Upload README.md with huggingface_hub

README.md

@@ -10,19 +10,35 @@ tags:
 - diffusion-single-file
 - convrot
 - int8
+- hadamard
 pipeline_tag: text-to-image
 ---
 
-# Anima DiT Base v1.0 - ConvRot INT8 Quantized
+# Anima DiT Collection - ConvRot INT8 Quantized Models
 
-This repository provides the **ConvRot INT8** quantized checkpoint of [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima) (Base v1.0, Cosmos 2 DiT architecture).
+Official high-fidelity **ConvRot INT8** quantized checkpoints for the **Anima DiT** family (Cosmos 2 DiT architecture) created by [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima).
 
-## Key Advantages of ConvRot INT8 vs MXFP8 / FP8
+## Why ConvRot INT8?
 
-- **Orthogonal Outlier Smoothing:** Unlike standard FP8/MXFP8 quantization which suffers from severe activation outliers in DiT channels, ConvRot applies a group-wise Regular Hadamard Transform (RHT, $N_0=256$) to disperse outliers evenly across all channels.
-- **Superior Semantic Fidelity (~90% Winrate vs MXFP8):** Generates outputs that are visually and structurally almost identical to the BF16 ground truth while eliminating color banding, blurry textures, or anatomical distortion.
-- **High Generation Speed:** Runs at **~2.59 it/s (~11.5s total)** on RTX 5060 Ti 16GB at 832x1216 (30 steps, batched CFG $B=2$) with native ComfyUI compile.
+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)
 
@@ -38,11 +54,9 @@ Quantized with ConvRot ($N_0=256$):
 
 ## Usage in ComfyUI
 
-This model embeds native ComfyUI `comfy_quant` metadata. It requires **ZERO custom node registration**:
-1. Download `anima-base-v1.0-convrot-int8.safetensors` to `ComfyUI/models/diffusion_models/`.
-2. Select it directly in the standard `UNETLoader` node.
+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`.
-4. ComfyUI and `comfy-kitchen` will automatically route the layers to hardware-fused `ck.int8_linear` kernels on your GPU.
 
 For the full optimization recipe and converter tools, see:
 https://github.com/ruwwww/anima-fastpath-recipe