languel/genereticore-sdxs-coreml

GeneretiCore sdxs · 256px + 512px

Converted local image models for text and sketch. Includes tokenizer and text encoder. Downloads run on your Mac; no hosted inference, prompts or drawings are uploaded. Requires macOS 14+ Apple silicon; tested on one macOS 27 Mac with ComfyUI. Students do not convert weights. First use compiles Core ML packages locally.

Install

Follow GeneretiCore setup. Choose the bundle in the generator's Download control, then explicitly press Start. The pinned download is verified by size and SHA-256. This is an evaluation release; older macOS and broad Manager installations remain untested.

Credits and terms

See Notice.txt, provenance, original model cards under notices, and full license copies under notices/licenses. SDXS and Sketch are by IDKiro, declared OpenRAIL++; DreamShaper teachers by Lykon retain OpenRAIL-M terms. Tiny VAE originates in TAESD by Ollin Boer Bohan (MIT). Text encoder implementation notices credit OpenAI CLIP and OpenCLIP contributors. Source use restrictions continue to apply to these converted packages.

Scope

Fixed 256px and 512px tensor shapes and FP16 conversion; sketch fuses controller/denoiser. The base SDXS UNet and text assets are shared runtime dependencies of Turbo. Canny/depth/pose packages are not included. Canny's upstream terms are unclear; use the original source and contributor conversion instructions if appropriate, or use this bundle's image mode. No compiled caches, credentials or user media.

Code: GeneretiCore.

Resolution selection

Use the existing SDXS or SD Turbo Comfy node and select 256/unet.mlpackage or 256/turbo_unet.mlpackage for 256px; select the corresponding 512/ path for 512px. No separate node or student conversion is needed. The 256/ and 512/ folders contain complete fixed-shape packages; text assets are shared. 256px packages reuse the original 512-trained weights with smaller fixed tensor shapes, so image quality/composition may differ from 512px.

Converted with coremltools 8.3.0, PyTorch 2.5.1, diffusers 0.32.2 and the same pinned upstream weights as 512px. Tested locally in Comfy on one Apple silicon Mac: output dimensions 256 × 256; SDXS text/sketch and Turbo image mode. Native first-use compilation/loading is separate from warm inference. Contributor converter: convert_native_images.py.