ruwwww/Anima-ConvRot

Anima DiT Collection - ConvRot INT8 Quantized Models

Official high-fidelity ConvRot INT8 quantized checkpoints for the Anima DiT family (Cosmos 2 DiT architecture) created by circlestone-labs/Anima.

Why ConvRot INT8?

Standard FP8/MXFP8 quantization on Diffusion Transformers introduces activation outlier truncation errors, often leading to color banding, blurry eye/facial details, or subtle anatomical deformities.

ConvRot (Convolution-like Regular Hadamard Rotation, arXiv:2512.03673) solves this by pre-rotating weight and activation coordinates with an orthogonal regular Hadamard matrix ($N_0=256$, $H H^\top = I$).

• ~90% Winrate vs MXFP8: Produces outputs nearly identical to BF16 ground truth in blind side-by-side evaluations.
• Fast Execution: Reaches 2.59 it/s (11.5s total) on RTX 5060 Ti 16GB (832x1216, 30 steps, batched CFG $B=2$)—only ~0.3s difference from raw MXFP8.
• VRAM Savings: File size reduced from 3.89 GB down to 2.41 GB (~38% reduction) with peak active VRAM under ~5.2 GB.
• Zero Custom Nodes: Standard ComfyUI comfy_quant metadata allows comfy-kitchen to dispatch ck.int8_linear kernels out of the box.

Available Checkpoints

Checkpoint File	Base Model	Suggested Steps	Suggested CFG	Description
anima-base-v1.0-convrot-int8.safetensors	Base v1.0	25 - 30	4.0 - 5.0	Standard production foundation model
anima-turbo-v1.0-convrot-int8.safetensors	Turbo v1.0	8 - 12	1.0 - 2.0	Distilled fast few-step generator
anima-turbo-v1.1-convrot-int8.safetensors	Turbo v1.1	8 - 12	1.0 - 2.0	Updated turbo with improved sharpness
anima-aesthetic-v1.0b-convrot-int8.safetensors	Aesthetic v1.0b	25 - 30	4.0 - 5.0	Fine-tuned for anime aesthetic fidelity
anima-preview-convrot-int8.safetensors	Preview 1	25 - 30	4.0 - 5.0	Early release checkpoint
anima-preview2-convrot-int8.safetensors	Preview 2	25 - 30	4.0 - 5.0	Preview edition v2
anima-preview3-base-convrot-int8.safetensors	Preview 3 Base	25 - 30	4.0 - 5.0	Preview edition v3

Preserved Layers (100% BF16 Fidelity)

Following the proven DiT quality preservation policy:

• blocks.0 and blocks.1 (Initial latent representation)
• blocks.27 (Final reconstruction block)
• All adaln_modulation modules (Timestep and text guidance projections)
• x_embedder and final_layer

Quantized with ConvRot ($N_0=256$):

• Blocks 2 through 26 Attention projections (q_proj, k_proj, v_proj, output_proj)
• Blocks 2 through 26 Feed-Forward Networks (mlp.layer1, mlp.layer2)

Usage in ComfyUI

1. Place any downloaded checkpoint in ComfyUI/models/diffusion_models/.
2. Select it directly in the UNETLoader node.
3. Launch ComfyUI with --fast --use-ck-attention.

For the full optimization recipe and converter tools, see: https://github.com/ruwwww/anima-fastpath-recipe