-38b

Anima 3.8B
https://civarchive.com/models/2880480/anima-38b

Anima 3.8B — Hugging Face
https://huggingface.co/lylogummy/Anima-3.8B

Anima_LiquidMix10138 Dual is an experimental derivative / adaptation model created for use with this existing architecture. It does not claim to be the origin of the Dual Encoder or Expanded Conditioning approaches.

🤝 Additional Acknowledgements

Special thanks to bestluner180 for providing technical advice and insights regarding the Anima 3.8B block architecture, Gen2 structure, and Qwen3.5 / Connector compatibility.

These insights were helpful during the LiquidMix reconstruction experiments.

No checkpoints or model weights from bestluner180 are included in or merged into Anima_LiquidMix.

🥚 Feedback

Feedback and generation examples are welcome.

In particular, comparisons involving the following would be very helpful:

• Original Anima_LiquidMix / 28-layer architecture

• Anima 2.9B / 40-layer LiquidMix version

• Anima 3.8B / 52-layer LiquidMix version

• Qwen 0.6B Native Conditioning vs. Qwen3.5 4B Expanded Conditioning

• Tag-only prompts vs. natural-language / hybrid prompts with Anima 3.8B