Z-Image-Turbo ComfyUI template (image_z_image_turbo): three files, 8 steps at CFG 1 with a zeroed negative, versus the INT8 template

Z-Image-Turbo text-to-image in ComfyUI: set up and run the official template Scope. This guide reads the official image_z_image_turbo.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 baaf937f...f1d3, 27,390 bytes). It compares that file with the sibling image_z_image_turbo_int8.json at the same commit. Both graphs were traced statically through their wrapper and subgraph. Nothing was run and no weights were downloaded. 1. Prerequisites - ComfyUI with subgraph support. All nodes are core nodes; the file was saved by frontend 1.42.15 with comfy-core 0.3.64 and 0.3.73 node versions. Those are saved strings, not a tested minimum. No custom node pack is needed. - Three files. The template's Model link note names the repo Comfy-Org/z_image_turbo and gives these sizes: models/diffusion_models/z_image_turbo_bf16.safetensors (11.46 GB) models/text_encoders/qwen_3_4b.safetensors (7.49 GB), loaded by CLIPLoader with type lumina2 models/vae/ae.safetensors (319.8 MB) 2. The controls you edit The root node "Text to Image (Z-Image-Turbo)" (node 57) wraps subgraph f2fdebf6-dfaf-43b6-9eb2-7f70613cfdc1 and promotes the prompt, width, height, seed (with its control mode), steps and the three file names. Inside the subgraph: - KSampler 3: 8 steps (promoted), CFG 1, sampler res_multistep, scheduler simple. The seed is promoted with control randomize. - ModelSamplingAuraFlow 11 applies shift 3 to the model. - EmptySD3LatentImage 13 is 1024 x 1024, batch 1, with width and height promoted. - The negative input is ConditioningZeroOut 33 applied to the positive prompt encoding. There is no negative prompt to write. Output goes through VAEDecode 8 to SaveImage 9 with prefix z-image-turbo. 3. How these settings relate to the model card The Z-Image-Turbo card record ("Z-Image-Turbo: 9 steps (8 forwards), guidance 0, 16 GB, vs base Z-Image") gives two statements. Its model table describes Turbo as 8 steps without CFG. Its diffusers example uses num_inference_steps=9, which the card says is 8 DiT forwards, with guidance_scale=0.0. The template sets 8 sampler steps at CFG 1 with a zeroed negative. CFG 1 is ComfyUI's no-guidance setting, so the template matches the card's 8-step, no-CFG description. Do not enter 9 into the KSampler expecting the diffusers behaviour; the two step counts are defined by different pipelines, and no run here compared them. The card does not state the shift 3 and res_multistep choices; they come from the template. 4. bf16 versus INT8 template image_z_image_turbo_int8.json uses the same subgraph id and the same sampler settings: 8 steps, CFG 1, res_multistep, simple, shift 3, 1024 x 1024, ae.safetensors. It swaps two files: the diffusion model z_image_turbo_int8_convrot.safetensors instead of z_image_turbo_bf16, and the text encoder qwen_3_4b_fp8_mixed.safetensors instead of qwen_3_4b. Only the saved seed also differs. Choose the template whose files you have; the graph is otherwise the same. The earlier record "A missing ComfyUI UUID node may be an embedded subgraph" covers import errors for the INT8 file. 5. Run it 1) Put the three files in the folders above, or the two INT8 files plus ae.safetensors for the INT8 template. 2) Open the template. 3) Write the prompt on the wrapper. 4) Set width and height, and the seed if you want a repeatable result. 5) Run. Expect one image saved with prefix z-image-turbo at 1024 x 1024 by default. 6. Failure cases and checks - A missing UUID-named node is the embedded subgraph; update ComfyUI rather than searching for a custom pack. - Mixing files between the templates: the bf16 template's loaders name qwen_3_4b and z_image_turbo_bf16. Select the INT8 file names on the wrapper if those are the files you have. - A non-repeatable output: the seed is set to randomize; set it to fixed and record it. - Writing a negative prompt has nowhere to go: the negative is the zeroed positive by design. - Raising CFG above 1 enables guidance, which the card says Turbo should not use; the template does not test this. Evidence and rights. Evidence status: source_inspected, not_run, unreviewed. This is shared-founder desk curation by the tutorial lane. The workflow repository declares the MIT licence; that covers this analysis of the templates, not the model weights, text encoder, VAE or generated images, which keep their own terms. The card's Apache-2.0 statement is reported by that record and was not re-checked here.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_z_image_turbo.json",
      "note": "Exact pinned template; node byte ranges and excerpt hashes are in the source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_z_image_turbo_int8.json",
      "note": "Sibling INT8 template at the same commit (existing native source); node byte ranges in the source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/LICENSE",
      "note": "Repository MIT licence at the same commit (existing native source)."
    }
  ]
}