Input-title/Anima_Turbo_V1.1_ACDM

ACDM — Anima Turbo v1.1 (INT4 Compressed)

⚠️ EXPERIMENTAL — RESEARCH IN PROGRESS

This is an experimental compressed model checkpoint. Quality, memory usage, compatibility, and performance are still being evaluated. It is not recommended for production use.

Overview

ACDM (Anchor-Compressed Delta Models) is an experimental model-compression approach designed to reduce the storage footprint of diffusion transformer models while attempting to preserve their original generation behavior.

This release applies ACDM compression to Anima Turbo v1.1, an anime-oriented text-to-image model. The goal is to make the model's weights more compact while retaining useful visual quality for local image generation.

The compressed checkpoint uses a reduced-precision representation of model weights. It is intended to work with the associated ComfyUI integration rather than serving as a standalone image-generation pipeline.

Current checkpoint

Property	Details
Model	ACDM — Anima Turbo v1.1
Source model	Anima Turbo v1.1
Format	Safetensors
Compression	Experimental INT4 weight compression
Original checkpoint size	Approximately 3.90 GB in the author's measurements
Compressed checkpoint size	Approximately 2.11 GB in the author's measurements
Primary interface	ComfyUI
Status	Experimental

File sizes are preliminary measurements from the author's environment. They do not guarantee equivalent reductions in peak VRAM usage or inference time.

Model Sources

• Anima Turbo v1.1 — Civitai: View model and version
• Anima — Hugging Face: circlestone-labs/Anima
• Upstream base model — NVIDIA Cosmos-Predict2-2B-Text2Image: View base model

ACDM is a compressed derivative of Anima Turbo v1.1, not a newly trained replacement for the original model.

ComfyUI Integration

⚠️ Important development note

The repository linked below provides the ComfyUI custom-node integration for loading the compressed checkpoint. The optimized custom loading/dequantization kernel intended to improve inference speed is still in development. Do not assume that the current compressed checkpoint generates images faster than the original model.

ComfyUI custom node repository:
ComfyUI_DualStreamCompression — GitHub

Basic Setup

1. Install the custom-node package using the instructions in the repository.

2. Place anima_turboV11_int4_compressed.safetensors in:

   ComfyUI/models/diffusion_models/

3. Restart ComfyUI.

4. Add the Load Compressed DiT (Dual-Stream) node.

5. Connect its MODEL output to your workflow's KSampler.

Use the repository's current installation instructions and compatible workflow. The appropriate text encoder, VAE, and other pipeline components may still be required from the original Anima setup.

Results and Model Comparison

The purpose of this release is to explore whether a smaller model checkpoint can retain useful image-generation quality. Results may vary with the prompt, seed, sampler, resolution, and inference settings.

Benchmark results will be added here as testing progresses.

Quantitative Comparison

Metric	Original Anima Turbo v1.1	ACDM INT4
Checkpoint size	~3.90 GB	~2.11 GB
More Comparasions will be added once the custom kernal is built!!	Null	Null

Generation speed and quality comparisons should be measured using the same hardware, prompt, seed, resolution, sampler, and other relevant settings.

Visual Comparison

Add your original and compressed model outputs below. For a meaningful comparison, use matching prompts and inference settings.

Original Anima Turbo v1.1	ACDM INT4 Compressed
Original checkpoint	Compressed checkpoint

Limitations

• This is an experimental compression of a specific model, not a universal compression solution for all diffusion transformers.
• Image quality and prompt adherence may differ from the original checkpoint.
• Actual VRAM requirements depend on the full inference pipeline, hardware, software, and memory-management behavior.
• Reduced checkpoint size does not necessarily mean faster inference.
• The optimized custom loading/dequantization kernel is still under development.
• Compatibility may change as the model and ComfyUI integration evolve.

Intended Use

This checkpoint is intended for experimental local image generation, model-compression evaluation, and testing through the associated ComfyUI integration.

It should not be treated as a production-ready or universally compatible replacement for the original Anima Turbo v1.1 model.

License and Usage Restrictions

NO LICENSE IS GRANTED FOR THE ACDM-SPECIFIC CHECKPOINT OR IMPLEMENTATION. All rights reserved.

No permission is granted to use, reproduce, modify, redistribute, sublicense, or deploy the ACDM-specific materials without explicit prior written authorization from the author.

The underlying Anima model is subject to the applicable CircleStone Labs Non-Commercial License and relevant upstream terms. Users are responsible for reviewing and complying with all applicable licenses.

Acknowledgements

This work builds upon the Anima Turbo v1.1 model from CircleStone Labs and its upstream NVIDIA Cosmos-Predict2 model family.

ACDM is an independent experimental compression project and is not an official release or endorsement by CircleStone Labs, Comfy Org, or NVIDIA.