Release validated HQ inpainting pipelines

coreml-tools/apple-no-mid-block.patch

@@ -0,0 +1,54 @@
+diff --git a/python_coreml_stable_diffusion/unet.py b/python_coreml_stable_diffusion/unet.py
+index 666f146..519bc6f 100644
+--- a/python_coreml_stable_diffusion/unet.py
++++ b/python_coreml_stable_diffusion/unet.py
+@@ -916,19 +916,22 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
+             self.down_blocks.append(down_block)
+ 
+         # mid
+-        assert mid_block_type == "UNetMidBlock2DCrossAttn"
+-        self.mid_block = UNetMidBlock2DCrossAttn(
+-            in_channels=block_out_channels[-1],
+-            transformer_layers_per_block=transformer_layers_per_block[-1],
+-            temb_channels=time_embed_dim,
+-            resnet_eps=norm_eps,
+-            resnet_act_fn=act_fn,
+-            output_scale_factor=mid_block_scale_factor,
+-            resnet_time_scale_shift="default",
+-            cross_attention_dim=cross_attention_dim,
+-            attn_num_head_channels=attention_head_dim[i],
+-            resnet_groups=norm_num_groups,
+-        )
++        if mid_block_type is None:
++            self.mid_block = None
++        else:
++            assert mid_block_type == "UNetMidBlock2DCrossAttn"
++            self.mid_block = UNetMidBlock2DCrossAttn(
++                in_channels=block_out_channels[-1],
++                transformer_layers_per_block=transformer_layers_per_block[-1],
++                temb_channels=time_embed_dim,
++                resnet_eps=norm_eps,
++                resnet_act_fn=act_fn,
++                output_scale_factor=mid_block_scale_factor,
++                resnet_time_scale_shift="default",
++                cross_attention_dim=cross_attention_dim,
++                attn_num_head_channels=attention_head_dim[i],
++                resnet_groups=norm_num_groups,
++            )
+ 
+         # up
+         reversed_block_out_channels = list(reversed(block_out_channels))
+@@ -1014,9 +1017,10 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin):
+             down_block_res_samples = new_down_block_res_samples
+ 
+         # 4. mid
+-        sample = self.mid_block(sample,
+-                                emb,
+-                                encoder_hidden_states=encoder_hidden_states)
++        if self.mid_block is not None:
++            sample = self.mid_block(sample,
++                                    emb,
++                                    encoder_hidden_states=encoder_hidden_states)
+ 
+         if self.support_controlnet:
+             sample = sample + additional_residuals[-1]

