Skip to content

Text to image

A prompt in, a picture out. The one to start with.

Nodes: Prompt → Text to image → Output.

Steps

  1. Flows → Text to image.
  2. Tap the prompt node and write what you want, and what you do not want in Negative.
  3. Optional: tap the Text to image node to choose the model, the size and the sampling settings (below).
  4. Run. The picture lands on the output node and in Results.

Writing prompts for each model family

The model families read prompts differently:

family write it as example
SD 1.5, SDXL short tags, most important first portrait of an old sailor, weathered face, rim light, 35mm
Anima, Anima DiT anime-style tags 1girl, silver hair, school uniform, cherry blossoms
FLUX.2, Z-Image, Qwen Image, Krea 2, Iris plain sentences An old sailor smiling at the camera, late afternoon light behind him.
  • For SD 1.5 and SDXL you can weight a phrase: (red scarf:1.3) for more, (hat:0.7) for less.
  • FLUX.2, Z-Image, Qwen Image, Krea 2, Iris and Anima DiT are run at CFG 1, where the negative prompt is not read at all. Put what you want in the prompt instead.
  • A prompt in Russian or Chinese can be translated to English on the phone with the 文A button on the prompt box (SD 1.5 and SDXL only read English). See Settings → Prompts → Translation.
  • A long prompt? The ⤢ button on the prompt box's top edge opens it in the whole screen; Done puts it back, Cancel leaves the box as it was.
  • No words for it? The picture button on the prompt box describes a picture you already have, either a render on the canvas or a photo from your gallery, as sentences or, for anime models, Danbooru tags, and writes that into the prompt (Prompt node).

The generator's settings

Tap the Text to image node:

The generator's settings sheet

setting what it does tip
Checkpoint The model Grouped by family; only installed models are listed
Steps How many refinement passes More is slower; each model opens on the number its author recommends
CFG How strictly the prompt is followed Too high burns colours; turbo/distilled models want ~1–2
Seed The starting noise 0 = a new picture every Run. Type a seed shown on a node to get that picture back
Scheduler The sampling method. SD 1.5 and SDXL: DPM++ 2M, DPM++ 2M SDE, DPM++ 3M, DPM++ 3M SDE, DPM++ SDE (two model runs a step, so twice as slow), Euler, Euler A, LCM. FLUX.2, Z-Image, Qwen Image, Krea 2, Iris and Anima DiT: Euler, Euler A, Heun, DPM++ 2M, DPM++ 2M SDE, ER-SDE, Res Multistep, iPNDM, LCM Use what the model's author recommends; turbo models usually want Euler (DiT) or LCM
Schedule How the noise steps are spaced: Normal, Karras or Exponential (SD 1.5, SDXL and the DiT models but Wan and Iris; not with LCM), and on the DiT models also SGM Uniform, Bong Tangent, Simple, Beta, Beta57, Linear Quadratic and Radiance Shift. Linear Quadratic spends half its steps near full noise, so it wants 20 steps or more. Normal is each model's own. On a 4- or 8-step DiT model Karras and Exponential come out murky at those steps: give them about 12 Many turbo and hyper SDXL models recommend Exponential
Resolution / Aspect / Width, Height The picture's size Depends on the family. Below

Size, per family

  • SD 1.5: pick a Resolution from the sizes the model supports (512 square up to 1024, portrait and landscape). A different size reloads the model on the next Run.
  • SDXL and Anima: these draw a fixed 1024 square; Aspect crops it to the ratio you choose (1:1, 9:7, 4:3, 3:2, 5:3, 16:9 and their portrait versions). No reload.
  • FLUX.2, Z-Image, Qwen Image, Krea 2, Anima DiT: Width and Height sliders, any size from 512 to 2048 in 64-pixel steps. Aspect above them sets both to a ratio at about the same size, and the sliders fine-tune it. No reload. Bigger takes longer, roughly in proportion to the area. FLUX.2 Klein 9B starts at 768: on a 12 GB phone, 1024 may be closed by Android when other apps hold memory.

Width and height sliders on a DiT model

How long it takes

Roughly, on a Snapdragon 8 Elite:

family a 1024-ish picture
SD 1.5 (512) a few seconds
SDXL, Anima tens of seconds to over a minute
Anima DiT about a minute at 1024, a quarter of that at 512
FLUX.2 Klein 4B (4 steps) under a minute
FLUX.2 Klein 9B (4 steps, 768²) about a minute
Qwen Image 2.1 (20 steps, 1024²) about 4 minutes
Qwen Image 2.1 Turbo (8 steps, 1024²) under 2 minutes

The first Run after choosing a model also loads it, which adds a few seconds (more for the large ones).