Z-Image Text-to-Image — ComfyUI workflow

Z-Image text-to-image on the RunPod image endpoint: the 6B base (CFG and negative prompt) or a Civitai Turbo finetune (CyberRealistic v8.0 bf16, Z-Image-Turbo BF16 AIO), with an optional Z-Image LoRA. Translated from Comfy-Org/workflow_templates@aaac56dd templates/image_z_image.json (base: KSampler res_multistep + simple, 25 steps, CFG 4, ModelSamplingAuraFlow shift 3, 1024x1024 from ResolutionSelector 1:1 at 1 MP, empty negative prompt) and templates/image_z_image_turbo.json (Turbo: 8 steps, CFG 1, negative zeroed with ConditioningZeroOut). Choosing a Turbo checkpoint selects the Turbo schedule through checkpoint_profile.

13Graph nodes

12Node classes

10Inputs

10Models

Tags: zimagerunpod-serverlessvolume-backedtext-to-image

Inputs (10)#

The typed parameter surface callers bind when they request this workflow. Enum options and numeric bounds are the values the workflow document declares.

promptstringrequireddefault Latina female with thick wavy hair, harbor boats and pastel houses behind. Breezy seaside light, warm tones, cinematic close-up.

Positive prompt in natural prose (the default is the prompt of the upstream Z-Image-Turbo template)

negative_promptstringdefault

Negative prompt, honoured by the base checkpoint (CFG-guided; the upstream base template leaves it empty). A Turbo checkpoint samples at CFG 1 with the negative zeroed (ConditioningZeroOut, as the upstream Turbo template does), so this text is not encoded for it.

checkpointenumdefault zimage-base-bf16

Z-Image checkpoint on the network volume. zimage-base-bf16 = Z-Image 6B base (UNETLoader, CFG and negative prompt). cyberrealistic-zimage-turbo-v8-bf16 = CyberRealistic Z-Image Turbo v8.0 bf16 (UNETLoader with the shared Qwen3 4B encoder and Flux ae). zimage-turbo-bf16-aio = Z-Image-Turbo BF16 all-in-one (CheckpointLoaderSimple; encoder and VAE come from the same file). The two Turbo checkpoints run the distilled schedule automatically: 8 steps at CFG 1 with the negative zeroed, whatever steps and cfg say.

options: zimage-base-bf16cyberrealistic-zimage-turbo-v8-bf16zimage-turbo-bf16-aio

stepsintegerdefault 25min 8max 60

Sampling steps for the base checkpoint (25 is the upstream base template's value; its note suggests 30-50). Not used by a Turbo checkpoint, which always runs 8.

cfgfloatdefault 4.0min 1.0max 10.0

Classifier-free guidance for the base checkpoint (4 is the upstream base template's value; its note suggests 3-5). Not used by a Turbo checkpoint, which always samples at 1.

seedintegerdefault -1

Random seed (-1 = random)

aspectenumdefault square_1024x1024

Output aspect. The sizes are what the upstream template's ResolutionSelector (1.0 MP, multiple of 8) computes for each of its eight ratios (ComfyUI v0.34.0 comfy_extras/nodes_resolution.py: scale = sqrt(1024^2 / (w*h)), each side rounded to a multiple of 8); they are literals here so the per-job caps can read the latent size.

options: square_1024x1024portrait_2x3_840x1256landscape_3x2_1256x840portrait_3x4_888x1184landscape_4x3_1184x888portrait_9x16_768x1368landscape_16x9_1368x768ultrawide_21x9_1568x672

batch_sizeintegerdefault 1min 1max 1

Images per run. Pinned to 1: the executor's resolution-scaled batch cap (C.12.19) allows a second latent only below ~0.5 MP, and every aspect here is ~1 MP.

loraenumdefault none

Optional Z-Image LoRA (A.03.02, all Civitai with every commercial flag; none of the five versions lists trained words, so no trigger is appended). photoreal-betternudes-v3 = PhotoReal BetterNudes v3 (explicit); realistic-snapshot-v5 = Realistic Snapshot ZIT v5; breast-slider and detail-slider = [ZIT] sliders, driven by strength in either direction; detaildaemonz-v1.1 = DetailDeamonZ v1.1 slider. All but DetailDeamonZ were trained on Z-Image Turbo (DetailDeamonZ lists Z-Image base and serves both); on the base checkpoint the Turbo-trained ones load onto the same architecture but were not trained for it. none leaves the graph without a LoRA loader.

options: nonephotoreal-betternudes-v3realistic-snapshot-v5breast-sliderdetail-sliderdetaildaemonz-v1.1

lora_strengthfloatdefault 1.0min -2.0max 2.0

LoRA strength on the model (LoraLoaderModelOnly.strength_model). The sliders are meant to be driven negative as well as positive.

