Clover Image Tiny Inpaint HQ

SD 1.5-class inpainting for local, prompt-guided image edits.
Image + mask + prompt · Clover's shared components · Diffusers and separate Core ML resources.

Example · Model · Run it · Evaluation & documentation

Inpainting example

The white area of a mask selects what to regenerate; black selects what to preserve. This published example uses the prompt “add blue sunglasses.”

Original image	Masked edit

Model

Inpaint HQ uses the full Stable Diffusion 1.5 inpainting denoiser with Clover's text encoder, tokenizer, VAE, and scheduler. It is a larger model than the compact regular Clover text-to-image checkpoint. Regular Clover generates from text; Inpaint HQ requires an image and mask.

On the release's 24-edit held-out evaluation, HQ improved masked prompt alignment by 4.8% and reduced masked target error by 11.1% versus the previous Clover inpainting release. These are evaluation-set results; individual edits still depend on the image, mask, and prompt. See the quality gate and selection record below.

Run an edit with Python

Install the project's pinned Python dependencies. This example uses an NVIDIA CUDA GPU:

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")

Start with 20 steps, guidance 6.0, and 96 pixels of surrounding context. Use a source image and same-size grayscale mask. Small selections benefit from a context crop so the model can see both the edited region and its surroundings. To use an already downloaded model offline, pass local_files_only=True to from_pretrained.

Apple silicon and CPU

For Apple silicon, change .to("cuda") to .to("mps"). For CPU, use torch_dtype=torch.float32 and .to("cpu"); generation is slower. Hardware, precision, and dependencies can change the result even with the same seed.

Styles and Apple devices

The Clover iOS app linked above provides a native masking workflow. Core ML resources are published separately.

Published Clover attention-only LoRAs are mechanically compatible in Diffusers, but they were trained for text-to-image, so evaluate each style's editing quality. The shipping iOS inpainting screen does not enable dynamic styles. Adapter and deployment details are included in Evaluation and provenance below.

Limits and license

Small text, hands, faces, exact logos, and very small masks can be difficult. Review generated edits before sharing. The model inherits the limitations and CreativeML Open RAIL-M obligations of Clover and Stable Diffusion 1.5 inpainting.

Evaluation and provenance

Quality gate, selection record, Core ML, and LoRA compatibility

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	Change vs. previous Clover	SD 1.5 inpaint teacher
Masked prompt CLIP similarity ↑	0.2642	0.2768	4.8% higher	0.2820
Masked target MAE ↓	0.2510	0.2231	11.1% lower error	0.2156
Changed pixels outside the mask ↓	0	0	Unchanged	0

↑ Higher is better; ↓ lower is better. Bold highlights HQ's improvements over the previous Clover release, not the best score across all models. The strongest measured improvement is 11.1% lower masked target error. The SD 1.5 teacher still scores better on both alignment and target error in this evaluation.

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.

In Diffusers, the published Clover style LoRAs are mechanically compatible with this nine-channel U-Net because they modify attention projections only; those tensor shapes are unchanged from the four-channel model. They were trained for text-to-image, however, so masked-edit quality should be evaluated per style. LoRAs that modify the four-channel input convolution are not compatible.

Dynamic Core ML loading is a separate deployment capability. The shipping iOS inpainting screen currently selects a stateless or chunked U-Net and does not enable styles. An adapter-aware stateful nine-channel export can provide up to three independently weighted slots, but it must include the matching adapter schema and be validated on its target devices. For a stateless deployment, fuse a compatible LoRA before conversion.

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

@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}
}

Created by Lukas Lozada Perez.