# PAN 2 single-file models for ComfyUI Scripts that turn the [IFM/PAN2](https://huggingface.co/IFM/PAN2) release (diffusers folders, about 81 GiB) into single `.safetensors` files in ComfyUI's quantized formats. ## What you get | Target name | File | Size | Notes | |---|---|---|---| | `transformer_int8_convrot` | `diffusion_models/PAN2_int8_convrot.safetensors` | 39.8 GiB | int8 ConvRot; the model's fp32 `norm1.linear` layers stay fp32 | | `transformer_int8_convrot_full` | `diffusion_models/PAN2_int8_convrot_full.safetensors` | 27.4 GiB | int8 ConvRot, those layers quantized too | | `transformer_nvfp4` | `diffusion_models/PAN2_nvfp4.safetensors` | 15.5 GiB | NVFP4 (4-bit); runs natively on RTX 50 series, emulated elsewhere | | `te_int8_convrot` | `text_encoders/PAN2_qwen3_5_int8_convrot.safetensors` | 8.4 GiB | Qwen3.5-9B text encoder | | `te_nvfp4` | `text_encoders/PAN2_qwen3_5_nvfp4.safetensors` | 4.6 GiB | Qwen3.5-9B text encoder | | `vae_fp16` | `vae/PAN2_vae_fp16.safetensors` | 1.3 GiB | the fp32 VAE cast to fp16 | Weight error against the original (relative, measured per layer): int8 ConvRot about 0.9%, NVFP4 about 9%. ## Requirements - Python 3.10+ with `torch` (2.1 or newer, CPU build is enough), `safetensors`, `numpy`, `huggingface_hub` - Disk: the download (transformer 63 GiB, text encoder 15 GiB, VAE 2.7 GiB) plus the outputs you choose - RAM: 16 GiB or more recommended. No GPU needed. ```bash pip install torch safetensors numpy huggingface_hub ``` ## Use Make everything, downloading what is needed into the Hugging Face cache: ```bash python convert_pan2.py --out /path/to/ComfyUI/models ``` Only some files: ```bash python convert_pan2.py --out /path/to/ComfyUI/models --only transformer_nvfp4 te_nvfp4 vae_fp16 ``` From a copy you already downloaded (a folder containing `transformer/`, `text_encoder/`, `vae/`): ```bash python convert_pan2.py --out /path/to/ComfyUI/models --src /path/to/PAN2 ``` Files that already exist are skipped; add `--overwrite` to redo them. Each transformer file takes roughly ten minutes on a desktop CPU. `--device cuda` does the math on the GPU. It is faster, but a few values can round differently, so the result may not be byte-identical to a CPU run. ## Check a result ```bash python verify_pan2.py /path/to/PAN2/transformer /path/to/ComfyUI/models/diffusion_models/PAN2_nvfp4.safetensors ``` It confirms that every tensor is present, that everything not quantized is bit-identical to the source, and prints the error of the quantized layers. ## Using the files These load in a ComfyUI build that has PAN 2 support (model code, the `pan2` text encoder type, and loading of diffusers-layout Wan VAEs). Stock ComfyUI without that support will not recognise them. One such build is . - Diffusion model: Load Diffusion Model - Text encoder: Load CLIP, type `pan2` - VAE: Load VAE ## What the script does - Quantizes the large 2-D weights of the transformer blocks and the text refiner (for the text encoder: of its layers) and copies every other tensor unchanged. - int8 ConvRot: each weight is rotated by a block Hadamard matrix of size 256, then stored as int8 with one scale per output row. - NVFP4: 4-bit e2m1 values with one fp8 scale per block of 16 and one fp32 scale per tensor. - NVFP4 text encoder: layers are marked `full_precision_matrix_mult`, and the token embedding table is stored as plain row-wise int8. - The model's `config.json` is kept in the file's metadata. PAN 2 itself is released by IFM under its own license; check the model page before redistributing converted files.