PS5SD

stable-diffusion.cpp on the PS5 GPU — a native title running ggml-vulkan on RADV, with an OpenAI/A1111-compatible image API, a browser UI and a TV dashboard.

Status: 0.8.0. The title runs the server, the web UI, the model downloader and the TV dashboard on the console. 0.2.0 added LoRAs; 0.3.0 several images per run, your own size buttons, a done sound and a switch that keeps images and prompts off the TV; 0.3.1 an API key (for access from outside your network) and a Presets tab; 0.4.0 the hires fix and an Upscale button. 0.5.0 adds image to image and inpainting: start from any image, paint a mask over what should change, and only that part is made again. 0.5.2 shows the console's address on the TV with a QR code: a phone's camera opens the web UI. 0.6.0 makes short videos with Wan 2.1 (2 seconds at 832x480 in about 15 minutes), played in the page and the gallery, and offers two LoRAs that make SDXL models fast. 0.7.0 adds ControlNet (a guide image whose edges set the layout, with SDXL) and CausVid, which makes a video in under 4 minutes. 0.7.1 adds a refine pass (the hires fix at ×1) and says why a run failed, out of memory for one. 0.8.0 rebuilds the web UI for clarity: five tabs, settings in groups, a count next to Generate that queues any number of images, a Models tab in three parts, and the measured time per image in the catalog. Downloads run at 50 to 75 MB/s from start to end and resume after a close; the web UI uploads models from the PC at about 80 MB/s, and its Help tab shows docs/MODELS.md and docs/API.md. Closing from the home screen during a download used to end in a crash report about 1 time in 5; that has not happened in 8 Closes since the download fix. The measurements are in docs/CONSOLE.md.

Install: unpack the release's PPSA99576.zip into /data/homebrew/ on the console (the zip holds the PPSA99576 folder) and let ShadowMountPlus scan, or install it from the homebrew store.

What comes next (longer videos, more ControlNets): ROADMAP.md.

Milestone: the first image made on a PS5

2026-10-19, FW 13.00: a photograph of an astronaut riding a horse, SDXS (Q8_0), 512x512, one step, seed 42, on the console's GPU through ggml-vulkan and RADV, in 2.6 s. Measurements in docs/CONSOLE.md.

The web UI

Open the address the TV shows (or scan its QR code) on a phone or a PC on the same network. Create picks the model, the mode (Image, Edit, Guide, Video) and the settings in groups, with the result next to them; the count next to Generate sends that many images, with new seeds, and more can be queued while it runs. Gallery keeps every image and video, Models installs, switches and removes models, LoRAs and upscalers, Settings holds the done signal, the TV and the API key, and Help explains the rest.

Requirements

• A jailbroken PS5 on firmware 7.00–13.60, with:
  • etaHEN and kstuff-lite
  • ShadowMountPlus
  • elfldr (port 9021)
  • ftpsrv (port 2121), optional: the model downloader and the web UI's uploads write their files through an FTP server, because the console slows a title's own writes to about 2 MB/s after 1.4 GB. When nothing answers on 2121, the title starts the copy of ftpsrv it carries (GPL-3.0+, John Törnblom and drakmor; third_party/ftpsrv at 1.16-ng-stable with patches/ftpsrv). That copy listens on 127.0.0.1:2122 only, so it opens nothing on the LAN. It runs until the console restarts. scripts/ps5sd-app.sh (deploy, close, log) still needs ftpsrv on 2121.
• A host with a Vulkan-capable toolchain for building the payload and the driver patches.

Models

The title downloads models itself, from the web UI: a starter catalog (SDXL Base 1.0, FLUX.3 Image Klein, Z-Image Turbo and a few smaller ones; one click installs the files and a preset) or any Hugging Face or Civitai file URL, with progress, cancel and resume. Downloads keep going while images are generated. A gated or private Hugging Face repository needs a token in /data/PS5SD/hf-token.txt, a Civitai model that asks for a login one in /data/PS5SD/civitai-token.txt (one line each).

LoRAs go the same ways (as LoRA in the upload or download form; a Civitai LoRA is recognised by itself) and are picked on the Create tab, several at once, each with a strength.

A model from anywhere else (your own merge or training, another site, a file already on the PC) goes onto the console with the web UI's upload button or by FTP; docs/MODELS.md has the steps, the preset format for models of several files, and which models fit. How the downloader works is in docs/PORTING.md.

API

Port 7860 serves the web UI and the A1111, OpenAI and sd-server APIs, plus PS5SD's own routes:

• the model library;
• switching and deleting models;
• the downloader;
• the job queue and cancel.

With no model yet, the title starts in standby, so a fresh install can fetch its first model from the web UI. Every route is in docs/API.md.

The API is open on your network until you turn on the API key (web UI, Create, Settings, API key). Then every client needs it, the TV shows it, and the console can be put on the internet behind a Cloudflare Tunnel or reached through a VPN; docs/API.md shows how. Do not forward port 7860 on your router: the console speaks plain HTTP.

Building

See docs/BUILDING.md.

Credits

• PS5LM by cobanov — the project PS5SD is modelled on.
• stable-diffusion.cpp by leejet, and ggml.
• PS5_Vulkan and PS5_Mesa by mihawk-99 — the Vulkan driver stack.
• Machete PS5SD by VenezuelanBiggie24 — the console address and QR code on the TV come from its version.
• QR Code generator library by Project Nayuki (MIT), ps5/app/qrcodegen.{hpp,cpp}, unchanged from commit 3c6d0b3.
• libwebp and libwebm by Google (BSD-3-Clause), through stable-diffusion.cpp's submodules: the WebM videos. Their licences are in the title's licenses folder.
• ps5-homebrew-ui by blackbearreloaded (GPL-3.0). The original repository is gone; the submodule uses the copy at Phononenom/ps5-homebrew-ui, pinned at the same commit.

License

PS5SD is licensed under the GNU General Public License v3.0 or later (GPL-3.0-or-later); see LICENSE. Third-party parts keep their own licences.