| --- |
| license: apache-2.0 |
| tags: |
| - ccsr |
| - super-resolution |
| - image-to-image |
| - upscaling |
| - controlnet |
| - tensorrt |
| - rtx |
| - comfyui |
| pipeline_tag: image-to-image |
| --- |
| |
| # CCSR: TensorRT RTX Acceleration Engine |
|
|
| Ultra-fast **TensorRT RTX** execution engine and auxiliary modules for **CCSR (Creative Content Super-Resolution)**, designed for real-time generative image upscaling in **ComfyUI**. |
|
|
| > 📦 **ComfyUI Loader & Upscaler Extension:** |
| > All nodes supporting TensorRT engine execution are available in: |
| > 👉 **[https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker](https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker)** |
|
|
| --- |
|
|
| ## 🌟 Overview |
|
|
| Creative Content Super-Resolution (CCSR) is a diffusion-based super-resolution framework leveraging a Controlled UNet and ControlNet structure to synthesize rich photorealistic textures and fine details. |
|
|
| This repository provides an optimized **NVIDIA TensorRT RTX Engine** implementation for CCSR: |
|
|
| - **Fused Denoising Engine (`ccsr_apply_f16io.rtxplan`)**: |
| - Fuses the ControlNet and Controlled UNet denoising computation into a single compiled TensorRT engine. |
| - Fixed 512px tile resolution (64×64 latent tile) executing at ~**24 ms/step** (~**4.7× speedup** over PyTorch FP16 at ~113 ms/step on modern RTX GPUs). |
| - Synchronized stream execution on current PyTorch CUDA streams to eliminate race conditions and deadlocks. |
| - **Engine-Only Deployment (`ccsr_trt_aux.safetensors`)**: |
| - Contains only the essential companion modules: FP16 AutoencoderKL (VAE encoder/decoder) and condition encoder. |
| - Automatically loaded alongside the engine, eliminating the need to download large full checkpoints (~3.2 GB saved). |
|
|
| --- |
|
|
| ## 📦 Available Files |
|
|
| | Filename | Description | Architecture / Components | File Size | Recommended Location | License | |
| | :--- | :--- | :--- | :--- | :--- | :--- | |
| | `ccsr_apply_f16io.rtxplan` | TensorRT Fused Denoising Engine | ControlNet + UNet fused RTX Engine (Tile 512px / Latent 64×64) | ~1.4 GB | `custom_nodes/.../nodes/CCSR/trt_engines/` | Apache-2.0 | |
| | `ccsr_trt_aux.safetensors` | TRT Auxiliary Weights | FP16 VAE AutoencoderKL + Condition Encoder | ~450 MB | `custom_nodes/.../nodes/CCSR/trt_engines/` | Apache-2.0 | |
|
|
| --- |
|
|
| ## ⚙️ Performance & Benchmark Comparison |
|
|
| Measurements conducted on an NVIDIA RTX 4090 / RTX 5090 environment: |
|
|
| | Execution Mode | Files Required | VRAM Overhead (Denoising) | Step Latency (Tile 512) | Speedup | |
| | :--- | :--- | :--- | :--- | :--- | |
| | **Stock CCSR (FP16 PyTorch)** | Full Checkpoint (~3.2 GB) | ~3.8 GiB | ~113 ms / step | 1.0× (Baseline) | |
| | **CCSR TensorRT RTX** | Engine + Aux (~1.85 GB total) | ~2.2 GiB | ~**24 ms / step** | **~4.7× faster** | |
|
|
| --- |
|
|
| ## 🚀 Usage in ComfyUI |
|
|
| TensorRT engine execution for CCSR is integrated natively into the **[ComfyUI-NunchakuFluxLoraStacker](https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker)** custom-node pack. |
|
|
| ### Workflow Example |
|
|
| <p align="center"> |
| <img src="https://huggingface.co/ussoewwin/CCSR-TensorRT-Engine/resolve/main/png/ccsrtensor.png" alt="CCSR TensorRT RTX Engine Workflow in ComfyUI" width="800"> |
| </p> |
|
|
| ### Installation & Setup |
|
|
| 1. **Install the Custom Node Pack**: |
| ```bash |
| cd ComfyUI/custom_nodes |
| git clone https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker.git |
| ``` |
|
|
| 2. **Place Engine & Aux Files**: |
| Download both `ccsr_apply_f16io.rtxplan` and `ccsr_trt_aux.safetensors` and place them directly into the engine directory: |
| ``` |
| ComfyUI/custom_nodes/ComfyUI-NunchakuFluxLoraStacker/nodes/CCSR/trt_engines/ |
| ├── ccsr_apply_f16io.rtxplan |
| └── ccsr_trt_aux.safetensors |
| ``` |
|
|
| 3. **In ComfyUI**: |
| - Add **`Load CCSR Model (TensorRT)`** (`LoadCCSRModelTensorRT`). The node automatically discovers `.rtxplan` files in `trt_engines/` and loads the companion `ccsr_trt_aux.safetensors`. |
| - Connect the `ccsr_model` output to **`CCSR Upscale (TRT)`** (`CCSR_Upscale_TRT`). |
| - Connect an input image to `image`. |
| - Configure upscale parameters: |
| - `tile_size`: 512 (fixed to match the compiled static engine shape) |
| - `tile_stride`: 256 (recommended for seamless blending) |
| - `color_fix_type`: `adain` (or `wavelet` / `none`) |
| - `steps`: Effective diffusion step count (densified schedule guarantees exact execution of requested step count) |
|
|
| --- |
|
|
| ## 📜 Credits & License |
|
|
| - **Original CCSR Implementation & Weights**: [csslc/CCSR](https://github.com/csslc/CCSR) (Apache-2.0 License) |
| - **Research Paper**: *"Creative Content Super-Resolution with Pre-trained Diffusion Model"* by Liang et al. |
| - **ComfyUI CCSR Node Foundation**: [kijai/ComfyUI-CCSR](https://github.com/kijai/ComfyUI-CCSR) & [Kijai/ccsr-safetensors](https://huggingface.co/Kijai/ccsr-safetensors) (Apache-2.0 License) |
| - **TensorRT Engine & ComfyUI Loader**: [ussoewwin/ComfyUI-NunchakuFluxLoraStacker](https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker) (CCSR nodes located under `nodes/CCSR/`) |
| - **License**: **Apache-2.0** |
|
|
