Model guide / 17 September 2026
Every model,
in plain words.
What each LocalSR model does, what it is genuinely best at, and what to check before you download it or use its results for anything commercial. The face-restoration deep-dive with measured results is at the end.
Start here
A model is a large trained network file, not a program. LocalSR runs it on your computer, and your images are never sent anywhere. Each model is built for one job at one scale — that is why the catalog is a short list, not a wall of options.
Not sure which to pick? Start with the Quick Start preset: SPAN 4x NomosUni. It downloads in seconds, runs on modest hardware, and its results cover most photos. Move to a bigger model only when you can name what the quick one is missing.
Pick by what you want to fix:
- “Make it bigger and sharper” — upscaling models. Choose by scale (2×, 4×) and by how much quality you need.
- “It’s grainy / noisy” — denoise models. 1× models; they keep the size and remove noise.
- “It’s blurry, or an old JPEG” — deblur and JPEG restoration. Also 1×; chain them before an upscaler.
- “It’s anime or a drawing” — one dedicated model, faster and cleaner on line art than photo models.
- “Recover a face” — optional face models. Stronger terms apply; read the warning first.
- “Restore a whole video” — SeedVR2, Labs. Experimental, heavy, hours-to-days per clip.
Before you download
Read this once before downloading any model.
- Models are third-party files with their own terms. LocalSR’s application is MIT-licensed; the models are not. Each model’s license and commercial-use status is listed in the table below. Personal, private use is broadly fine; publishing or selling outputs may require more.
- Two models have unclear rights. The face-restoration checkpoints were fine-tuned for this project, but their training-data and weight rights are unresolved. LocalSR does not auto-download them. If you supply a copy, you are responsible for having the right to use it. No commercial-use claim is made for any face model.
- Results are reconstructions, not recovered truth. These networks invent plausible texture. Edges, text, eyes, and skin can look convincing and still be wrong. Keep your originals — LocalSR never overwrites them — and inspect results before you rely on or republish them.
- Download sizes are large. From 5 MB (SPAN) to 6.8 GB per SeedVR2 file. LocalSR downloads only from the pinned source, over HTTPS, and verifies the exact size and SHA-256 hash before a model is installed (see how verification works).
Making images bigger
| Model | Scale | Best for | Cost |
|---|---|---|---|
| SPAN 4x NomosUni Quick Start | 4× | The default first choice. Tiny download, very fast, good general photo quality. Also the model behind the website’s worked example. | Very low · ~5 MB · ~0.5 GB VRAM |
| RealPLKSR 4x NomosWebPhoto Best Quality preset | 4× | When the quick result visibly falls short: finer texture in photos, better fidelity on faces and fabric. Slower than SPAN, still moderate. | Medium · ~30 MB · ~1.2 GB VRAM |
| Real-ESRGAN ×2 | 2× | You genuinely want 2× — for a webpage, a print at modest enlargement, or a 4× model’s jump being too much. Native 2× avoids the downsample a 4× model needs. | Medium · ~67 MB · ~1.8 GB VRAM |
| HAT-S ×4 | 4× | A stronger, slower alternative for photos; the lightest of the HAT family and the base of the face-aware pair. | High · ~81 MB · ~2 GB VRAM |
| HAT-L ×4 ImageNet | 4× | Maximum quality stills on capable hardware — the largest official HAT variant. Roughly 3× slower than HAT-S; needs a strong GPU and patience. | Very high · ~166 MB · ~6 GB VRAM |
| Real-ESRGAN ×4 — Classic | 4× | The well-known real-world photo upscaler. Robust on photos with unknown blur, noise and compression; smoother than NomosWebPhoto on fine texture. | Medium · ~67 MB · ~1.8 GB VRAM |
| SwinIR ×4 Real — Transformer | 4× | A transformer trained on real-world degradation, between SPAN and HAT in cost. Sharp and strong on real photographs with mixed blur and compression. | High · ~67 MB · ~2.5 GB VRAM |
A practical path: SPAN first; if a specific image deserves more, try NomosWebPhoto; reserve HAT-L for hero shots. For 2× outputs, Real-ESRGAN ×2 beats running a 4× model and shrinking.
Noisy photos
| Model | Scale | Best for | Cost |
|---|---|---|---|
| RealPLKSR Denoise | 1× | Fast, light cleanup of grainy snapshots — evening phone photos, high-ISO shots. Cleans while keeping size and character. | Low · ~30 MB · ~1 GB VRAM |
| NAFNet SIDD Width64 | 1× | The strongest denoise option: trained on real smartphone camera noise (SIDD). Highest-fidelity preset; large download, noticeably slower. | High · ~464 MB · ~4 GB VRAM |
| SCUNet Color — Blind Denoise | 1× | Real photos with unknown mixed noise and mild compression. No strength setting; this variant keeps texture rather than inventing it. | Medium · ~72 MB · ~1.5 GB VRAM |
All three are 1×: they don’t enlarge. Denoise first, then upscale — clean input makes every upscaler better. Two-stage jobs (denoise → HAT/SPAN) are a normal LocalSR workflow.
