/turbo#

/turbo is accelerated image generation. Instead of picking a raw checkpoint like /imagine does, you pick a turbo preset, and the preset decides which acceleration backend runs your prompt. It is text-to-image only.

/turbo prompt: a fox in a snowy forest

Like every generation command it works in servers, in DMs, and through a user install in servers where Eimi is not added.

When to use /turbo instead of /imagine#

Use `/turbo` when	Use `/imagine` when
You want the shortest wait per image	You want a specific checkpoint from the full 53-model list
You are exploring prompt ideas and want fast iterations	You need img2img (`/turbo` has no `input_image`)
The preset list already contains the look you want	You want full control of every sampling parameter at full resolution
You want big output sizes cheaply (MrFlow samples small, then upscales)	You want the model to do all its work at the final resolution

/turbo shares almost everything else with /imagine: the same access rules, the same 15-job concurrent limit, the same queue card, the same result card and buttons.

The two backends#

Every preset declares one backend. This is the single most important thing to understand about /turbo, because the backend decides which of your parameters are even allowed.

	TensorRT	MrFlow
How it works	A pre-compiled engine for one exact model, resolution and batch size	Samples at roughly half scale, upscales with a Real-ESRGAN 2x model, then runs a short refine pass
Speed	Fastest per image, on its fixed sizes	Much less work than a full-resolution `/imagine` run at the same output size
Aspect ratio	Only the orientation names its engines were built for: `portrait`, `landscape`	Full `/imagine` grammar: any `W:H` ratio or `WxH` pixel size
Batch size	Fixed per preset (Echo Dream 1-6, Cat Carrier 1-4), default 2	1-9, default 1
LoRA	Not supported	Supported (with two preset exceptions, see below)
`denoise`	Not supported	Supported, sets the refine strength
`upscaler`	Not supported	Supported, replaces the pipeline's upscale model
Steps / CFG / sampler / scheduler	Applied to the single pass	Applied to the stage-1 generation only
Embeddings, enhancer, negative prompt	Preset carries its own baked prompt scaffolding	Near-full `/imagine` parity

TensorRT engines are static

A TensorRT preset only runs if a compiled engine exists on disk for the exact combination of batch size, width and height you asked for. If one is missing you get an ephemeral warning reading TensorRT engine not available for this configuration. with the batch size and orientation it tried, and a suggestion to change one of them.

All 20 parameters#

Only prompt is required.

Parameter	Type	Default	What it does
`prompt`	text (required)	--	Describe what you want to create
`model`	text (autocomplete)	Z-Image Turbo (MrFlow 9+1)	Which turbo preset, and therefore which backend, runs the job
`negative`	text	your saved negative	What to avoid
`batch_size`	number	TensorRT 2, MrFlow 1	Number of images to generate
`aspect_ratio`	text (autocomplete)	your saved ratio, else the preset's size	Ratio for MrFlow, orientation name for TensorRT
`enhance`	true/false	from your AI features setting	Rewrite your prompt with AI first
`cfg`	number	the preset's stage-1 CFG	CFG applied to stage-1 generation
`steps`	number	the preset's stage-1 steps	Steps applied to stage-1 generation
`scheduler`	text (autocomplete)	the preset's stage-1 scheduler	Scheduler for stage-1 generation
`sampler`	text (autocomplete)	the preset's stage-1 sampler	Sampler for stage-1 generation
`seed`	number	random	Fixes the noise pattern for a reproducible run
`lora`	text (autocomplete)	your saved LoRA slots 1-4	LoRA to apply, MrFlow only
`lora_strength`	number	the LoRA's own default	Strength of the LoRA, -10 to 10
`upscaler`	text (autocomplete)	the preset's own upscale model	Override the pipeline upscaler, MrFlow only
`denoise`	number	the preset's refine strength	Override the refine strength, MrFlow only, typical 0.08-0.20
`attention`	text (autocomplete)	`auto`	Attention backend used for the run
`nsfw`	true/false	your saved setting	Removes the safe-content prompt additions
`raw`	true/false	your saved setting	Disables all default prompt scaffolding and the enhancer
`debug_backend`	choice	--	Experimental. Force a backend on any model, bypassing the preset list
`debug_model`	text (autocomplete)	--	Experimental. Checkpoint to run through the forced backend

Parameters that behave exactly as on /imagine (negative, enhance, seed, lora_strength, attention, nsfw, raw) are documented in full on the /imagine page.