coreml-tools/build_hq_manifest.py

@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+"""Build the pinned HQ inpainting download manifest from release artifacts."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+COMPILED = (
+    ROOT
+    / "coreml-inpaint-hq-quant"
+    / "compiled"
+    / "Stable_Diffusion_version_clover-image-tiny-inpaint-hq_unet_int8.mlmodelc"
+)
+SCHEMA = ROOT / "coreml-inpaint-hq-quant" / "adapter-schema.json"
+OLD_MANIFEST = Path("/tmp/inpaint-coreml-manifest.json")
+RELEASE_PATH = "hq-v4-int8"
+OUTPUT = ROOT / "release" / RELEASE_PATH / "manifest.json"
+
+
+def record(path: Path, *, local_path: str, remote_path: str) -> dict[str, object]:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        while chunk := handle.read(1024 * 1024):
+            digest.update(chunk)
+    return {
+        "path": local_path,
+        "remote_path": remote_path,
+        "size": path.stat().st_size,
+        "sha256": digest.hexdigest(),
+    }
+
+
+def main() -> None:
+    previous = json.loads(OLD_MANIFEST.read_text())
+    resources = [
+        item
+        for item in previous["resources"]
+        if item["path"].startswith("VAEEncoder.mlmodelc/")
+    ]
+    for path in sorted(COMPILED.rglob("*")):
+        if path.is_file():
+            relative = path.relative_to(COMPILED).as_posix()
+            resources.append(
+                record(
+                    path,
+                    local_path=f"Unet.mlmodelc/{relative}",
+                    remote_path=f"{RELEASE_PATH}/Unet.mlmodelc/{relative}",
+                )
+            )
+    resources.append(
+        record(
+            SCHEMA,
+            local_path="adapter-schema.json",
+            remote_path=f"{RELEASE_PATH}/adapter-schema.json",
+        )
+    )
+    manifest = {
+        "schema_version": 2,
+        "model": "neonforestmist/Clover-Image-Tiny-Inpaint",
+        "base_model": "neonforestmist/Clover-Image-Tiny",
+        "minimum_ios": "18.0",
+        "requires_base_model": True,
+        "resolution": [512, 512],
+        "max_adapter_count": 3,
+        "weight_compression": "per-channel symmetric int8",
+        "unet_input": {
+            "shape": [1, 9, 64, 64],
+            "channels": ["noisy_latent", "mask", "masked_image_latent"],
+            "mask_semantics": "white=regenerate, black=preserve",
+        },
+        "recommended_inference": {
+            "scheduler": "dpm_solver_multistep",
+            "steps": 20,
+            "guidance_scale": 6.0,
+            "mask_crop_padding": 96,
+            "composite_outside_mask": "exact source pixels",
+        },
+        "validation": {
+            "samples": 24,
+            "masked_clip_similarity": 0.27676651254296303,
+            "masked_target_mae": 0.2230850402265787,
+            "fp16_base_psnr_db": 78.55727510619732,
+            "fp16_three_style_psnr_db": 78.4737949968816,
+            "int8_base_psnr_db": 59.3153859192934,
+            "int8_three_style_psnr_db": 59.30338086829437,
+        },
+        "resources": sorted(resources, key=lambda item: item["path"]),
+    }
+    OUTPUT.parent.mkdir(parents=True, exist_ok=True)
+    OUTPUT.write_text(json.dumps(manifest, indent=2) + "\n")
+    print(OUTPUT)
+    print(sum(int(item["size"]) for item in resources))
+
+
+if __name__ == "__main__":
+    main()

coreml-tools/constraints.txt

@@ -0,0 +1,5 @@
+coremltools==9.0
+numpy==1.23.5
+scikit-learn==1.5.1
+scipy==1.11.4
+torch==2.7.0

coreml-tools/convert_stateful_multi_lora_unet.py

@@ -18,6 +18,7 @@ import argparse
 import gc
 import json
 import os
+import re
 import time
 from pathlib import Path
 
@@ -89,21 +90,49 @@ def inject_stateful_lora(
     template: dict[str, torch.Tensor],
     scale: float,
     slot_count: int,
-) -> None:
+) -> dict[str, str]:
     target_names = sorted(
         {
             key.removeprefix("unet.").split(".lora.")[0]
             for key in template
         }
     )
-    for target_name in target_names:
+    target_mapping: dict[str, str] = {}
+    for source_target_name in target_names:
+        down = template[source_key(source_target_name, "lora_down")]
+        up = template[source_key(source_target_name, "lora_up")]
+        candidates = [source_target_name]
+        up_match = re.match(r"up_blocks\.(\d+)\.(.+)", source_target_name)
+        if up_match:
+            # The HQ inpainting U-Net has an additional deepest UpBlock2D at
+            # index zero. Clover's three cross-attention up blocks therefore
+            # align with HQ indices one through three.
+            candidates.append(
+                f"up_blocks.{int(up_match.group(1)) + 1}.{up_match.group(2)}"
+            )
+        target_name = ""
+        base: torch.nn.Conv2d | None = None
+        for candidate in candidates:
+            try:
+                module = model.get_submodule(candidate)
+            except AttributeError:
+                continue
+            if not isinstance(module, torch.nn.Conv2d):
+                continue
+            if (
+                down.shape[1] == module.in_channels
+                and up.shape[0] == module.out_channels
+            ):
+                target_name = candidate
+                base = module
+                break
+        if base is None:
+            raise AttributeError(
+                f"No channel-compatible HQ projection for {source_target_name}; "
+                f"tried {candidates}"
+            )
         parent_name, child_name = target_name.rsplit(".", 1)
         parent = model.get_submodule(parent_name)
-        base = model.get_submodule(target_name)
-        if not isinstance(base, torch.nn.Conv2d):
-            raise TypeError(f"Expected Conv2d for {target_name}, found {type(base)}")
-        down = template[source_key(target_name, "lora_down")]
-        up = template[source_key(target_name, "lora_up")]
         setattr(
             parent,
             child_name,
@@ -115,8 +144,12 @@ def inject_stateful_lora(
                 slot_count=slot_count,
             ),
         )
+        target_mapping[target_name] = source_target_name
     if len(target_names) != 72:
         raise RuntimeError(f"Expected 72 Clover LoRA targets, found {len(target_names)}")
+    if len(target_mapping) != len(target_names):
+        raise RuntimeError("LoRA target mapping contains duplicate HQ targets")
+    return target_mapping
 