Blurry or JPEG-damaged
| Model | Scale | Best for | Cost |
|---|---|---|---|
| NAFNet GoPro Deblur | 1× | Camera-shake blur from moving subjects or long handheld exposures. Trained on the GoPro motion set; won’t fix a missed focus point, but handles motion well. | Medium · ~272 MB · ~1.5 GB VRAM |
| FBCNN Color | 1× | Old or re-shared JPEGs: blocking, ringing, mosquito noise. A restoration pass, not an upscaler — chain it before upscaling old downloads and forwarded images. | Medium · ~288 MB · ~2.5 GB VRAM |
For 2004-era camera clips and heavily compressed files, FBCNN → SPAN is the workhorse order.
Anime & illustration
| Model | Scale | Best for | Cost |
|---|---|---|---|
| RealPLKSR 4x HFA2k — Anime | 4× | Anime cels, manga scans, illustrations, clean line art. Photo models soften lines; this one keeps them crisp. Fast enough for batches. | Low · ~30 MB · ~0.8 GB VRAM |
| Real-ESRGAN ×4 Anime — Fast | 4× | A compact anime and illustration model: clean lines and flat colour in a fraction of the photo models’ time. Good for long batches of frames or scans. | Low · ~18 MB · ~0.9 GB VRAM |
HFA2k is published by Philip Hofmann under CC BY 4.0 on its model card, so commercial use is allowed with attribution. (Its GitHub release spells the license “CC-BY-0.4”, a typo for the same license.) Real-ESRGAN Anime is BSD-3-Clause.
Faces
Warning — unresolved rights. The face-restoration checkpoints below were fine-tuned for this project, but their training-data and weight rights are not established. LocalSR does not download them automatically; you must supply a copy you are permitted to use, which LocalSR verifies by size and SHA-256. No commercial-use claim is made for either model or its outputs. Full details and measured trade-offs are in the deep-dive below.
| Model | Scale | Best for | Cost |
|---|---|---|---|
| HAT-S ×4 Face Labs | 4× face pass | Small, damaged, or degraded faces within photos. Runs as a companion to stock HAT-S: non-face regions keep the primary model’s output. Blended to recover damaged faces without regressing clean ones. | High · ~40 MB · ~2 GB VRAM |
| HAT-L ×4 Face Labs | 4× face pass | The same idea on the larger base — more recovery power, more cost. Also a companion; also SDR-trained (HDR quality unverified). | Very high · ~166 MB · ~6 GB VRAM |
Both are experimental blends, not magic: they can smooth texture or alter fine features. Inspect eyes, teeth, and hair before keeping a result.
Video Labs
| Model | Best for | Cost |
|---|---|---|
| SeedVR2-3B (FP16 and FP8 variants) | Whole-clip video restoration and upscaling. The FP8 variant trades a little fidelity for a smaller download on NVIDIA CUDA / AMD ROCm. Requires its own supported GPU backend — it does not run through the image models’ paths. | Extreme · 6.8 GB + 0.5 GB per file · 16 GB+ memory · hours to days per clip |
Experimental: SDR output only, memory use depends on clip length and output size, and 16 GB does not guarantee a fit. Test on a short segment before committing an evening. Everyday video upscaling in LocalSR uses the image models frame-by-frame — this family is the separate, heavier Labs route.
License table
Every downloadable model, in one place. “Personal use” means private restoration of your own media; “commercial” means selling or publishing outputs as part of a product or service.
| Model | Author | License | Commercial use |
|---|---|---|---|
| SPAN 4x NomosUni | Philip Hofmann | CC BY 4.0 | Allowed with attribution |
| RealPLKSR 4x NomosWebPhoto | Philip Hofmann | CC BY 4.0 | Allowed with attribution |
| RealPLKSR 4x HFA2k — Anime | Philip Hofmann | CC BY 4.0 | Allowed with attribution |
| RealPLKSR Denoise | Philip Hofmann | CC BY 4.0 | Allowed with attribution |
| Real-ESRGAN ×2 / ×4 / ×4 Anime | Xintao Wang et al. | BSD-3-Clause | Allowed under BSD terms |
| SwinIR ×4 Real | Jingyun Liang et al. | Apache-2.0 | Allowed under Apache terms |
| SCUNet Color | Kai Zhang et al. | Apache-2.0 | Allowed under Apache terms |
| HAT-S ×4 / HAT-L ×4 ImageNet | XPixel Group | Apache-2.0 | Allowed under Apache terms |
| HAT-S ×4 Face / HAT-L ×4 Face | HerRei / XPixel Group | Checkpoint rights unverified | Unclear — user-supplied; no commercial claim |
| NAFNet SIDD / GoPro Deblur | Liangyu Chen et al. (megvii-research) | MIT | Allowed under MIT terms |
| FBCNN Color | Jiaxi Jiang et al. | Apache-2.0 | Allowed under Apache terms |
| SeedVR2-3B (FP16/FP8) | ByteDance Seed · numz | Apache-2.0 | Allowed under Apache terms |
| YuNet face detector (helper, not a picker model) | Shiqi Yu · OpenCV Zoo | MIT | Allowed under MIT terms |
Licenses cover the models, not your photos: your media stays yours in every case. When in doubt — especially before any use of the face-model rows — check the linked source page for each model in the app’s catalog, or get advice. This guide is plain-language orientation, not legal advice.