ComfyUI node graph (13)#

The executable ComfyUI prompt graph: 13 nodes across 12 distinct node classes, wired by 6 data dependencies. Nodes tinted green come from a custom node pack this workflow declares; the rest are ComfyUI core / baked-community classes.

flowchart TD n0["1: UNETLoader"] n1["2: CLIPLoader"] n2["3: VAELoader"] n3["4: CheckpointLoaderSimple"] n4["5: LoraLoaderModelOnly"] n5["6: ModelSamplingAuraFlow"] n6["7: CLIPTextEncode"] n7["8: CLIPTextEncode"] n8["9: ConditioningZeroOut"] n9["10: EmptySD3LatentImage"] n10["11: KSampler"] n11["12: VAEDecode"] n12["13: SaveImage"] n6 -->|conditioning| n8 n5 -->|model| n10 n6 -->|positive| n10 n9 -->|latent_image| n10 n10 -->|samples| n11 n11 -->|images| n12

Data-flow DAG — scroll to zoom, drag to pan.

Nodes (13)#

1UNETLoadercore

unet_name = {{checkpoint_map[checkpoint]}} tmplweight_dtype = default

outputs: MODEL

2CLIPLoadercore

clip_name = qwen_3_4b.safetensorstype = lumina2device = default

outputs: CLIP

3VAELoadercore

vae_name = ae.safetensors

outputs: VAE

4CheckpointLoaderSimplecore

ckpt_name = {{checkpoint_map[checkpoint]}} tmpl

outputs: MODELCLIPVAE

5LoraLoaderModelOnlycore

model = {{loader_refs[checkpoint_profile[checkpoint].loader].model}} tmpllora_name = {{lora_map[lora]}} tmplstrength_model = {{lora_strength}} tmpl

outputs: MODEL

6ModelSamplingAuraFlowcore

model = {{lora_ref if lora != 'none' else loader_refs[checkpoint_profile[checkpoint].loader].model}} tmplshift = 3

outputs: MODEL

7CLIPTextEncodecore

text = {{constructed_prompt}} tmplclip = {{loader_refs[checkpoint_profile[checkpoint].loader].clip}} tmpl

outputs: CONDITIONING

8CLIPTextEncodecore

text = {{negative_prompt}} tmplclip = {{loader_refs[checkpoint_profile[checkpoint].loader].clip}} tmpl

outputs: CONDITIONING

9ConditioningZeroOutcore

conditioning = ◂ node 7 · out[0]

outputs: CONDITIONING

10EmptySD3LatentImagecore

width = {{aspect_map[aspect].width}} tmplheight = {{aspect_map[aspect].height}} tmplbatch_size = {{batch_size}} tmpl

outputs: LATENT

11KSamplercore

model = ◂ node 6 · out[0]positive = ◂ node 7 · out[0]negative = {{checkpoint_profile[checkpoint].negative}} tmpllatent_image = ◂ node 10 · out[0]seed = {{seed}} tmplsteps = {{distilled_schedule.steps if checkpoint_profile[checkpoint].distilled else steps}} tmplcfg = {{distilled_schedule.cfg if checkpoint_profile[checkpoint].distilled else cfg}} tmplsampler_name = res_multistepscheduler = simpledenoise = 1.0

outputs: LATENT

12VAEDecodecore

samples = ◂ node 11 · out[0]vae = {{loader_refs[checkpoint_profile[checkpoint].loader].vae}} tmpl

outputs: IMAGE

13SaveImagecore

images = ◂ node 12 · out[0]filename_prefix = isis/zimage-txt2img

Prompt construction#

template

{base_prompt}

Variables (1)#

base_prompt

{{prompt}} tmpl

Parameter banks (9)#

The prompt / configuration lookup tables this workflow keys into from its inputs — the vocabulary that turns a style / palette / preset selection into graph parameters.

