- Home
- Skills
- Image & Video
- SDXL on AMD / Intel GPUs via Vulkan (No CUDA or ROCm)
SDXL on AMD / Intel GPUs via Vulkan (No CUDA or ROCm)
Vendor PyTorch builds skip many AMD and Intel cards on Windows and ROCm is Linux-first, so people with a capable 16 GB card are told they cannot run SDXL locally.
$29
Works with the AI Agents you already use
SDXL on AMD / Intel GPUs via Vulkan (No CUDA or ROCm)
Example session with this skill installed
I have a Radeon RX 6800 XT on Windows and every SDXL guide wants CUDA or ROCm. Get Stable Diffusion XL running locally on this card, fast, without installing a Python stack.
- Read your context and instructions
- Compiled the sdxl on amd
Summary
SDXL runs on the RX 6800 XT through Vulkan: 8 seconds per 1024x1024 image on the fast preset, 47 seconds on quality, nothing installed system-wide, every downloaded byte hash-verified.
What was done
pwsh -File scripts/Test-Scripts.ps1: 15 checks passed (no download)..\scripts\setup.ps1: plan printed (tool release master-1024-77b72a2, three models, 14.5 GB)..\scripts\setup.ps1 -Install: four files downloaded with resume, each hashed and moved into place only on a match;sd-cli.exe --helpanswered..\scripts\generate.ps1 "a red fox in a snowy pine forest at dawn, wildlife photography":Done in 8.3 s: out\20261029-141502-fast-184719-*.png, sidecar JSON beside it..\scripts\generate.ps1 "an old fisherman mending a net, harbour at dusk" -Preset quality -Width 832 -Height 1216 -Seed 7:Done in 49.1 s.
Why it was slow before
The card fails the direct-convolution self-check, so the tool decoded each image on the slow path: 34 s per image. --conv-direct auto,vae=on (the default in generate.ps1) decodes in 1.5 s and matched the default decode at 50.7 dB PSNR. Flash attention was measured 7x slower on this card and stays off.
Licences
SDXL base and Lightning: CreativeML OpenRAIL++-M (commercial use allowed with use restrictions). Tool and VAE: MIT. Face-identity adapters left out: their face models are non-commercial.
Connects securely to your tools. The creator never sees your data.
About this skill
Vendor PyTorch builds skip many AMD and Intel cards on Windows and ROCm is Linux-first, so people with a capable 16 GB card are told they cannot run SDXL locally. They can: stable-diffusion.cpp over Vulkan runs it on anything with a Vulkan driver. This kit is the pinned install, the presets and the measurements that make it fast instead of painfully slow.
setup.ps1: plan mode prints what will be downloaded and where; install mode fetches the pinned stable-diffusion.cpp Vulkan build and three models (SDXL base 1.0, SDXL-Lightning 4-step, the fp16-fix VAE; 14.5 GB) with resume, hashes each .part file and moves it into place only when it matches the approved SHA-256, deletes a mismatch, and re-runs download nothing already correct. Nothing system-wide, no driver changes.
generate.ps1: fast (Lightning 4-step, guidance 1, about 8 s at 1024x1024, 6 s each in a batch) and quality (SDXL base, 30 steps, guidance 7, about 47 s); size validation (multiples of 64, 512 to 1536); seed, count, negative prompt; a .json sidecar beside every PNG; -DryRun prints the exact command.
Why it is fast (reference/speed-settings-and-why.md, measured on an RX 6800 XT): the card fails the direct-convolution self-check so decode took 34 s; forcing it on for the VAE only decodes in 1.5 s and matched the default at 50.7 dB PSNR, while tiled decode (21 dB) and CPU decode (60 s) were rejected; flash attention measured 7x slower; the Lightning LoRA matched 0 of 2,364 tensors so the full checkpoint is used; the fp16-fix VAE avoids black images.
Licences (reference/models-and-licences.md): every file with source, revision, licence and hash; OpenRAIL++-M use restrictions; why face adapters are left out. 15 offline checks (Test-Scripts.ps1) that download nothing. PowerShell 5.1 or 7.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Video Skills.
Requires GPU with 8GB+ VRAM and Vulkan support. Windows only. No face-identity adapters included. Downloads are 14.5 GB. One generation job at a time to avoid VRAM exhaustion.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 19 days ago
Windows only
- Software
- PowerShell 5.1 or 7curl.exeGPU with Vulkan and 8 GB+ VRAM
16 GB of VRAM is comfortable.
Permissions
Shell to run the PowerShell scripts and tests; read/write files for the tool, models and output folders. Network access is used only by setup.ps1 -Install to download from github.com and huggingface.co; generation makes no network calls. No environment variables.