Upscaler Catalog#
The eight upscaler models Eimi can run, and the settings that surround them. This page is the catalog -- which model to pick. For the four upscaling methods and how to launch them, see Upscaling Methods.
The eight upscalers#
Listed in the order the picker shows them by default.
| Name | Scale | Best for | Notes |
|---|---|---|---|
| SeedVR2 | AI restore, 2x-4x | Restoring soft, damaged or heavily compressed images | Not an ESRGAN model. A diffusion transformer with its own pipeline. Output is capped at 4096 px on the long side, so 4x is the practical ceiling. Eimi clears VRAM before it runs. |
| AnimeSharp (4x) | 4x | Anime and illustration, sharp linework | The global fallback. Any model without its own default upscaler uses this one. |
| UltraSharp (4x) | 4x | Anything, when you are not sure | The safe all-rounder. |
| NMKD-Siax_200k (4x) | 4x | Realistic and photographic images | The default for the Krea 2 models. |
| Remacri (4x) | 4x | Realistic images, textures and skin | The default for every Z-Image model. |
| ESRGAN (8x) | 8x | General purpose, when you need 8x | The only 8x model in the catalog. |
| PureScale 4x (SAFMN) | 4x | High-fidelity 4x on images with JPEG damage | Handles compression artifacts rather than sharpening them. |
| R-ESRGAN 4x+ Anime6B | 4x | Anime hi-res fix | The recipe WAI Anima (v1.0) uses for its own hi-res fix pass, on the model author's recommendation. |
SeedVR2 always sorts first
SeedVR2 is pinned to the top of the picker regardless of your sort setting. The other seven can be reordered or hidden per user in /settings.
Licenses#
| Upscaler | License |
|---|---|
| AnimeSharp (4x) | CC BY-NC-SA 4.0 |
| UltraSharp (4x) | CC BY-NC-SA 4.0 |
| NMKD-Siax_200k (4x) | WTFPL |
| Remacri (4x) | CC BY-NC-SA 4.0 |
| ESRGAN (8x) | Apache 2.0 |
| PureScale 4x (SAFMN) | Custom |
| R-ESRGAN 4x+ Anime6B | BSD 3-Clause |
Scale options#
/upscale offers three scales:
| Scale | Effect |
|---|---|
| 2x | Doubles each side. 4x the pixels. |
| 4x | Quadruples each side. 16x the pixels. |
| 8x | 8x each side. 64x the pixels. |
SeedVR2 disappears at 8x
Pick 8x and SeedVR2 is filtered out of the list -- its output cap makes 4x the maximum it can honour. The ESRGAN-family models all accept up to 8x.
Ultimate SD Upscale tiling#
The tiled diffusion method re-renders your image in overlapping tiles instead of running a single upscale model over the whole frame. What that means for you:
| Behaviour | Value |
|---|---|
| Input size cap | The source is resized down to 4 MP before tiling |
| Tile size | Half the input dimensions, clamped to 256-1024 px per side |
| Denoise | 0.2 by default |
| Sampler, steps, CFG | Inherited from the source job, then from the model's own defaults, falling back to deis / beta at 20 steps and CFG 5 |
Your prompt is replaced on most models
Only SDXL-family models get your original scene prompt fed into each tile. On every other architecture -- Krea 2, Flux 2, Z-Image, Anima, ERNIE, Qwen, NetaYume, Mage Flow, Flux Dev, Mugen -- the tile prompt is swapped for a fixed generic one (highly detailed, sharp focus, high quality, intricate details, clean, negative blurry, low quality, low resolution, artifacts, jpeg artifacts). Prompt-obedient models otherwise paint a complete miniature scene into every tile.
Three families refuse tiled upscaling entirely
Z-Image, NetaYume and Ideogram 4 reject this method with an error. Use the model-based upscale or SeedVR2 on those.
Hi-res fix factor clamp#
The hi-res fix inside /imagine picks its upscaler from the model's own default (or from a model-specific hi-res recipe where one exists), then re-denoises the enlarged image.
| Setting | Default | Range |
|---|---|---|
| Upscale factor | 1.5, or the model's own value | 1.25 to 2.0, raised to 1.25 to 8.0 only when /upscale's refine method is driving it from a 2x / 4x / 8x choice |
| Denoise | 0.35, or the model's own value | 0.1 to 0.7 |
| Steps | The model's own value, else 75% of the parent job's steps | 1 to 50 |
| CFG | The parent job's CFG, else the model's CFG, else 7.0 | 0.5 to 20.0 |
1.25 is the floor
Ask for a smaller factor and it is clamped up to 1.25. Below that the pass costs time without visibly changing anything.
Unlike the tiled and detailer paths, no architecture is blocked from hi-res fix. See Hi-Res Fix and Detailer.
Each model's default upscaler#
These models pin their own upscaler for hi-res fix and for upscales launched from a result card. Everything else falls back to AnimeSharp (4x).
| Model | Default upscaler |
|---|---|
| Krea, Krea (Quality), Krea (Exp) | NMKD-Siax_200k (4x) |
| Z Image Turbo, Z Image Base, Cyber Realistic Turbo, ZiT-ANI | Remacri (4x) |
| Z Anime Base, Z Anime 8-Step, Z Anime 4-Step | Remacri (4x) |
| WAI Anima (v1.0) | R-ESRGAN 4x+ Anime6B (via its own hi-res fix recipe) |
| everything else | AnimeSharp (4x) |
You can always override this by passing an upscaler: of your own.
One upscaler you cannot pick
The MrFlow presets in /turbo use a fixed intermediate upscaler between their two sampling stages. It is internal to those presets and never appears in the picker.
Related#
- Upscaling Methods -- model-based, SeedVR2, tiled and refine, and when to use each
- /upscale -- the command
- Hi-Res Fix and Detailer
- Model overview