VioletXF/submaroon-manga-refinement-coreml

Optional manga color refinement for Autocomic

This is a modified Core ML conversion of the SD 1.5 ControlNet-manga-recolor15 checkpoint plus AnyLoRA UNet/VAE components. It is intended to refine an already colored v2 image, conditioned separately by the original grayscale drawing. It is not the SDXL checkpoint.

Sources and modifications

• ControlNet: SubMaroon/ControlNet-manga-recolor, revision e8f57f799f0c46bc894b10257bed66c982bca28c; file ControlNet-manga-recolor15.safetensors. Upstream model card declares MIT. Author SubMaroon; training compute credited to Flanayt Pulsar; based on Diffusers and lllyasviel's ControlNet.
• Base: Lykon/AnyLoRA, revision 5dac4354b7f5d32f140e1f0fbc849e92b744f347; CreativeML Open RAIL-M. Derived from Stable Diffusion v1.5. Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors.
• Conversion architecture: apple/ml-stable-diffusion, revision ea2805dc1945be20561c77e5f6d1d9a5a637cda2 (Apple license included).
• All four .mlpackage files are modified conversions by VioletXF: fixed 512-square batch-one tensors; FP16 convolution/linear weights/compute, FP32 other operations; VAE encoder returns scaled posterior mode, decoder accepts scaled latent; UNet accepts 13 ControlNet residual tensors. Target iOS 18 or newer.
• positive.bin and negative.bin are fixed prompt embeddings; noise.bin is fixed seed-42 FP32 noise; schedule.json contains the nine DDIM steps for strength 0.30 of 30 nominal steps, epsilon prediction, eta 0. Use CFG 5 and ControlNet scale 0.8.

The pack's base-derived components and use are governed by the included CreativeML Open RAIL-M license, including Section III paragraph 5 and Attachment A use restrictions, which apply to all users. The ControlNet additionally retains its upstream MIT license declaration and attribution. See the included licenses and upstream model cards; this conversion does not change those terms.

Intended use and limitations

Run v2 first, then img2img refinement with the original grayscale condition. Apply only generated chroma to the full-resolution v2 luminance to retain text and line detail. Raw decoded output can redraw lettering and small faces. Colors may drift or be wrong; this is an optional experimental refinement. No prompt editing or remote inference service is supplied by this pack.

The four components passed numerical comparisons on a Mac. A full native Swift page run closely matched the PyTorch DDIM reference. Device-specific memory, latency and quality vary; the pack is approximately 2.61 GB before device compilation. Complete per-file sizes/hashes are in assets.json and source/precision details in provenance.json. No private comic fixtures or credentials are included.