Ir al contenido
Invoke

Upscaling

Esta página aún no está disponible en tu idioma.

The Upscale panel enlarges an image. An upscaler model, run through Spandrel (such as a Real-ESRGAN variant), enlarges it first. In Refine mode a main model then runs a short image-to-image pass over the result to add detail that fits your prompt.

Open the panel from the left region, where it sits behind Generate in the Compose and Edit layouts and behind Video in the Video layout. To upscale a gallery image without opening it, use the image menu (see Upscaling from the image menu).

How an upscale runs
  1. Source image
  2. Upscaler modelRun through Spandrel
    • Model only
    • Refine
      1. Sharpen
      2. Detail passMain model and prompt
  3. Result

When no model of a kind the panel needs is installed (a Spandrel upscaler, a supported main model, or the Tile or Union ControlNet a model requires), its picker says so and offers Install a model, which opens Add Models in the Model Manager filtered to that kind. Users who cannot manage models are asked to contact an administrator instead.

The Mode switch at the top of the panel picks how the image is upscaled.

Refine

The upscaler, then a detail pass with a main model and your prompt. Adds new detail; needs a main model of a supported family. The default, also for projects saved before the switch existed.

Model only

The upscaler alone: no main model, prompt or detail pass. Faster and faithful to the source. Always queues once whatever Iterations says, because the upscaler gives the same image every time.

Switching modes keeps your Refine settings, so switching back restores them. Model only does not need the model capability list Refine loads, so it works while that list is loading or could not be loaded.

Model only: an illustrated girl in a red scarf in a snowy village at dusk, enlarged by the upscaler alone. Her eyes and the lantern beside her are soft and smeared.
Refine: the same crop after the detail pass. Her eyes, hair and scarf have crisp line work, and the lantern and wood grain show clean new detail.

A crop of the same 2× upscale of a 1024 × 1024 SDXL illustration with the RealESRGAN x4plus upscaler. Refine ran with the panel’s defaults: the Balanced preset, 30 steps, CFG 2 and an SDXL Tile ControlNet.

  1. Refine or Model only. Model only hides everything below Scale.
  2. Pick an image from the gallery, drag one in, or upload one.
  3. The input size, Scale, output size and output area. Model only adds the upscaler’s Native scale. From 50 megapixels the output is marked Large output: it takes much longer and needs more VRAM.
  4. The upscaler. Choose one suited to your image: photography, illustration or linework.
  5. The output size is the source size times this, from 1× to 16× in steps of 0.5. Default 4×.
  6. Refine only.

    • Treatment preset sets Creativity and Structure together: Conservative (−5, 5), Balanced (0, 0), Creative (5, −2) or Artistic (8, −5).
    • Creativity, −10 to 10: higher values invent more detail and give the prompt more influence.
    • Structure, −10 to 10: higher values keep the source composition more strictly. Shown only when a Tile ControlNet is in use.
  7. Refine only, collapsed until you open them.

    • Detail guidance: the prompt, marked Shared because Generate uses the same project prompt. Describe the medium, texture and style rather than new content.
    • Generation: the Main model for the detail pass, Steps, CFG scale, Scheduler, Seed, and concepts (LoRAs) that match the main model.
    • Advanced: the Tile ControlNet and encoders the main model needs, Tile size, Tile overlap, VAE, VAE precision, and CLIP skip for SD1.5.

The panel says what is missing before you can queue, for example “Select a compatible Tile or Union ControlNet.” or “This model needs a T5 encoder selected.” Iterations sits with the Invoke button in the top bar.

Refine works with SD1.5, SDXL, FLUX.1, FLUX.2, Z-Image, Anima, Krea-2 and Qwen Image main models.

FamilyTile ControlNetNegative promptDetail pass
SD1.5, SDXLRequiredYesTile by tile
FLUX.1, FLUX.2––Whole image
Z-ImageOptionalWhen CFG > 1Whole image
Anima, Krea-2, Qwen Image–When CFG > 1Whole image

Where a family has no Tile ControlNet, Structure is hidden. Where it ignores the negative prompt, the prompt is hidden and not recorded.

  • Tile ControlNet: required, a Tile or Union ControlNet of the same family, under Advanced. One is picked for you if installed.
  • CFG scale: kept as you set it. It goes back to 2 when you switch from a family that sets its own.
  • Components: taken from the main model. A VAE can replace the bundled one.
  • Scheduler: any of the family’s.

A value “set when you pick the model” is the model’s default setting, which you can change in the Model Manager. Switching from FLUX.2 to another family sets Steps back to 30.

