--- title: "AI modes" url: https://phatdev.net/en/docs/pixgamma/che-do-ai lang: en updated: 2026-10-22 alternate: vi: https://phatdev.net/docs/pixgamma/che-do-ai markdown: https://phatdev.net/en/docs/pixgamma/che-do-ai.md --- # AI modes Part of the doc set: [PixGamma](https://phatdev.net/en/docs/pixgamma.md) > Auto, Standard, High Fidelity, Art & CGI, Low Resolution, sizing, adjustments, output and speed. PixGamma has six modes. **Auto** measures the photo and picks the model and strengths; the others are there to choose by hand.
Nature: real feathers, no invented textureNature: real feathers, no invented texture
Nature: real feathers, no invented texture
## Modes | Mode | Model | Use for | |---|---|---| | **Auto** | Chosen for you | Default. Measures noise (Immerkær), sharpness (Laplacian variance), JPEG blocking, flat areas and faces, and recognizes artwork | | **Standard** | PixGamma Photo v2 (4x), SPAN (2x) | Everyday photos: products, landscapes, animals. True materials, no invented texture | | **High Fidelity** | RealESRNet | Strict faithfulness; the **Detail** slider blends in detail from Standard | | **Art & CGI** | Real-ESRGAN anime | Paintings, anime, graphics, renders | | **Low Resolution** | realesr-general | Very small, noisy images; continuous denoising | | **Recover** | Diffusion (SD 1.5 + ControlNet Tile) | Badly blurred photos and portraits: redraws skin, hair and fur. See [Recover](https://phatdev.net/en/docs/pixgamma/recover) | Auto never sharpens on its own and adds no contrast or grain; those sliders are yours to turn on. White-background studio product shots are recognized as photos, not as artwork. Old, grainy scans do best in **Low Resolution**; paintings and anime in **Art & CGI** when Auto does not pick it.
Old family photo (1910): Low Resolution cleans the grain and keeps the original tonesOld family photo (1910): Low Resolution cleans the grain and keeps the original tones
Old family photo (1910): Low Resolution cleans the grain and keeps the original tones
Illustration: Art & CGI keeps clean lines and flat colorIllustration: Art & CGI keeps clean lines and flat color
Illustration: Art & CGI keeps clean lines and flat color
## Size - By factor **1x–16x**, or by **width** or **height** (aspect kept). - PixGamma picks the smallest native scale that reaches the target (2x or 4x) and stacks passes above 4x. - Images are split into 256 px tiles with surrounding context, so no seams show. ## Adjustments | Slider | Effect | |---|---| | Denoise | Removes grain before upscaling | | Sharpen | Adds crispness after upscaling | | JPEG cleanup | Removes 8×8 blocks and ringing | | Dehalo | Clips bright halos around high-contrast edges (default 80) | | Local contrast | Adds depth and saturation | | Texture | Adds fine natural grain to areas that are too smooth | | Faces | **Realistic** (CodeFormer, with an identity slider) or **Creative** (GFPGAN). Black-and-white and sepia photos keep their tone, no blue eyes or pink skin painted in | After upscaling PixGamma **keeps the original color**: low-frequency color and tone are pulled back to the source, so the model cannot shift the palette. Transparent images keep their alpha channel. ## Output PNG, JPEG, TIFF, HEIC or the source format. EXIF, GPS and IPTC are kept; DPI can be written; prefix, suffix and model name can be added to file names. **The original is never overwritten.** ## Compute device GPU by default. `all` (GPU + Neural Engine), `ane` and `cpu` are available. PixGamma Photo runs on the GPU only (the Neural Engine compiler cannot build the DAT2 architecture), and PixGamma switches to the GPU automatically when `ane` or `all` is chosen. | Image (Mac Studio M2 Ultra) | Time | |---|---| | 800 × 600 to 4x, Standard | about 5–8 s | | 800 × 600 to 2x | about 1 s | | 3000 × 2000 to 2x | about 3 s | | Recover, 1536 px output | about 35 s |