 
 def main() -> None:
@@ -134,15 +167,19 @@ def main() -> None:
         raise ValueError("--max-adapter-count must be positive")
     template = load_file(str(template_path))
     original_class = coreml_unet.UNet2DConditionModel
+    target_source_names: dict[str, str] = {}
 
     class StatefulMultiLoRAUNet2DConditionModel(original_class):
         def load_state_dict(self, state_dict, *args, **kwargs):
             result = super().load_state_dict(state_dict, *args, **kwargs)
-            inject_stateful_lora(
-                self,
-                template=template,
-                scale=wrapper_args.lora_scale,
-                slot_count=wrapper_args.max_adapter_count,
+            target_source_names.clear()
+            target_source_names.update(
+                inject_stateful_lora(
+                    self,
+                    template=template,
+                    scale=wrapper_args.lora_scale,
+                    slot_count=wrapper_args.max_adapter_count,
+                )
             )
             return result
 
@@ -210,7 +247,7 @@ def main() -> None:
             )
         for (torch_name, state_value), exported in zip(buffers, exported_states):
             target_name, state_suffix = torch_name.rsplit(".", 1)
-            key = source_key(target_name, state_suffix)
+            key = source_key(target_source_names[target_name], state_suffix)
             source_value = template[key]
             source_shape = list(source_value.shape)
             source_shape.extend([1] * (state_value.ndim - source_value.ndim))

coreml-tools/validate_stateful_multi_lora.py

@@ -6,6 +6,7 @@ from __future__ import annotations
 import argparse
 import json
 import math
+import re
 from pathlib import Path
 
 import coremltools as ct
@@ -54,10 +55,34 @@ def inject_loras(model, adapters, scales):
         }
     )
     for target_name in target_names:
-        parent_name, child_name = target_name.rsplit(".", 1)
-        parent = model.get_submodule(parent_name)
-        base = model.get_submodule(target_name)
         prefix = f"unet.{target_name}.lora"
+        down_shape = states[0][f"{prefix}.down.weight"].shape
+        up_shape = states[0][f"{prefix}.up.weight"].shape
+        candidates = [target_name]
+        up_match = re.match(r"up_blocks\.(\d+)\.(.+)", target_name)
+        if up_match:
+            candidates.append(
+                f"up_blocks.{int(up_match.group(1)) + 1}.{up_match.group(2)}"
+            )
+        resolved_name = ""
+        base = None
+        for candidate in candidates:
+            try:
+                module = model.get_submodule(candidate)
+            except AttributeError:
+                continue
+            if (
+                isinstance(module, torch.nn.Conv2d)
+                and down_shape[1] == module.in_channels
+                and up_shape[0] == module.out_channels
+            ):
+                resolved_name = candidate
+                base = module
+                break
+        if base is None:
+            raise AttributeError(f"No compatible projection for {target_name}")
+        parent_name, child_name = resolved_name.rsplit(".", 1)
+        parent = model.get_submodule(parent_name)
         components = [
             (
                 state[f"{prefix}.down.weight"],
@@ -116,6 +141,9 @@ def main():
     pipeline = DiffusionPipeline.from_pretrained(
         args.model_version,
         local_files_only=True,
+        torch_dtype=torch.float16,
+        variant="fp16",
+        use_safetensors=True,
     )
     reference = coreml_unet.UNet2DConditionModel(**pipeline.unet.config).eval()
     reference.load_state_dict(pipeline.unet.state_dict())

inpainting/evaluate_benchmark.py

@@ -45,6 +45,46 @@ def _model_specs(values: list[str]) -> list[tuple[str, str]]:
 