Creativity sets how far down the noise schedule the detail pass starts: the higher, the noisier the start and the more the main model redraws. The panel offers only schedulers that start where Creativity puts the pass, and hides the field where only one is left.

Schedulers and steps for each family
FamilySchedulersSteps the pass runs
SD1.5, SDXLAll of the family’sThe part of Steps after the start: about half at Balanced
FLUX.1Euler, HeunThe part of Steps after the start: Balanced at 30 steps runs about the last 8 from 1024 px up
FLUX.2Euler onlyStarts where FLUX.1 does, then spreads all of Steps over the rest of the pass, so 30 steps really run 30 times
Z-ImageEuler, HeunStarts at FLUX.1’s noise level; Balanced at 30 steps on a 2048 px result runs the last 7
AnimaEuler, DPM++ 2M, DPM++ 2M SDE, ER-SDEAs Z-Image; the last 8
Krea-2Euler onlyAs Z-Image; the last 7. Turbo and Raw both keep your Steps: Turbo, distilled for 8 steps, runs the end of a 30-step schedule
Qwen ImageEuler onlyAs Z-Image; the last 9

Z-Image, Anima, Krea-2 and Qwen Image count the start in steps rather than in noise level, so the panel translates Creativity for them: it starts on the step whose noise level is closest to where FLUX.1 would start. A preset therefore means the same on every family. These families need at least 2 steps.

LCM builds its own noise schedule, so it would treat the enlarged image as pure noise. Anima’s Heun starts below Creativity’s point and leaves noise in the image.

SD1.5, SDXLOther families
Detail passIn tiles, so the output can be much larger than the model generates in one goOver the whole image at once: memory grows with the output size
Tile sizeThe detail pass’s tiles: larger tiles can improve continuity but need more VRAMOnly the VAE encode and decode tiles (Anima’s decode picks its own)
Tile overlapMore overlap reduces seams but costs time and memoryNo effect

Encoding and decoding the enlarged image with the VAE runs in tiles for every family. If you still run out of memory, see Low-VRAM mode.

Each upscaler has a native scale, such as 4× for 4x-UltraSharp. In Model only mode the upscaler runs until the image is at least the Scale you asked for, then resizes it down to that size. The size preview shows the Native scale and a note on how it gets there:

Scale2× upscaler4× upscaler
1.5×Runs 2× once, then resizes to 1.5×Runs 4× once, then resizes to 1.5×
2×Runs 2× onceRuns 4× once, then resizes to 2×
4×Runs 2× 2 timesRuns 4× once
8×Runs 2× 3 timesRuns 4× 2 times, then resizes to 8×

When Scale equals the native scale the note is left out. A 1× model (a restoration model) keeps the input size, so Scale has no effect. In Refine mode the panel shows no native scale. FLUX.1, FLUX.2, Z-Image, Anima, Krea-2 and Qwen Image resize the upscaler’s result to Scale before the detail pass; SD1.5 and SDXL keep the upscaler’s size, so with a 1× model their result stays at the source size.

A Model only result’s metadata is the source image’s metadata plus the upscaler and the Scale, so Recall on it restores how the source was made.

Upscalers record their native scale when they are installed. One installed with an earlier version has none until it is re-identified: Model only then shows no Native scale row, and a note under the Spandrel model says the native scale is unknown. To record it:

  1. Under the Spandrel model, press Open in Model Manager, or open the upscaler’s page there yourself.

  2. Press Re-identify model and confirm. Invoke re-probes the file, and the page then shows the native scale. Re-identifying also resets the model’s custom variant and default settings.

Only users who manage models can re-identify; others are asked to contact an administrator.

Right-click an image in the gallery, or open Edit in the Preview’s header, and open Upscale:

  • Send to Upscale opens the Upscale panel with the image as its source.
  • Upscale now, model only lists each installed upscaler by name, with its native scale when it is known. Picking one of its factors queues a Model only upscale straight away, without changing the Upscale panel or the current Invoke settings.
Upscaler’s native scaleFactors offered
1× (restoration)1×
2×1.5×, 2× (native)
3×1.5×, 2×, 3× (native)
4×1.5×, 2×, 4× (native)
8×1.5×, 2×, 8× (native)
UnknownNative scale, which runs it once at whatever size it produces; Other factors is disabled with a note to re-identify it

Each factor runs as Model only runs it at that Scale (see Native scale), so the result has the same size as one from the panel. For other factors, or a fractional Scale such as 2.5×, use Send to Upscale. When no upscaler is installed, the entry is disabled.

Results go to the board the Gallery names under Saving to, where new results go. Their metadata is the source image’s metadata plus the upscaler and its scale. If the source image’s metadata cannot be read, nothing is queued and a notification says so.