The 8 presets#

The default is marked with ★ in the autocomplete, and presets sort default first, then alphabetically.

Preset	Backend	Base model	Steps	Notes
Z-Image Turbo (MrFlow 9+1) ★	MrFlow	Z-Image Turbo	9 stage-1 at CFG 1.0 (euler / simple) + 1 refine step at strength 0.11	The default. Much faster than 9+9 while keeping the stage-1 structure intact
Z-Image Turbo (MrFlow 9+9)	MrFlow	Z-Image Turbo	9 stage-1 + 9 refine steps at strength 0.11	Highest quality of the Z-Image trio, and the slowest
Z-Image Turbo (MrFlow 6+1)	MrFlow	Z-Image Turbo	6 stage-1 + 1 refine step at strength 0.13	The most aggressive setting. Fastest, but can lose coherence on complex prompts
Qwen-Image (MrFlow)	MrFlow	Qwen-Image	12 stage-1 at CFG 2.5 + 1 refine step at denoise 0.12, CFG 4.0	LoRAs are silently dropped on this preset
Krea	MrFlow	Krea 2 Turbo	8 stage-1 at CFG 1.0 + 3 refine steps at strength 0.13	Experimental. The refine pass is 3 steps so the progress bar visibly moves
Ideogram 4 (MrFlow)	MrFlow	Ideogram 4	12 stage-1 at CFG 7.0 (res_multistep) + 3 refine steps at strength 0.11, CFG 4.0	Keeps Ideogram's dual-model path end to end. LoRAs are not supported yet
Echo Dream v1.0	TensorRT	Echo Dream v1.0	12 steps at CFG 1 (euler_ancestral / simple)	Batch 1-6. Portrait 1152x1536, landscape 1536x1152. Recommended steps 8-15, CFG 1-1.8
Cat Carrier v8.0	TensorRT	Cat Carrier v8.0	30 steps at CFG 6 (euler_ancestral / normal)	Batch 1-4. Same two fixed resolutions. Recommended steps 20-40, CFG 5-7

Note

TensorRT presets carry their own baked positive, negative and safe-content text. Cat Carrier v8.0, for example, always contributes masterpiece, best quality, absurdres, very aesthetic to the positive and (safe:1.2) when NSFW mode is off.

If no presets are configured at all you get a "No Turbo Model" message telling you to ask an admin.

Presets are hand-built by the maintainer, and TensorRT ones in particular depend on compiled engine files sitting on a specific machine. They get retuned, replaced and thrown out as models change, so treat this table as what the list looked like when the page was written rather than a fixed menu.

The speed and quality ladder#

Within the Z-Image family the three presets are the same model with a different amount of refine work. That is the whole ladder, from fastest to best:

Rank	Preset	Stage-1 steps	Refine steps	What you trade
1 (fastest)	Z-Image Turbo (MrFlow 6+1)	6	1	Least structure. Complex or crowded prompts can fall apart
2 (default)	Z-Image Turbo (MrFlow 9+1)	9	1	Full stage-1 structure, minimal refine. The sensible everyday choice
3 (best)	Z-Image Turbo (MrFlow 9+9)	9	9	Crispest detail after the upscale, and the longest wait of the three

Reading it in general terms:

More stage-1 steps buys composition and coherence, because that is where the image is actually decided.
More refine steps buys crispness after the 2x upscale. Fewer refine steps look softer.
TensorRT skips this ladder entirely: one pre-compiled pass, no upscale stage, fastest per image but locked to two resolutions and no LoRAs.

Parameters in detail#

model#

The list is the presets from the turbo config, not the /imagine model list. Backend selection is automatic, so picking a preset is also picking what you are allowed to override.

aspect_ratio#

Behaviour depends on the backend of the preset you picked, and the autocomplete adapts to it:

Backend	Autocomplete offers	Free text accepted
TensorRT	Only the orientations that preset has engines for: `portrait`, `landscape`	No. Anything else gives `TensorRT model only supports these orientations: portrait, landscape`
MrFlow	`1:1`, `3:4`, `4:3`, `2:3`, `3:2`, `9:16`, `16:9`, `4:5`, `5:4`	Yes, the full `/imagine` grammar including `WxH` pixel sizes

steps, cfg, sampler, scheduler#

On a MrFlow preset these all apply to the stage-1 generation, the low-resolution pass that decides the image. The refine pass keeps the preset's own settings unless you change denoise.