 
 def _load_pipeline(model_path: str, device: torch.device):
+    if model_path.startswith("blend@"):
+        _, alpha_text, candidate_path = model_path.split("@", 2)
+        alpha = float(alpha_text)
+        if not 0.0 <= alpha <= 1.0:
+            raise ValueError("blend alpha must be in [0, 1]")
+        pipeline = StableDiffusionInpaintPipeline.from_pretrained(
+            "neonforestmist/Clover-Image-Tiny-Inpaint",
+            torch_dtype=torch.float16,
+            safety_checker=None,
+            requires_safety_checker=False,
+        )
+        candidate = load_file(candidate_path)
+        blended = pipeline.unet.state_dict()
+        for name, value in blended.items():
+            if name not in candidate:
+                raise KeyError(f"Candidate is missing U-Net tensor {name}")
+            update = candidate[name].to(dtype=value.dtype)
+            blended[name] = value.lerp(update, alpha)
+        pipeline.unet.load_state_dict(blended, strict=True)
+        return pipeline.to(device)
+    if model_path.startswith("clover-hybrid::"):
+        teacher_path = model_path.removeprefix("clover-hybrid::")
+        pipeline = StableDiffusionInpaintPipeline.from_pretrained(
+            teacher_path,
+            torch_dtype=torch.float16,
+            safety_checker=None,
+            requires_safety_checker=False,
+        )
+        clover = StableDiffusionInpaintPipeline.from_pretrained(
+            "neonforestmist/Clover-Image-Tiny-Inpaint",
+            torch_dtype=torch.float16,
+            safety_checker=None,
+            requires_safety_checker=False,
+        )
+        pipeline.tokenizer = clover.tokenizer
+        pipeline.text_encoder = clover.text_encoder
+        pipeline.vae = clover.vae
+        pipeline.scheduler = clover.scheduler
+        del clover
+        return pipeline.to(device)
     pipeline_path, separator, unet_state_path = model_path.partition("::")
     pipeline = StableDiffusionInpaintPipeline.from_pretrained(
         pipeline_path,

modal_inpaint.py

@@ -48,7 +48,12 @@ app = modal.App(
 )
 
 
-@app.function(gpu="A10", timeout=24 * 60 * 60, cpu=8, memory=32768)
+@app.function(
+    gpu=os.environ.get("CLOVER_MODAL_GPU", "A10"),
+    timeout=24 * 60 * 60,
+    cpu=8,
+    memory=32768,
+)
 def train(
     *,
     base_model: str = "neonforestmist/Clover-Image-Tiny",
@@ -72,6 +77,7 @@ def train(
     lr_warmup_steps: int | None = None,
     train_batch_size: int = 1,
     gradient_accumulation_steps: int = 4,
+    gradient_checkpointing: bool = True,
     lora_rank: int = 0,
     teacher_loss_weight: float = 0.75,
     ground_truth_loss_weight: float = 0.25,
@@ -136,7 +142,6 @@ def train(
         str(gradient_accumulation_steps),
         "--train_batch_size",
         str(train_batch_size),
-        "--gradient_checkpointing",
         "--mixed_precision",
         "bf16",
         "--random_flip",
@@ -165,6 +170,8 @@ def train(
         "--output_dir",
         str(output_dir),
     ]
+    if gradient_checkpointing:
+        command.append("--gradient_checkpointing")
     if initial_inpaint_revision:
         command.extend(["--initial_inpaint_revision", initial_inpaint_revision])
     if lora_rank > 0:
@@ -221,6 +228,7 @@ def main(
     lr_warmup_steps: int | None = None,
     train_batch_size: int = 1,
     gradient_accumulation_steps: int = 4,
+    gradient_checkpointing: bool = True,
     object_fraction: float = 0.70,
     lora_rank: int = 0,
     teacher_loss_weight: float = 0.75,
@@ -230,7 +238,7 @@ def main(
     boundary_loss_weight: float = 2.0,
     resume: bool = False,
 ) -> None:
-    """Launch a smoke test or the credit-compatible bounded A10 v3 recipe."""
+    """Launch a smoke test or a bounded v3 recipe on the selected Modal GPU."""
 
