File size: 3,978 Bytes
be06992 8158957 be06992 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 | # Clover Inpaint HQ — technical reference
[Back to the model card](https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint).
## 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`](https://huggingface.co/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
```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.
|