`option_policy` (2)#

checkpoint

{"zimage-base-bf16": {"distribution": "service", "content": "neutral"}, "cyberrealistic-zimage-turbo-v8-bf16": {"distribution": "service", "content": "racy"}, "zimage-turbo-bf16-aio": {"distribution": "service", "content": "neutral"}}

lora

{"photoreal-betternudes-v3": {"distribution": "service", "content": "explicit"}, "realistic-snapshot-v5": {"distribution": "service", "content": "neutral"}, "breast-slider": {"distribution": "service", "content": "racy"}, "detail-slider": {"distribution": "service", "content": "neutral"}, "detaildaemonz-v1.1": {"distribution": "service", "content": "neutral"}}

`checkpoint_map` (3)#

zimage-base-bf16

z_image_bf16.safetensors

cyberrealistic-zimage-turbo-v8-bf16

cyberrealistic_zimage_turbo_v8_bf16.safetensors

zimage-turbo-bf16-aio

zimage_turbo_bf16_aio.safetensors

`checkpoint_profile` (3)#

zimage-base-bf16

{"loader": "unet", "distilled": false, "negative": ["8", 0]}

cyberrealistic-zimage-turbo-v8-bf16

{"loader": "unet", "distilled": true, "negative": ["9", 0]}

zimage-turbo-bf16-aio

{"loader": "aio", "distilled": true, "negative": ["9", 0]}

`distilled_schedule` (2)#

steps

8

cfg

1.0

`loader_refs` (2)#

unet

{"model": ["1", 0], "clip": ["2", 0], "vae": ["3", 0]}

aio

{"model": ["4", 0], "clip": ["4", 1], "vae": ["4", 2]}

`lora_ref` (2)#

50

`lora_map` (6)#

none

photoreal-betternudes-v3

zimage-photoreal-betternudes-v3.safetensors

realistic-snapshot-v5

zimage-realistic-snapshot-v5.safetensors

breast-slider

zimage-breast-slider.safetensors

detail-slider

zimage-detail-slider.safetensors

detaildaemonz-v1.1

zimage-detaildaemonz-v1.1.safetensors

`aspect_map` (8)#

square_1024x1024

{"width": 1024, "height": 1024}

portrait_2x3_840x1256

{"width": 840, "height": 1256}

landscape_3x2_1256x840

{"width": 1256, "height": 840}

portrait_3x4_888x1184

{"width": 888, "height": 1184}

landscape_4x3_1184x888

{"width": 1184, "height": 888}

portrait_9x16_768x1368

{"width": 768, "height": 1368}

landscape_16x9_1368x768

{"width": 1368, "height": 768}

ultrawide_21x9_1568x672

{"width": 1568, "height": 672}

`requires_families` (3)#

zimagezimage-cyberrealisticzimage-aio

Models & dependencies#

Models required (10)#

z_image_bf16.safetensors

cyberrealistic_zimage_turbo_v8_bf16.safetensors

zimage_turbo_bf16_aio.safetensors

qwen_3_4b.safetensors

ae.safetensors

zimage-photoreal-betternudes-v3.safetensors

zimage-realistic-snapshot-v5.safetensors

zimage-breast-slider.safetensors

zimage-detail-slider.safetensors

zimage-detaildaemonz-v1.1.safetensors

Output contract#

What a successful run of this workflow returns.

type

image

format

png

Taxonomy & routing#

How the control plane classifies this workflow — from the committed workflow-taxonomy-registry.json. It drives the consistency / control surface the agentic director can exercise over the workflow.

assetFamily

portrait-hero-image-bundle

outputPackageProfile

image-single-profile

controlModalities

model-locksampler-scheduler-lockseed-lockprompt-template-lock

consistencyDimensions

identitylightinglens

notes

Z-Image base or Turbo finetune text-to-image on RunPod serverless; a Turbo checkpoint selects the 8-step schedule.

Inputs (10)#

ComfyUI node graph (13)#

Nodes (13)#

Prompt construction#

Variables (1)#

Parameter banks (9)#

option_policy (2)#

checkpoint_map (3)#

checkpoint_profile (3)#

distilled_schedule (2)#

loader_refs (2)#

lora_ref (2)#

lora_map (6)#

aspect_map (8)#

requires_families (3)#

Models & dependencies#

Models required (10)#

Output contract#

Taxonomy & routing#

`option_policy` (2)#

`checkpoint_map` (3)#

`checkpoint_profile` (3)#

`distilled_schedule` (2)#

`loader_refs` (2)#

`lora_ref` (2)#

`lora_map` (6)#

`aspect_map` (8)#

`requires_families` (3)#