     if smoke:
         steps = min(steps, 4)
@@ -250,6 +258,7 @@ def main(
         lr_warmup_steps=lr_warmup_steps,
         train_batch_size=train_batch_size,
         gradient_accumulation_steps=gradient_accumulation_steps,
+        gradient_checkpointing=gradient_checkpointing,
         object_fraction=object_fraction,
         lora_rank=lora_rank,
         teacher_loss_weight=teacher_loss_weight,

modal_inpaint_benchmark.py

@@ -43,7 +43,12 @@ app = modal.App(
 )
 
 
-@app.function(gpu="A10", timeout=4 * 60 * 60, cpu=8, memory=32768)
+@app.function(
+    gpu=os.environ.get("CLOVER_MODAL_GPU", "A10"),
+    timeout=4 * 60 * 60,
+    cpu=8,
+    memory=32768,
+)
 def benchmark(*, models: list[str], output_name: str, samples: int = 24) -> str:
     output_dir = OUTPUT_ROOT / output_name
     if output_dir.exists():

release/hq-v3/README-coreml.md

@@ -0,0 +1,67 @@
+---
+library_name: coreml
+pipeline_tag: image-to-image
+base_model: neonforestmist/Clover-Image-Tiny-Inpaint
+license: creativeml-openrail-m
+tags:
+  - coreml
+  - ios
+  - inpainting
+  - lora
+  - stable-diffusion
+---
+
+# Clover Image Tiny Inpaint HQ — 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 HQ U-Net is under `hq-v3/Unet.mlmodelc`. Shared Clover resources remain at
+the repository root so app upgrades only download files whose checksums
+changed. `hq-v3/adapter-schema.json` maps each small `.safetensors` style into
+one of three independent state slots.
+
+## Validation
+
+PyTorch/Core ML parity on deterministic inputs:
+
+| Configuration | PSNR |
+|---|---:|
+| Base HQ inpainting U-Net | **78.56 dB** |
+| Monet 0.70 + Pointillism 0.45 + Watercolor Anime 1.10 | **78.47 dB** |
+
+The release gate is 35 dB. The three-style result validates the exact
+block-concatenated LoRA sum rather than a UI-only approximation.
+
+Held-out 24-case inpainting quality relative to the previous release:
+
+| 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** |
+
+## 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
+
+## Citation
+
+```bibtex
+@software{lozadaperez2026cloverimagetinyinpaintcoreml,
+  author = {Lukas Lozada Perez},
+  title = {Clover Image Tiny Inpaint HQ 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.

