PromptZone - AI Prompts, Guides and Tools for Builders

Cover image for HiDream-I1 in ComfyUI: Model Files and Sampling Settings Guide
Ishaan Nair
Ishaan Nair

Posted on Edited on

HiDream-I1 in ComfyUI: Model Files and Sampling Settings Guide

HiDream-I1 is HiDream-ai's text-to-image model family with a 17-billion-parameter image model and Full, Dev, and Fast checkpoints. ComfyUI supports it natively through a workflow that loads a diffusion model, four text encoders, and a VAE. The official guide links the required Comfy-Org files and provides different sampling settings for each checkpoint. Project repository ComfyUI guide

What are the key facts about HiDream-I1 in ComfyUI?

Field Verified information
Developer HiDream-ai develops the model; ComfyUI supplies the documented integration. ComfyUI guide
Released April 7, 2025, for HiDream-I1. Project repository
Type Text-to-image diffusion model with Full and distilled Dev/Fast variants. Project repository
Size or parameters 17 billion parameters for HiDream-I1; the workflow additionally loads four text encoders. ComfyUI guide
License and access MIT transformer weights; component licenses apply separately. ComfyUI links repackaged files. Model card ComfyUI guide
Where it runs Local ComfyUI; the official guide also links Comfy Cloud templates. ComfyUI guide

What does the native HiDream-I1 ComfyUI workflow support?

The native workflow makes component selection explicit. Its QuadrupleCLIPLoader holds the four text encoders, while the diffusion model and VAE use separate loaders. That separation is useful when checking whether a saved graph really uses the intended files. ComfyUI guide

HiDream's model card describes photographic, cartoon, and artistic generation and publishes prompt-following evaluations. For a practical first test, choose a brief with a clear subject, a material, and a spatial relationship. Save the result before adding adapters or downstream processing, so you have a reference for later changes. Model card

Full, Dev, and Fast are documented choices with different sampling baselines. ComfyUI provides a graph for each, which gives you a concrete way to evaluate the variant that fits your task. Compare complete workflows rather than changing only the displayed model filename. ComfyUI guide

The sibling HiDream-I1 prompt evaluation guide covers how to judge those outputs. This page concentrates on making sure the graph and checkpoint agree before you interpret visual differences.

What hardware and configuration limits affect HiDream-I1?

The published model size does not make HiDream a small ComfyUI extension. The documented graph loads the image model plus four text encoders and a VAE. ComfyUI's guide flags substantial VRAM requirements and uses FP8 files in its examples to reduce the memory burden. ComfyUI guide

The guide lists more than 16 GB VRAM for its FP8 examples and more than 27 GB for the higher-precision versions. Treat those as the documentation's workflow guidance, not a measured requirement for every installation or a guarantee of throughput on a particular GPU. ComfyUI guide

Dev and Fast use different guidance conventions from Full. ComfyUI specifies cfg=1.0 for the distilled variants, while the developer's Diffusers example uses guidance_scale=0.0 for them. Follow the convention of the interface you are actually using. ComfyUI guide Project repository

Native support also depends on a compatible ComfyUI installation. When a loader is missing, check the application's version and startup output before searching for an unrelated custom node with a similar name. The official guide identifies native support as a prerequisite. ComfyUI guide

How do you configure HiDream-I1 files and sampling in ComfyUI?

Update ComfyUI and open the official HiDream-I1 tutorial. Choose the Full, Dev, or Fast workflow from that page. Keep the downloaded graph as an unchanged reference while preparing a working copy. ComfyUI guide

Download the shared components linked by Comfy-Org. The documented directory structure is:

Directory under ComfyUI/models Files
text_encoders clip_l_hidream.safetensors, clip_g_hidream.safetensors, t5xxl_fp8_e4m3fn_scaled.safetensors, llama_3.1_8b_instruct_fp8_scaled.safetensors
vae ae.safetensors from the distribution linked in the guide
diffusion_models The selected HiDream-I1 Full, Dev, or Fast diffusion file

These filenames and locations come from the official workflow instructions. Use the linked distribution so a generic filename such as ae.safetensors has a known source. ComfyUI guide Comfy-Org distribution

For example, with the Hugging Face CLI installed, download the Dev FP8 diffusion file from your ComfyUI directory and copy it into the loader's folder. This command handles only that diffusion file; install the shared components listed above as well. Comfy-Org distribution Dev FP8 file CLI documentation

hf download Comfy-Org/HiDream-I1_ComfyUI \
  split_files/diffusion_models/hidream_i1_dev_fp8.safetensors \
  --local-dir ./hidream-download
mkdir -p models/diffusion_models
cp ./hidream-download/split_files/diffusion_models/hidream_i1_dev_fp8.safetensors \
  models/diffusion_models/
Enter fullscreen mode Exit fullscreen mode

Load the selected graph. Confirm the diffusion file in Load Diffusion Model, all four encoder selections in QuadrupleCLIPLoader, and the VAE in Load VAE. Then check the sampling values against the official table below. ComfyUI guide

Variant Steps ComfyUI CFG ModelSamplingSD3 shift
Full 50 5.0 3.0
Dev 28 1.0 6.0
Fast 16 1.0 3.0

All values are from the corresponding ComfyUI workflows. The guide also offers lcm sampling and the normal scheduler as options; retain the downloaded workflow's configuration for your baseline before experimenting. ComfyUI guide

Enter a short original prompt, such as a blue glass bottle beside a folded white cloth on a wooden table. Run the graph and save both the image and workflow. Check the bottle, cloth, and relative placement before adding a more demanding subject.

When switching variants, use the matching graph or update the checkpoint, steps, CFG, and shift together. Preserve the same creative brief for comparison, but do not assume a checkpoint switch reproduces the same composition. Record the full workflow with every selected output.

How do HiDream-I1 and Qwen-Image ComfyUI workflows differ?

Qwen-Image also has a native ComfyUI workflow, but its documented component list uses a Qwen text encoder and its own VAE. A Qwen graph is therefore not interchangeable with HiDream's four-encoder graph. Compare the models through their respective official templates. Qwen ComfyUI guide

The ComfyUI pillar explains how to organize these experiments. For the model decision, compare accepted images and measured runtime while keeping each family's documented baseline intact.

What should you check when setting up HiDream-I1 in ComfyUI?

Is HiDream-I1 a custom ComfyUI extension?

HiDream-I1 is an image model with native ComfyUI support. Its workflow loads the diffusion model, four text encoders, and VAE through documented nodes. ComfyUI guide

Why does HiDream-I1 need four text encoders in ComfyUI?

The documented HiDream-I1 ComfyUI workflow integrates CLIP-L, CLIP-G, T5, and Llama text encoders. Select all four corresponding files in QuadrupleCLIPLoader. ComfyUI guide

What CFG should I use for HiDream-I1 Dev and Fast?

The official ComfyUI instructions specify CFG 1.0 for HiDream-I1 Dev and Fast. The developer's Python example uses guidance_scale=0.0 for these distilled variants, so follow the value documented for your interface. ComfyUI guide Project repository

Can I switch HiDream-I1 variants by changing only the checkpoint?

Switching HiDream-I1 variants requires checking the checkpoint, inference steps, CFG, and sampling shift together. The official ComfyUI configurations use 50/5.0/3.0 for Full, 28/1.0/6.0 for Dev, and 16/1.0/3.0 for Fast, in steps/CFG/shift order. ComfyUI guide

Sources

Top comments (0)