Validation limits are:

Parameter	Accepted range
`cfg`	0 to 30
`steps`	1 to 150
`denoise`	0.0 to 1.0
`seed`	0 to 4294967295
`batch_size` (MrFlow)	1 to 9
`batch_size` (TensorRT)	The preset's own min and max

Beyond those hard limits, each preset was tuned as a whole. Doubling the stage-1 steps on a 9+1 preset mostly buys you a longer wait.

denoise#

MrFlow only, and it means "how hard the refine pass repaints the upscaled image". The typical range is 0.08 to 0.20. Higher values add detail but can also add artefacts and drift away from your stage-1 composition. On a TensorRT preset it is rejected outright.

lora and lora_strength#

MrFlow only. Same rules as /imagine: up to 4 LoRAs, one from the command plus your saved slots, and passing lora on the command clears your saved slots for that run.

Two MrFlow presets are exceptions: Qwen-Image silently drops LoRAs, and Ideogram 4 does not support them yet. On a TensorRT preset the command refuses with LoRAs are only supported for MrFlow-backed models (TensorRT engines are static).

Browse what is available on the LoRAs page.

upscaler#

MrFlow only, and unlike /imagine it is live: it replaces the model used for the pipeline's 2x upscale between stage 1 and the refine pass, so it directly changes how the final image looks. See Upscalers.

debug_backend and debug_model#

Experimental

These two bypass the preset list and force any checkpoint through a chosen MrFlow variant. They exist for testing new models and can fail in ways normal presets do not. debug_backend without debug_model is rejected with a message saying it requires debug_model to be set.

The four forced backends are:

Choice	Meant for
`mrflow (Z-Image / generic flow-match, no reference_latents)`	Z-Image and other generic flow-match models
`mrflow (Qwen / with reference_latents)`	Qwen-Image
`mrflow (Krea 2 / Flux-family flow-match)`	Krea 2 and Flux-family models
`mrflow (Ideogram 4 / single-model simplified)`	Ideogram 4

Error messages you may hit#

Message	Why
`LoRAs are only supported for MrFlow-backed models (TensorRT engines are static).`	You passed `lora` on a TensorRT preset
`denoise only applies to MrFlow-backed models.`	You passed `denoise` on a TensorRT preset
`upscaler only applies to MrFlow-backed models.`	You passed `upscaler` on a TensorRT preset
`Batch size must be between {min} and {max} for this TensorRT model`	Outside that engine's compiled batch sizes
`Batch size must be between 1 and 9`	MrFlow batch out of range
`TensorRT model only supports these orientations: portrait, landscape`	A ratio or pixel size on a TensorRT preset
`CFG must be between 0 and 30` / `Steps must be between 1 and 150` / `Denoise must be between 0.0 and 1.0` / `Seed must be between 0 and 4294967295`	Value out of range
`LoRA {i}: '{name}' is not compatible with {preset}`	That LoRA does not match the preset's base model
`TensorRT engine not available for this configuration.`	No compiled engine for that batch size and orientation combination

Worked examples#

Default preset, nothing else:

/turbo prompt: a fox in a snowy forest

Fastest possible Z-Image run, four images:

/turbo prompt: quick concept sketch of a treehouse model: Z-Image Turbo (MrFlow 6+1) batch_size: 4

Best-quality Z-Image run in widescreen:

/turbo prompt: a storm over a mountain lake model: Z-Image Turbo (MrFlow 9+9) aspect_ratio: 16:9

MrFlow with a LoRA and a stronger refine pass:

/turbo prompt: a knight in a rainstorm lora: <pick> lora_strength: 0.7 denoise: 0.16

A TensorRT preset, which only accepts an orientation:

/turbo prompt: a cat on a windowsill model: Echo Dream v1.0 aspect_ratio: portrait batch_size: 4