release/hq-v3/README-diffusers.md

@@ -0,0 +1,121 @@
+---
+library_name: diffusers
+pipeline_tag: image-to-image
+base_model:
+  - neonforestmist/Clover-Image-Tiny
+  - stable-diffusion-v1-5/stable-diffusion-inpainting
+license: creativeml-openrail-m
+tags:
+  - clover-image
+  - inpainting
+  - stable-diffusion
+  - coreml
+---
+
+# Clover Image Tiny Inpaint HQ
+
+Clover Image Tiny Inpaint HQ is the high-quality, context-aware inpainting
+pipeline for Clover. It combines the complete Stable Diffusion 1.5 inpainting
+U-Net with Clover Image Tiny's tokenizer, text encoder, VAE, and scheduler.
+This preserves Clover compatibility while replacing the compact inpainting
+denoiser that frequently produced blurry or unrecognizable masked objects.
+
+The pipeline uses the standard nine-channel inpainting contract:
+
+```text
+[noisy latent (4), mask (1), masked-image latent (4)]
+```
+
+## Diffusers example
+
+```python
+import torch
+from diffusers import AutoPipelineForInpainting, DPMSolverMultistepScheduler
+from diffusers.utils import load_image
+
+pipe = AutoPipelineForInpainting.from_pretrained(
+    "neonforestmist/Clover-Image-Tiny-Inpaint",
+    torch_dtype=torch.float16,
+).to("cuda")
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+
+image = pipe(
+    prompt="a glossy red enamel kettle resting naturally on the countertop",
+    negative_prompt="blurry, distorted, low detail",
+    image=load_image("source.png"),
+    mask_image=load_image("mask.png"),
+    num_inference_steps=20,
+    guidance_scale=6.0,
+    padding_mask_crop=96,
+).images[0]
+image.save("clover-inpaint.png")
+```
+
+Recommended interactive defaults are DPM-Solver++, 20 steps, CFG 6.0, and a
+96-pixel context crop. Composite the generated result through the exact binary
+mask when unchanged source pixels must remain byte-for-byte untouched.
+
+## Quality gate
+
+The release was evaluated on 24 deterministic, held-out, human-rated
+InpaintCOCO edits. Every output was also reviewed in three visual contact
+sheets before release.
+
+| Metric | Previous Clover inpaint | HQ release | SD 1.5 inpaint teacher |
+|---|---:|---:|---:|
+| Masked prompt CLIP similarity (higher) | 0.2642 | **0.2768** | 0.2820 |
+| Masked target MAE (lower) | 0.2510 | **0.2231** | 0.2156 |
+| Changed pixels outside the mask | 0 | **0** | 0 |
+
+The HQ release improves prompt alignment by 4.8% and reduces masked target
+error by 11.1% relative to the previous Clover inpainting release. The visual
+gate showed recognizable buses, dogs, trains, furniture, signs, and
+scene-consistent lighting where the compact candidates often collapsed into
+amorphous fills.
+
+## Selection provenance
+
+The release process compared the existing checkpoint, a 30,000-step full-U-Net
+distillation run, two fused context-LoRA refinements, partial weight blends,
+the full Stable Diffusion inpainting reference, and this Clover-component
+hybrid. The 30,000-step and context-LoRA candidates were rejected because they
+did not beat the existing release across both visual and quantitative gates.
+The published HQ architecture was the only Clover-compatible candidate that
+materially improved both prompt alignment and reconstruction.
+
+- Inpainting U-Net revision:
+  `stable-diffusion-v1-5/stable-diffusion-inpainting@8a4288a76071f7280aedbdb3253bdb9e9d5d84bb`
+- Clover components: `neonforestmist/Clover-Image-Tiny`
+- Evaluation dataset: `phiyodr/InpaintCOCO@1ffac84be2dfc5ad9afccad868522fad64457435`
+- Selection platform: Modal H100
+- Evaluation seed: `20260813`
+
+## Core ML and style mixing
+
+The companion iOS resources are published at
+[`neonforestmist/Clover-Image-Tiny-Inpaint-CoreML`](https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint-CoreML).
+Its batch-one stateful U-Net supports up to three Clover styles simultaneously
+with independent strengths. The style tensors remain separate downloads and
+are composed exactly at runtime; they are not fused into three full 1.6 GB
+models.
+
+## Limitations
+
+Small text, hands, faces, exact logos, and masks below latent resolution can
+still fail. Output quality depends on the source, mask, prompt, scheduler,
+guidance, seed, and step count. This release inherits the limitations and
+license obligations of Clover Image Tiny and Stable Diffusion 1.5 inpainting.
+
+## Citation
+
+```bibtex
+@software{lozadaperez2026cloverimagetinyinpaint,
+  author = {Lukas Lozada Perez},
+  title = {Clover Image Tiny Inpaint HQ: Local Context-Aware Image Inpainting},
+  year = {2026},
+  url = {https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint}
+}
+```
+
+Designed and developed independently by Lukas Lozada Perez. Open weights under
+the CreativeML Open RAIL-M license; complete local inference is supported.

