Wan2.2-FastWan-5B-Paiton-RDNA4

Wan2.2 and FastWan 5B — Paiton RDNA4 runtime package

Two qualified silent-video settings for one AMD Radeon AI PRO R9700, 32 GB, gfx1201. This repository provides runtime integration, pinned container recipes, compiled Paiton artifacts, benchmarks and retained clips. Original model weights download from their upstream publishers at pinned revisions. Private compiler source is not included.

One-command ComfyUI launch

Linux, Docker with Compose, working AMD GPU device access, and Git LFS are required. Install Git LFS and run git lfs install before cloning so the compiled artifact and media are materialized.

FastWan:

git clone --depth 1 https://huggingface.co/EliovpAI/Wan2.2-FastWan-5B-Paiton-RDNA4 paiton-wan22 && cd paiton-wan22 && chmod +x models/Wan2.2/*.sh models/FastWan/*.sh && ./models/FastWan/launch.sh

Wan2.2 with optional image upload:

git clone --depth 1 https://huggingface.co/EliovpAI/Wan2.2-FastWan-5B-Paiton-RDNA4 paiton-wan22 && cd paiton-wan22 && chmod +x models/Wan2.2/*.sh models/FastWan/*.sh && ./models/Wan2.2/launch.sh

Choose one command for the initial checkout. From an existing checkout, run the other model's launch.sh to add that preset. Both share persistent caches and one ComfyUI service on 0.0.0.0:8192. Open http://localhost:8192/?paiton=1&preset=fast or http://localhost:8192/?paiton=1&preset=base, edit the prompt and click Run. Stock/Paiton selection, duration and resolution are in the connected workflows; the base workflow accepts image uploads.

First launch builds the runtime locally and downloads weights. Allow 100 GB free disk; retained model files occupy about 28.15 GB for FastWan, 18.15 GB for base, or 38.15 GB together. Qualification used about 15 GiB usable host RAM and 4 GiB swap; 32 GB RAM provides more headroom. First-run loading and compilation are slower than warm clips. Outputs default to ~/paiton-videos.

Measurements and examples

Means of two measured runs after two warmups, including fresh conditioning and file encoding. FastWan's measured Paiton latency reduction is 2.8–4.7%; base image cases did not improve end to end. Distillation is identical between engines and is not a compiler gain. All paired measured denoised latents were bitwise equal; decoded pixels have small rounding differences.

Full benchmark data · Quality assessment and paired clips · Comparison demo · Pinned inputs

At 24 fps, the two/five-second controls produce 49/121 frames, encoding as 2.0417/5.0417 seconds. 720 class is 1280×704. FastWan image input is not qualified. Both models generate silent video; no audio synthesis, stitching or interpolation is included.

Licenses and reproduction

See third-party notices and retained licenses. Runtime bindings and recipes are Apache-2.0; dependencies retain their own terms. ComfyUI is assembled locally from pinned source. The combined ComfyUI image is not published. Model originals are preserved; FastWan conversion only renames keys and records provenance. SHA256SUMS covers the packaged files. Terminal generation and benchmark commands are in the model guides.