darkmaniac7/TokForge-DreamShaper-8-CoreML-6bit

Published by TokForge for on-device inference.

TokForge

• Website: https://tokforge.ai
• Discord: https://discord.gg/Acv3CBtfVm
• Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
• iOS TestFlight: https://testflight.apple.com/join/jnufjzRr

Runs on-device in the TokForge app.

TokForge: DreamShaper 8 · CoreML 6-bit (Apple Neural Engine)

A 6-bit palettized Apple CoreML conversion of DreamShaper 8 (Lykon, SD-1.5 realistic finetune), built for on-device image generation in the TokForge iOS app. Converted with Apple ml-stable-diffusion (torch2coreml) using SPLIT_EINSUM_V2 attention and --quantize-nbits 6 (6-bit palettized weights), so it compiles fast on the Apple Neural Engine. It is the default on iOS, replacing an FP16 finetune whose ANE graph compile was too slow.

Part of the TokForge iOS · CoreML Image Models collection.

Files

File	Size	Contents
DreamShaper-8_palettized_split_einsum_v2_compiled.zip	~874 MB	The compiled Swift-CLI resource bundle (a single ZIP of Resources/)
Resources/	~913 MB	The unzipped tree: TextEncoder.mlmodelc / Unet.mlmodelc / VAEDecoder.mlmodelc / VAEEncoder.mlmodelc + vocab.json + merges.txt

The ZIP expands to one wrapper folder (Resources/) holding the compiled .mlmodelc models plus the CLIP vocab.json + merges.txt, the exact layout Apple's StableDiffusionPipeline (and the TokForge installer) loads.

Recommended render settings (standard SD-1.5)

attention:    split_einsum_v2 (Apple Neural Engine)
compute:      .cpuAndNeuralEngine  (palettized -> fast ANE compile)
steps:        20   (8 = fast floor, 40 = extra refinement)
cfg-scale:    7.5
resolution:   512x512  (SD-1.5 native; baked into the compiled model)

How this was built

1. Loaded Lykon/dreamshaper-8 (SD-1.5 diffusers format).
2. Converted UNet + text encoder + VAE decoder + VAE encoder to CoreML with Apple ml-stable-diffusion python_coreml_stable_diffusion.torch2coreml, --attention-implementation SPLIT_EINSUM_V2 (ANE-shaped attention).
3. Applied 6-bit palettization (--quantize-nbits 6): the iOS-17 ANE runtime feature that makes the graph compile fast on the Neural Engine.
4. Bundled the compiled resources for the Swift CLI (--bundle-resources-for-swift-cli).

Conversion peaked at ~9.96 GB RAM (no --chunk-unet needed); runs on Apple silicon, iOS 17+ (6-bit palettized weights require the iOS-17 runtime).

License & attribution

• License: CreativeML OpenRAIL-M (inherited from DreamShaper 8 / Stable Diffusion 1.5). Use is subject to the OpenRAIL-M use restrictions.
• Base model: DreamShaper 8 by Lykon: https://huggingface.co/Lykon/dreamshaper-8 (a Stable Diffusion 1.5 finetune). All credit for the model weights is Lykon's.
• Conversion tooling: Apple ml-stable-diffusion: https://github.com/apple/ml-stable-diffusion (6-bit palettization, SPLIT_EINSUM_V2 attention).
• Built on top of Stable Diffusion 1.5 (Runway/CompVis/Stability).

This repository is a redistribution for on-device use, a format conversion (PyTorch -> CoreML) and 6-bit palettization of Lykon's DreamShaper 8. No weights were retrained. The original OpenRAIL-M terms and attribution requirements propagate to this conversion and any images generated with it. No additional restrictions are imposed by this repackaging.

Credits

• DreamShaper 8 by Lykon (Civitai, Hugging Face). Used under the CreativeML OpenRAIL-M license and the creator's Civitai terms. Lykon's Civitai terms: no selling the model or merges of it.
• Stable Diffusion v1.5 by Runway and CompVis (Hugging Face). Used under the CreativeML OpenRAIL-M license.
• Converted to Core ML with ml-stable-diffusion by Apple (MIT).