release/hq-v3/manifest.json

@@ -0,0 +1,103 @@
+{
+  "schema_version": 2,
+  "model": "neonforestmist/Clover-Image-Tiny-Inpaint",
+  "base_model": "neonforestmist/Clover-Image-Tiny",
+  "minimum_ios": "18.0",
+  "requires_base_model": true,
+  "resolution": [
+    512,
+    512
+  ],
+  "max_adapter_count": 3,
+  "unet_input": {
+    "shape": [
+      1,
+      9,
+      64,
+      64
+    ],
+    "channels": [
+      "noisy_latent",
+      "mask",
+      "masked_image_latent"
+    ],
+    "mask_semantics": "white=regenerate, black=preserve"
+  },
+  "recommended_inference": {
+    "scheduler": "dpm_solver_multistep",
+    "steps": 20,
+    "guidance_scale": 6.0,
+    "mask_crop_padding": 96,
+    "composite_outside_mask": "exact source pixels"
+  },
+  "validation": {
+    "samples": 24,
+    "masked_clip_similarity": 0.27676651254296303,
+    "masked_target_mae": 0.2230850402265787,
+    "base_psnr_db": 78.55727510619732,
+    "three_style_psnr_db": 78.4737949968816
+  },
+  "resources": [
+    {
+      "path": "Unet.mlmodelc/analytics/coremldata.bin",
+      "remote_path": "hq-v3/Unet.mlmodelc/analytics/coremldata.bin",
+      "size": 243,
+      "sha256": "0776280c592bd31e478eecd0cac5918984d48e629d1ffa5a9d6432a698448c9e"
+    },
+    {
+      "path": "Unet.mlmodelc/coremldata.bin",
+      "remote_path": "hq-v3/Unet.mlmodelc/coremldata.bin",
+      "size": 14009,
+      "sha256": "4cb24ed826e3e090138e99e3eb334e706d6b8bf733771fe034355268456e6834"
+    },
+    {
+      "path": "Unet.mlmodelc/metadata.json",
+      "remote_path": "hq-v3/Unet.mlmodelc/metadata.json",
+      "size": 50802,
+      "sha256": "54a5d730ce8790f028faacfe44cf9b555fb3109153cb5ffd8016ff5deb0375d8"
+    },
+    {
+      "path": "Unet.mlmodelc/model.mil",
+      "remote_path": "hq-v3/Unet.mlmodelc/model.mil",
+      "size": 1206218,
+      "sha256": "2f7238a5825ae16c9dc054536d9a5dd7cf0963af2423979b4f84ae9121c6606a"
+    },
+    {
+      "path": "Unet.mlmodelc/weights/weight.bin",
+      "remote_path": "hq-v3/Unet.mlmodelc/weights/weight.bin",
+      "size": 1719146496,
+      "sha256": "0f6a88bcd1ae4f36c076d46d404373c2f5c44bf85975ed2737b23643b9040826"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/analytics/coremldata.bin",
+      "size": 243,
+      "sha256": "ba122c04ae4c8d28c8b8805049e22fb7dfeac3eada7bde68c2366a14e712d1d6"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/coremldata.bin",
+      "size": 883,
+      "sha256": "fe1d98a943015c7ac3d29ef2d573b37610e62c97e5724d2f85c555960832629d"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/metadata.json",
+      "size": 2674,
+      "sha256": "f18ce0aabcf1ad6695c8d7bf418dd30bdac1d031ef453fe1f2b655487dff019d"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/model.mil",
+      "size": 132453,
+      "sha256": "52c0b22663abb7890a76932f8c8b2467908bfc5b6cdef92e4a2cbcde9a7d07a4"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/weights/weight.bin",
+      "size": 68338112,
+      "sha256": "71bf826e5ee1b455462aff0b34e3a469f5f58955ae715708962b1d81db92dc64"
+    },
+    {
+      "path": "adapter-schema.json",
+      "remote_path": "hq-v3/adapter-schema.json",
+      "size": 55773,
+      "sha256": "b356bae467a4e60cb26151c5b2cad2d2dc298b057f44b71479096036ca53c1d9"
+    }
+  ]
+}

release/hq-v4-int8/README-coreml.md

@@ -0,0 +1,70 @@
+---
+library_name: coreml
+pipeline_tag: image-to-image
+base_model: neonforestmist/Clover-Image-Tiny-Inpaint
+license: creativeml-openrail-m
+tags:
+  - coreml
+  - ios
+  - inpainting
+  - lora
+  - stable-diffusion
+---
+
+# Clover Image Tiny Inpaint HQ — 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 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.
+
+## Validation
+
+PyTorch/Core ML parity on deterministic inputs:
+
+| 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 release gate is 35 dB. The three-style result validates the exact
+block-concatenated LoRA sum rather than a UI-only approximation.
+
+Held-out 24-case inpainting quality relative to the previous Diffusers release:
+
+| 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** |
+
+## 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
+
+## Citation
+
+```bibtex
+@software{lozadaperez2026cloverimagetinyinpaintcoreml,
+  author = {Lukas Lozada Perez},
+  title = {Clover Image Tiny Inpaint HQ 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.

