Skip to content

Parameters#

Every knob /imagine gives you, what it actually changes, and which way to move it. If you only read one section, read How ranges and defaults are decided -- it explains why the number you typed is not always the number you got.


How ranges and defaults are decided#

Almost nothing on this page has a single global value. Steps, CFG, sampler, scheduler, resolution, batch size, quality tags, negative prompt and default upscaler all come from the entry for the model you picked. Switching models silently changes all of them at once.

Those per-model entries are hand-written by the maintainer and get retuned as models come and go, so the exact numbers on this page describe how things are set up right now rather than anything fixed. The result card and the details panel always show what actually ran.

Each model also declares a recommended range for steps and CFG. Your accepted range is derived from it:

Parameter Lowest you may pass Highest you may pass
steps: a quarter of the model's recommended minimum, rounded up, never below 1 double the model's recommended maximum
cfg: a fifth of the model's recommended minimum, rounded to 1 decimal double the model's recommended maximum

Out-of-range values are clamped, not rejected

If you ask for steps: 200 on a model that tops out at 60, you do not get an error. You get 60 steps and a card that says 60. The same applies to CFG, denoise and batch size. Check the parameters on your result card if the output looks unlike what you asked for.

If a model declares no recommended block at all, the fallback range is 5 to 100 steps and 0.2 to 20.0 CFG, with defaults of 25 steps and CFG 5.0.


Steps#

What it does: the number of denoising passes the sampler makes. Each pass removes a little more noise and commits a little more detail.

Option steps:
Default the model's own value
Range a quarter of the model's recommended minimum up to double its recommended maximum

Worked examples of what that formula gives you:

Model Recommended Default You may pass
WAI (v16.0) 15--30 30 4--60
Krea 8--16 12 2--32
Flux 2 Klein Fast 2--8 4 1--16
Z Image Base 25--55 35 7--110

Which way to move it: up if fine detail (hair strands, fabric, small background objects) looks mushy; down if you only want a faster preview. Past the model's recommended maximum the picture stops improving and only the render time grows.

Tip

Do not raise steps on a distilled model. Z Anime 4-Step, Flux 2 Klein Fast, Mage Flow Turbo, ERNIE Turbo, Echo Dream and the Krea turbo build are trained to land in their step count. Running them at 40 steps makes them worse, not better.


CFG#

What it does: how hard the sampler is pushed toward your prompt and away from the negative prompt. Low CFG wanders and looks soft; high CFG obeys but eventually burns contrast and saturation.

Option cfg:
Default the model's own value
Range a fifth of the model's recommended minimum up to double its recommended maximum
Model Recommended Default You may pass
WAI (v16.0) 5--7 5 1.0--14.0
Hassaku (v3.0) 5--7 7 1.0--14.0
Krea 1--2 1 0.2--4.0
Z Image Turbo 1--2 1 0.2--4.0

Values are clamped and then rounded to one decimal place.

Which way to move it: up by 1 at a time when the model keeps ignoring part of your prompt; down by 1 when colours look scorched, skin looks plastic, or edges pick up a hard black outline.

Distilled and lightning models run at CFG 1

Every low-step model in the library sits at CFG 1 (Echo Dream at 1.5). Guidance is baked into the distillation, so raising CFG degrades the image instead of tightening prompt adherence. On Ideogram 4 (Fast) the value is pinned to 1.0 inside the graph, so a cfg: you pass has no effect at all. On Flux Dev your cfg: is not CFG either -- it is re-used as the guidance number (that model's real quality lever, default 3.5) while the sampler itself stays at 1.0.


Sampler#

What it does: the algorithm that walks the image from noise to finished picture. Different samplers reach different results from the same seed.

Option sampler:
Default the model's own value, global fallback euler

The autocomplete is filled live from ComfyUI, so the picker is the authoritative list. These are the ones the shipped models actually use:

Sampler Character
euler Clean and deterministic. Same seed, same result.
euler_ancestral Adds fresh noise each step, so results vary more and often look livelier. Changing the step count changes the image.
dpmpp_2m Good detail, converges tidily, common on realistic models.
dpmpp_2m_sde Stochastic version of the above. More texture, slightly slower.
res_multistep Multi-step residual method, used by Z Image Base.
er_sde Stochastic method used by one of the Anima previews.
deis The generic fallback for the USDU tile pass.

Which way to move it: stay on the model's default unless you have a reason. If you want the same image back every time from a fixed seed, prefer a non-ancestral sampler (euler, dpmpp_2m). If your outputs feel flat and samey, try euler_ancestral.

A LoRA can change your sampler

The first LoRA in your stack that declares a sampler wins, and it does so silently. An explicit sampler: on the command always beats both the LoRA and the model.


Scheduler#

What it does: the noise schedule -- how much noise comes off at each step, rather than which algorithm takes it off. It changes how the render splits its effort between composition (early steps) and detail (late steps).

Option scheduler:
Default the model's own value, global fallback normal
Scheduler Character
normal Standard schedule, the default on most SDXL-family models.
simple Flat schedule that suits distilled and DiT models. Used by nearly every Flux 2, Krea 2, Ideogram 4, Mage-Flow and ERNIE entry.
sgm_uniform Even spacing, often paired with the dpmpp samplers.
karras Front-loads the noise removal. Used by Cyber Realistic Pony.
beta Used by the Z Anime entries and Cyber Realistic Turbo.
exponential Available in the picker.

Which way to move it: this is the last knob to touch, not the first. If the model's default pairing looks wrong, change the sampler before you change the scheduler. A LoRA can override the scheduler under the same first-one-wins rule as the sampler.


Seed#

What it does: picks the starting noise. The same seed plus identical everything else gives you the identical image.

