Forge Neo Apple Accelerator

Normal Anima generation in 23.9% less time. Hi-Res Fix in 35.4% less time.

Keep the Forge interface, prompts, samplers, checkpoints, and workflows you already use. Install the extension, restart Forge, and its qualified Apple routes activate automatically. There are no separate nightly or performance switches.

Results at a glance

Measured on the M5 Pro development Mac with fresh Forge processes, identical prompts and seeds, counterbalanced run order, and nominal thermal checks:

Accelerated workflow	Add-on disabled	Add-on enabled	Measured improvement
Normal Anima generation — 1024×1024, 10 Euler a steps	22.74s	17.30s	23.9% less time / 1.31×
Complete Anima Hi-Res Fix — 1024→1536, 10+10 steps	134.49s	86.88s	35.4% less time / 1.55×
SwinIR stage inside that Hi-Res request	42.57s	12.40s	70.9% less time / 3.43×

Normal-generation evidence · Hi-Res evidence

Important

Normal text-to-image acceleration is currently Anima-only. SDXL models such as Nova Comic XL do not receive a base-generation speedup yet.

What is accelerated

Workflow	Support	Released behavior
Normal text-to-image	Anima	Metal RoPE + guarded MPSGraph attention
Hi-Res Fix	Anima + SwinIR_4x.pth	Stable Torch 2.14 VAE + BF16/compiled SwinIR + GPU compositing
Normal text-to-image	SDXL / Illustrious / Nova XL	No denoiser speedup yet
Tiled / Ultimate upscale	—	Not released yet

Unsupported models, shapes, and dtypes fall back to Forge. See MODEL_SUPPORT.md for the exact boundary.

Benchmark details

All public comparisons use fresh Forge processes, identical prompts and seeds, and explicit thermal checks on the M5 Pro development Mac.

Normal Anima generation

The packaged v0.2 lifecycle was tested at 1024×1024 and 10 Euler a steps in accelerated → vanilla → vanilla → accelerated order. All thermal boundaries were nominal.

Add-on disabled median	Add-on enabled median	Improvement
22.74s	17.30s	23.9% less / 1.31×

Same-route repeats were pixel-exact. Cross-route SSIM was 0.93084, with the same composition, detail level, and overall visual quality under full review.

Machine-readable v0.2 normal-generation matrix

Hi-Res Fix: installed add-on vs disabled add-on

Counterbalanced accelerated → vanilla → vanilla → accelerated test, 1024→1536, 10+10 Euler a steps, SwinIR_4x.pth:

Measurement	Add-on disabled	Add-on enabled	Improvement
Complete request median	134.49s	86.88s	35.4% less / 1.55×
SwinIR median	42.57s	12.40s	70.9% less / 3.43×

The two vanilla outputs were pixel-exact to each other, and the two accelerated outputs were pixel-exact to each other. Cross-route SSIM was 0.91819; visual review found the same subject, composition, detail level, and overall quality, with minor line/highlight placement drift. This is visual parity, not strict same-seed identity across runtimes.

The very first Hi-Res request on a clean install builds the SwinIR graph cache: 112.47s total and 35.26s in SwinIR in the measured run. After restart, the cached samples were 91.78s and 81.98s. Clean-cache and cached outputs were pixel-exact.

On the 48 GB test Mac, a 1024 source uses one full-image MPS SwinIR call. It is normal for the old tile progress bar to disappear: no CPU tile loop runs on that path. Lower-memory Macs automatically retain a conservative tiled route.

Machine-readable Hi-Res release matrix

Full timings, route proof, and caveats are in BENCHMARKS.md.

Install

1. Open Extensions → Install from URL in Forge Neo.
2. Paste https://github.com/Gohro/forge-neo-apple-accelerator and install.
3. In Installed, choose Apply and restart UI, then fully stop and restart the Forge backend once.

Installation downloads the pinned official stable torch 2.14.0 and torchvision 0.29.0 Apple wheels into an extension-owned overlay, verifies their SHA-256 hashes, and builds an exact-runtime native bridge. It does not replace the Torch packages in Forge's virtual environment.

While the extension is enabled, a small startup bootstrap selects that verified overlay before Forge imports Torch. Disable the extension in Installed and restart to return to Forge's original runtime and behavior.

Requirements

• Apple Silicon Mac (arm64)
• Forge Classic / Forge Neo neo branch
• Xcode Command Line Tools
• Forge's current Python 3.13 environment
• Internet access during the first install for the pinned PyTorch wheels

Tested Forge commit: e91c293170e4451fc27490827f9aa62ccc77e6dc.

Verify

Open Settings → Apple Accelerator for a read-only runtime/route status table, or query:

GET /internal/forge-apple-accelerator/status

Native self-test:

PYTHONPATH=extensions/forge_apple_accelerator \
  venv/bin/python -m forge_apple_accelerator.self_test

Current limits

• SDXL base generation is not accelerated in v0.2.
• Released Hi-Res measurements cover Anima with the exact SwinIR_4x.pth checkpoint on the 48 GB development Mac; lower-memory Macs use a conservative 512 tile automatically and need a separate published hardware matrix.
• The first use of a new SwinIR input shape builds a persistent graph cache.
• Ultimate SD Upscale is not yet claimed as accelerated.

Project docs

• Model support
• Benchmark evidence
• Changelog
• License and notices

License

GNU Affero General Public License v3.0. See LICENSE and NOTICE.md.