release/hq-v4-int8/manifest.json

@@ -0,0 +1,106 @@
+{
+  "schema_version": 2,
+  "model": "neonforestmist/Clover-Image-Tiny-Inpaint",
+  "base_model": "neonforestmist/Clover-Image-Tiny",
+  "minimum_ios": "18.0",
+  "requires_base_model": true,
+  "resolution": [
+    512,
+    512
+  ],
+  "max_adapter_count": 3,
+  "weight_compression": "per-channel symmetric int8",
+  "unet_input": {
+    "shape": [
+      1,
+      9,
+      64,
+      64
+    ],
+    "channels": [
+      "noisy_latent",
+      "mask",
+      "masked_image_latent"
+    ],
+    "mask_semantics": "white=regenerate, black=preserve"
+  },
+  "recommended_inference": {
+    "scheduler": "dpm_solver_multistep",
+    "steps": 20,
+    "guidance_scale": 6.0,
+    "mask_crop_padding": 96,
+    "composite_outside_mask": "exact source pixels"
+  },
+  "validation": {
+    "samples": 24,
+    "masked_clip_similarity": 0.27676651254296303,
+    "masked_target_mae": 0.2230850402265787,
+    "fp16_base_psnr_db": 78.55727510619732,
+    "fp16_three_style_psnr_db": 78.4737949968816,
+    "int8_base_psnr_db": 59.3153859192934,
+    "int8_three_style_psnr_db": 59.30338086829437
+  },
+  "resources": [
+    {
+      "path": "Unet.mlmodelc/analytics/coremldata.bin",
+      "remote_path": "hq-v4-int8/Unet.mlmodelc/analytics/coremldata.bin",
+      "size": 243,
+      "sha256": "fe4fd138e2daa944e7ce8aabdcc7fd371bea3f0fbd3add90c4534af74645c641"
+    },
+    {
+      "path": "Unet.mlmodelc/coremldata.bin",
+      "remote_path": "hq-v4-int8/Unet.mlmodelc/coremldata.bin",
+      "size": 14009,
+      "sha256": "cb75c645b04342e3c99a57577b88a74a006bde0bbf7ae758174868083717856c"
+    },
+    {
+      "path": "Unet.mlmodelc/metadata.json",
+      "remote_path": "hq-v4-int8/Unet.mlmodelc/metadata.json",
+      "size": 50871,
+      "sha256": "6485785471f213843cdb8133e07a0e76d1e6dc9667e59047ac4a67c75db899ea"
+    },
+    {
+      "path": "Unet.mlmodelc/model.mil",
+      "remote_path": "hq-v4-int8/Unet.mlmodelc/model.mil",
+      "size": 1256410,
+      "sha256": "f47ba899b3b7b47dcea95109d4ab8f59372b6cf24df25f83b094dd770d60afb2"
+    },
+    {
+      "path": "Unet.mlmodelc/weights/weight.bin",
+      "remote_path": "hq-v4-int8/Unet.mlmodelc/weights/weight.bin",
+      "size": 860709760,
+      "sha256": "142f9f803d75ea2379ca3793c7d03b5da9eb29ac59fbae1a6dc5fe38b864e8a5"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/analytics/coremldata.bin",
+      "size": 243,
+      "sha256": "ba122c04ae4c8d28c8b8805049e22fb7dfeac3eada7bde68c2366a14e712d1d6"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/coremldata.bin",
+      "size": 883,
+      "sha256": "fe1d98a943015c7ac3d29ef2d573b37610e62c97e5724d2f85c555960832629d"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/metadata.json",
+      "size": 2674,
+      "sha256": "f18ce0aabcf1ad6695c8d7bf418dd30bdac1d031ef453fe1f2b655487dff019d"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/model.mil",
+      "size": 132453,
+      "sha256": "52c0b22663abb7890a76932f8c8b2467908bfc5b6cdef92e4a2cbcde9a7d07a4"
+    },
+    {
+      "path": "VAEEncoder.mlmodelc/weights/weight.bin",
+      "size": 68338112,
+      "sha256": "71bf826e5ee1b455462aff0b34e3a469f5f58955ae715708962b1d81db92dc64"
+    },
+    {
+      "path": "adapter-schema.json",
+      "remote_path": "hq-v4-int8/adapter-schema.json",
+      "size": 55773,
+      "sha256": "b356bae467a4e60cb26151c5b2cad2d2dc298b057f44b71479096036ca53c1d9"
+    }
+  ]
+}