Option seed:
Default random, from 0 to 4,294,967,295
  • Leave it empty and every run gets a fresh random seed. That is what the 🎲 re-roll button does for you.
  • Set it and you can change one other thing at a time and see only that change.
  • Seed 0 has no special meaning. It is a seed like any other.

Tip

To reproduce a picture exactly you need the seed and the model, steps, CFG, sampler, scheduler, dimensions, LoRAs and the final prompt text. Enhancement and character detection rewrite the prompt, so an enhanced job re-run at the same seed with a fresh enhancement pass will not match. See AI Enhance.


Denoise#

What it does: how much of the starting image is destroyed before the model rebuilds it. It only means something when you attach an input_image:. In plain text-to-image it stays at 1.0, because there is nothing to preserve.

Option denoise:
Default 0.7 with an input image, 1.0 without
Range 0.05--0.98 with an input image, 0.05--1.0 without
Denoise Effect
0.1--0.3 Barely changed. Cleanup and small corrections.
0.3--0.5 Same composition, redrawn surfaces and detail.
0.5--0.7 Serious reinterpretation, still recognisably the same scene.
0.7--0.9 The original is only a rough suggestion.
0.9+ Effectively a fresh image.

Your input image is resized first

On every architecture that does real image-to-image (the SDXL family, Anima, Z-Image, Mugen, Qwen), the attachment is normalised to roughly 1 megapixel before it is encoded. A 4K upload does not give you a 4K starting point.

Krea 2 does not do image-to-image

An attachment on a Krea model becomes a vision reference handed to the text encoder, not a starting latent. Denoise does nothing there.


Batch size#

What it does: how many images the one job renders. They share the prompt and every parameter, and differ only by seed.

Option batch_size:
Default 2 unless the model overrides it
Range 1 to the model's maximum, fallback 9

Time scales roughly with the count, so a batch of 4 takes about four times as long as a batch of 1. Two presets pin their own: Krea (Quality) renders 1 at a time, Ideogram 4 (Turbo) renders 2.

Which way to move it: up while you are still exploring a prompt and want variety to choose from; down to 1 once you have locked a seed and are refining, since only one image matters.


Resolution and aspect ratio#

What it does: sets the output size. One option covers two very different input styles.

Option aspect_ratio:
Default the model's own default size, global fallback 864x1152 (3:4)

Ratio form (3:4, 16:9, 21:9)#

Resolved in this order:

  1. The model's own aspect ratio table, when it lists that ratio. You get the exact pixel pair the model was tuned for. This always wins.
  2. Otherwise the size is computed from the model's pixel budget -- its default width times its default height, or 1,000,000 pixels when it declares none. The result is rounded to a multiple of 8 and clamped to 4096 per side on models that carry their own pixel budget, 2048 on everything else.

Ratios more extreme than 32:1 are clamped to 32:1 before anything is computed.

Pixel form (1024x768)#

The literal size is used. It bypasses the model's pixel budget entirely, snaps down to a multiple of 8, and never goes below 64. The longest side is capped at 4096, and if you exceed that the whole image is scaled down proportionally so your ratio survives.

This is how you get a size the ratio table will not give you. It is also how you ask for something the model was never trained on, so expect duplicated limbs and repeated background elements once you push well past the model's normal size.

Defaults by model subtype#

When a model does not ship explicit dimensions, its subtype decides:

Subtype Default size Approx. pixels
Illustrious 1152x1536 1.77 MP
NoobAI, Pony, plain SDXL, Mugen 896x1152 1.03 MP

Models that do ship their own tables use those instead:

Family Default 1:1 16:9 21:9
Krea 2 (Default preset) 1024x1024 1024x1024 1280x720 1536x656
Krea 2 (Quality preset) 1408x1408 1408x1408 1920x1088 2112x896
Z-Image / Z-Anime 1024x1536 1280x1280 1600x896 1680x720
Cyber Realistic Turbo (Z) 1280x1920 1568x1568 2096x1168 2400x1024
Flux 2 Klein, Mage-Flow, ERNIE 1024x1024 1024x1024 1344x768 1536x656
Qwen Image 1328x1328 1328x1328 1664x928 --
Anima 896x1152 1024x1024 -- --
Flux Dev 1024x1024 1024x1024 1344x768 1536x640

Two families will not give you what you asked for

Ideogram 4 clamps every side to 256--2048 in multiples of 16, and wide ratios shrink the short side instead of growing the long side, so all three of its presets land on the same 2048-wide image at 21:9. Krea (Quality) outputs roughly double the size you requested, because its decoder upscales 2x and the compensating downscale step is removed for that preset.

Note

The ratio printed on your result card is fuzzy-matched. If the real width over height lands within 0.02 of a standard ratio (9:16, 4:5, 1:1, 3:4, 4:3, 3:2, 16:9, 21:9) that name is shown; otherwise you get the simplified fraction.


Shift#

What it does: shift moves where in the schedule the model spends its effort. A higher shift puts more of the compositional decision-making into the early steps, which is what the flow-matching architectures want.

You cannot set it. There is no shift: option. It is a per-model value baked into the configuration, listed here so you know why two models at the same steps and CFG behave differently.

Family Shift
Z-Image / Lumina2 3
Cyber Realistic Turbo (Z) 6
Qwen Image 3.1
Ideogram 4 5.0
Ideogram 4 (Fast) 3.0
Mugen 12

The SDXL-family models have no shift stage at all.


  • /imagine -- where all of these live
  • LoRAs -- the stack that can override your sampler and scheduler
  • Hi-res fix and detailer -- the second pass, with its own steps and denoise
  • Upscalers -- the four ways to make the finished image bigger
Categories: Imagine | Models | LoRAs | Settings