Which files do I need and where do they go?

Place the HiDream I1 Full UNet weights (e.g., hidream_i1_full_fp16.safetensors or the fp8 variant) in ComfyUI/models/unet. Load the HiDream text encoders via QuadrupleCLIPLoader (store them in your models/clip or text_encoders directory, matching your setup). Load a compatible VAE via VAELoader.

What do the official sampling settings do?

They match HiDream’s recommended configuration: steps=50, sampler=uni_pc, scheduler=simple, cfg=5.0, and shift=3.0 (set in ModelSamplingSD3). This balance typically yields strong detail and prompt adherence with stable convergence.

Can I change resolution, steps, or sampler?

Yes. Increase width/height (divisible by 8) for more detail at the cost of VRAM/time. Fewer steps (e.g., 30–40) are faster but may reduce fidelity. You can try other samplers/schedulers in KSampler, but UniPC + Simple is the recommended baseline for HiDream Full.

How do I reduce artifacts or oversharpening?

Use a clear negative prompt (e.g., "bad, ugly, jpeg artifacts"). Keep CFG near 5.0, and avoid overly long prompts. If images look harsh, try a slightly lower CFG or reduce resolution. Ensure the correct VAE is loaded to prevent color or contrast issues.

HiDream I1 Full - ComfyUI Workflow

This ComfyUI tutorial workflow generates high-quality images with the HiDream I1 Full model using a clean, SD3-style pipeline. It loads the HiDream UNet via UNETLoader, the matching text encoders with QuadrupleCLIPLoader, and prepares the canvas with EmptySD3LatentImage. Prompts are processed by CLIPTextEncode (positive and negative), sampling is configured through ModelSamplingSD3 (including the model-specific shift parameter), and KSampler performs 50 inference steps with the UniPC sampler and Simple scheduler for stable, detailed outputs. Finally, VAEDecode converts latents back to RGB and SaveImage writes the results to disk.

The included settings mirror HiDream’s official recommendations: steps=50, sampler=uni_pc, scheduler=simple, cfg=5.0, and shift=3.0. This combination is optimized for fidelity and prompt adherence, making it a strong starting point for cinematic, moody, or stylized scenes. Because the workflow uses EmptySD3LatentImage and QuadrupleCLIPLoader, it stays compatible with SD3-style prompts and latent geometry, while VAELoader ensures proper decoding for consistent color and detail.

API

Use this workflow from code

Every Comfy workflow is a JSON graph. The payload below is this workflow, exactly as ComfyUI runs it — fetch it from the URL, keep it in version control, or load it in ComfyUI and run it node by node.

hidream_i1_full.json

Fetching workflow JSON…

Run it from TypeScript or Python with the Comfy SDK. The same code targets Comfy Cloud or a ComfyUI you host yourself — only the base URL changes.

// Install (beta)
npm i @comfyorg/sdk

// Run "HiDream I1 Full" (TypeScript)
import { Comfy } from "@comfyorg/sdk";

const client = new Comfy({ apiKey: "comfyui-..." });

// This workflow, exported in API format (see note below)
const wf = await client.workflows.fromFile("hidream_i1_full_api.json");

wf.setInput("91", "text", "your prompt here");  // CLIPTextEncode

const job = await client.run(wf);
await job.getOutputs("90")[0].toFile("output.png");  // SaveImage

The SDK takes a workflow in API format: open this workflow in ComfyUI and use File → Export Workflow (API).

Comfy Cloud API access requires a plan with an API key.

SDK docs Get an API key

FAQ

Frequently Asked Questions