How LocalSR verifies downloads
Every model the app offers is pinned: an exact source URL, an exact file size, and an exact SHA-256 hash recorded before release. Downloads are HTTPS-only, checked against those pins, and installed atomically — a partial or altered file never becomes selectable. Files you import yourself are verified against the same hashes where a model is curated; unverified pickle or TorchScript checkpoints are blocked by default rather than loaded.
That protects you from corrupted and tampered downloads. It is not a rights check: a verified file is still governed by its license.
Deep-dive: availability
Both face checkpoints are optional Labs companions in the current beta. Neither is downloaded automatically or published: their checkpoint and training-data rights are unresolved. If you have a copy you are permitted to use, LocalSR verifies it before loading. Stock HAT remains the general-purpose choice within each pair.
Deep-dive: HAT-S Face · the lighter model
The selected blend combines 90% stock HAT-S with 10% of the face-tuned 95k checkpoint (α = 0.10). Interpolation recovered a useful model after the original GAN/perceptual fine-tune regressed on clean images.
| Bucket | Stock | Selected | Change |
|---|---|---|---|
| Clean | 32.59 | 31.97 | −0.62 dB |
| Mild | 28.70 | 28.83 | +0.13 dB |
| Hard | 25.78 | 25.86 | +0.08 dB |
Hard-bucket SSIM improved from 0.6842 to 0.7012. These are the HAT-S recovery report's measurements, not a new benchmark.
Exact file: base_95k_interp_a0p1.pth
40,484,805 bytes · 4× native scale
SHA-256: 92277daf002214307bea6f1e06b4fa745acdb7690728a0a9a619076e7bc8d7f2
Deep-dive: HAT-L Face · the larger model
Three L1-only stages ran for 250k, 250k, and 299k iterations, without GAN or perceptual loss. Training plateaued in a roughly 0.1 dB band. Raw face-tuned weights improved mildly degraded faces by about 1 dB but lost about 1.9 dB on clean inputs.
From 14 candidates, the selected blend combines 75% stock ImageNet HAT-L and 25% face-tuned run 3 (α = 0.25). Its reported results balance recovery with a much smaller clean-image cost.
| Bucket | Stock | Selected | Change |
|---|---|---|---|
| Clean | 32.78 | 32.58 | −0.19 dB |
| Mild | 28.69 | 29.13 | +0.44 dB |
| Hard | 25.72 | 25.97 | +0.25 dB |
Deltas come from unrounded report values; the displayed clean scores subtract to −0.20 dB. SSIM and ArcFace were also evaluated, but their numeric results are not reproduced here. The VGG perceptual detail stage was not launched.
Exact file: hat_l_x4_face_task4.pth
165,676,233 bytes · 4× native scale
SHA-256: 8a5548208310fcc7195abd4e1cf17ed87faaf3e5c38e2edeb63a45ac5b9c2af4
Deep-dive: use the matching companion
- Select stock HAT-S or stock HAT-L as the primary model.
- Choose Faces under Fix first.
- Acknowledge the checkpoint terms and import the exact copy you are permitted to use. LocalSR checks the file’s size and SHA-256.
- Adjust face fidelity and inspect the result. Non-face regions keep the primary model’s output.
No installer bundles either checkpoint.
Deep-dive: measured gains, with limits
HDR preservation (Labs) keeps float precision and adapts HAT detail to the source’s light and colour. Both face forks remain SDR-trained; their HDR detail quality is unverified. The feature exports 10-bit HLG/PQ as a Labs experiment, with SDR display previews.
These reports use synthetic degradation and checkpoint-selection sets. They do not establish a controlled HAT-S versus HAT-L comparison, general photographic improvements, or video quality. The larger model costs more memory and processing time.
Neither interpolation nor L1-only training guarantees accurate missing detail or preserved identity. Results can look smooth or contain incorrect texture. Keep the original image and inspect eyes, teeth, hair, and other fine features.
Apache-2.0 applies to the HAT source code; it does not independently establish rights to fine-tuned weights or their training images. No commercial-use claim is made. Neither checkpoint is published.