File size: 2,699 Bytes
139852f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | # JoyAI-Image-Edit-ComfyUI (NVFP4 Quantized)
This repository provides **NVFP4** quantized weights derived from the original BF16 weights of [jdopensource/JoyAI-Image-Edit-ComfyUI](https://huggingface.co/jdopensource/JoyAI-Image-Edit-ComfyUI). The conversion enables efficient usage of the JoyAI Image Edit model within ComfyUI.
## Quantization Details
The quantization was performed using:
- [Comfy Kitchen](https://github.com/Comfy-Org/comfy-kitchen)
- [comfy-dit-quantizer](https://github.com/bedovyy/comfy-dit-quantizer)
Quantization configuration file: `config_nvfp4.json`
Two weight variants are provided:
| Filename | Description |
|----------|-------------|
| `joyai_image_edit_nvfp4_uncalib.safetensors` | Uncalibrated (zero‑point free) NVFP4 weights |
| `joyai_image_edit_nvfp4.safetensors` | Calibrated NVFP4 weights |
The calibrated weights were generated using `CALIB_DATA.json` – a calibration dataset produced by running only **3 editing tasks**.
## Important Notes
To use this quantized version, you **must** either:
1. Switch to the official pull request submitted by JoyAI:
[Comfy-Org/ComfyUI#14428](https://github.com/Comfy-Org/ComfyUI/pull/14428)
or
2. Use my custom node repository:
[judian17/ComfyUI-JoyAI-Image-Edit](https://github.com/judian17/ComfyUI-JoyAI-Image-Edit)
## Purpose
- Reduce memory footprint and potentially improve inference speed on GPUs that support NVFP4.
- Maintain compatibility with the original model’s behavior while offering a compressed alternative.
- Provide an easy integration path for ComfyUI users.
## Usage
1. Download the desired weight file (`joyai_image_edit_nvfp4.safetensors` or `joyai_image_edit_nvfp4_uncalib.safetensors`) from this repository.
2. Place the weights into the appropriate ComfyUI model directory (e.g., `ComfyUI/models/joyai/`).
3. Ensure your ComfyUI installation includes the required custom node or the specific PR.
4. Load the model using the corresponding node – it will automatically recognise the NVFP4 format.
## Links
- Original BF16 model: [HuggingFace](https://huggingface.co/jdopensource/JoyAI-Image-Edit-ComfyUI)
- Required PR: [Comfy-Org/ComfyUI#14428](https://github.com/Comfy-Org/ComfyUI/pull/14428)
- Alternative custom node: [ComfyUI-JoyAI-Image-Edit](https://github.com/judian17/ComfyUI-JoyAI-Image-Edit)
- Comfy Kitchen: [https://github.com/Comfy-Org/comfy-kitchen](https://github.com/Comfy-Org/comfy-kitchen)
- comfy-dit-quantizer: [https://github.com/bedovyy/comfy-dit-quantizer](https://github.com/bedovyy/comfy-dit-quantizer)
## License
Please refer to the original model’s license for terms of use. |