Update production Core ML model card

README.md

@@ -7,64 +7,92 @@ tags:
   - coreml
   - ios
   - inpainting
-  - lora
   - stable-diffusion
 ---
 
-# Clover Image Tiny Inpaint HQ — Core ML
+# Clover Image Tiny Inpaint — Core ML
 
-Core ML resources for the high-quality Clover Image Tiny inpainting pipeline.
-The iOS 18 batch-one U-Net accepts `[1, 9, 64, 64]`, runs classifier-free
-guidance as two serial passes, and exposes 144 mutable Core ML states for exact
-runtime composition of up to three Clover styles.
+The production Core ML conversion of
+[`neonforestmist/Clover-Image-Tiny-Inpaint`](https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint)
+for fully local 512 × 512 inpainting on iPhone and iPad. It uses the same
+trained nine-channel inpainting checkpoint as the Diffusers release; Core ML
+changes the execution format, not the learned weights.
 
-The production U-Net is under `hq-v4-int8/Unet.mlmodelc`. Its immutable
-weights use per-channel symmetric int8 compression, reducing the complete
-inpainting add-on from about 1.79 GB to **931 MB** while leaving mutable LoRA
-state in FP16. Shared tokenizer, text encoder, and VAE decoder resources are
-hard-linked from the required main Clover installation instead of downloaded
-again. `hq-v4-int8/adapter-schema.json` maps each small `.safetensors` style
-into one of three independent state slots.
+Use this model with
+[`neonforestmist/Clover-Image-Tiny-iOS`](https://github.com/neonforestmist/Clover-Image-Tiny-iOS).
+The app downloads, verifies, and installs the resources automatically.
 
-## Validation
-
-PyTorch/Core ML parity on deterministic inputs:
+## Recommended release
 
-| Configuration | FP16 | Int8 production |
-|---|---:|---:|
-| Base HQ inpainting U-Net | 78.56 dB | **59.32 dB** |
-| Monet 0.70 + Pointillism 0.45 + Watercolor Anime 1.10 | 78.47 dB | **59.30 dB** |
+The production entry point is [`manifest-pipeline.json`](manifest-pipeline.json).
+It describes a **1.79 GB**, two-stage FP16 Core ML pipeline for iOS 18 or newer.
+Core ML owns the handoff between the stages, which reduces the lifetime of
+large intermediate tensors without changing the U-Net calculation.
 
-The release gate is 35 dB. The three-style result validates the exact
-block-concatenated LoRA sum rather than a UI-only approximation.
+| Resource | Purpose | Size |
+|---|---|---:|
+| `pipeline-v1/UnetPipeline.mlmodelc` | Full nine-channel FP16 inpainting U-Net | 1.72 GB |
+| `VAEEncoder.mlmodelc` | Encodes the masked source image | 68.5 MB |
+| `manifest-pipeline.json` | Pinned paths, sizes, and SHA-256 checksums | — |
 
-Held-out 24-case inpainting quality relative to the previous Diffusers release:
+The separate Clover installation is still required. Its approximately
+**994.9 MB** runtime provides the tokenizer, text encoder, and VAE decoder, so
+this repository does not download duplicate copies of those components.
 
-| Metric | Previous | HQ |
-|---|---:|---:|
-| Masked prompt CLIP similarity | 0.2642 | **0.2768** |
-| Masked target MAE | 0.2510 | **0.2231** |
-| Changed pixels outside the mask | 0 | **0** |
+The repository also retains compressed and stateful research artifacts for
+comparison. The shipping iOS app uses `manifest-pipeline.json`, not the older
+`manifest.json` entry point.
 
 ## Runtime contract
 
 - Minimum OS: iOS 18
-- Resolution: 512×512
-- U-Net input: noisy latent (4) + mask (1) + masked image latent (4)
-- Mask: white regenerates, black preserves
-- Maximum simultaneous styles: 3
-- Safety checker: not constructed by the Clover iOS pipeline
+- Resolution: 512 × 512
+- U-Net sample input: Float16 `[1, 9, 64, 64]`
+- Timestep input: Float16 `[1]`
+- Text input: Float16 `[1, 768, 1, 77]`
+- Noise prediction output: Float32 `[1, 4, 64, 64]`
+- Channels: noisy latent (4) + mask (1) + masked-image latent (4)
+- Mask semantics: white regenerates, black preserves
+- Recommended scheduler: DPM-Solver++ multistep
+- Recommended settings: 20 steps, guidance scale 6.0
+
+The selected source region is replaced by neutral gray before VAE encoding,
+which becomes zero after normalization to `[-1, 1]`. For a small mask, the iOS
+runtime uses a focused 512 × 512 crop with surrounding source context and then
+composites the result through the exact user mask. Pixels outside the mask are
+copied directly from the source image.
+
+## LoRA compatibility
+
+LoRAs made for the regular Clover Create model target its four-channel U-Net
+and cannot be applied directly to this nine-channel inpainting U-Net. To ship
+an inpainting style, train or adapt it for the inpainting checkpoint and fuse
+it before Core ML conversion.
+
+## Validation
+
+The production bundle is checked in three ways:
+
+- every file is verified against the byte count and SHA-256 in the schema-v3 manifest;
+- the compiled pipeline passes a native macOS Core ML prediction smoke test with finite output and the expected tensor shape;
+- the complete app flow was exercised on a physical iPhone 15 running iOS 26.6 using a 30-step masked edit.
+
+The physical-device run completed successfully in about 73 seconds in the
+XCTest/debug validation environment. That timing is a release smoke test, not
+a formal performance benchmark; first-run Core ML compilation and device
+thermal state can materially change latency.
 
 ## Citation
 
 ```bibtex
 @software{lozadaperez2026cloverimagetinyinpaintcoreml,
   author = {Lukas Lozada Perez},
-  title = {Clover Image Tiny Inpaint HQ Core ML},
+  title = {Clover Image Tiny Inpaint Core ML},
   year = {2026},
   url = {https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint-CoreML}
 }
 ```
 
-Designed and developed independently by Lukas Lozada Perez. Open weights under
-the CreativeML Open RAIL-M license; inference runs completely on device.
+Designed and developed independently by Lukas Lozada Perez. Model weights are
+available under the CreativeML Open RAIL-M license. Inference runs completely
+on device after installation.