HEISS UI

Image model

How to run HiDream I1 locally.

HiDream I1 runs in ComfyUI with six files: the model, four text encoders (HiDream’s own CLIP-L and CLIP-G, T5-XXL and Llama 3.1 8B) and the Flux VAE. The fp8 model is 17.1 GB and needs a 16 GB graphics card or more, and the encoders add 15.9 GB of downloads. The model is MIT licensed.

Updated 29 Sep 20269 min read

Maker
HiDream.ai
Released
Apr 2025Full, Dev and Fast together
Licence
MITLlama 3.1 licence for one encoder
Memory
16 GB and upin fp8. Full precision 27 GB

Full, Dev or Fast.

HiDream I1 is a 17B sparse mixture-of-experts image model by HiDream.ai, released in April 2025. It reads the prompt through four text encoders at once. All three versions are the same size; they differ in how many steps they take and whether the negative prompt works.

  • Full is undistilled: 50 steps at CFG 5, and the negative prompt works. The best quality and the slowest.
  • Dev is distilled: 28 steps at CFG 1, no negative prompt.
  • Fast is distilled further: 16 steps at CFG 1, no negative prompt.

HiDream also made E1 and E1.1 for editing and a newer pixel-space model, O1. Those are different models and not covered here.

Files you need.

All six are in Comfy-Org/HiDream-I1_ComfyUI. The four encoders go into one QuadrupleCLIPLoader node.

  • Model hidream_i1_full_fp8.safetensors ComfyUI/models/diffusion_models/ or hidream_i1_dev_fp8 / hidream_i1_fast_fp8, same size. Full precision: 34.2 GB
    17.1 GB Download
  • Text encoder clip_l_hidream.safetensors ComfyUI/models/text_encoders/
    0.2 GB Download
  • Text encoder clip_g_hidream.safetensors ComfyUI/models/text_encoders/
    1.4 GB Download
  • Text encoder t5xxl_fp8_e4m3fn_scaled.safetensors ComfyUI/models/text_encoders/
    5.2 GB Download
  • Text encoder llama_3.1_8b_instruct_fp8_scaled.safetensors ComfyUI/models/text_encoders/
    9.1 GB Download
  • VAE ae.safetensors ComfyUI/models/vae/
    0.3 GB Download
ComfyUI/models 
models/
├── diffusion_models/
│   └── hidream_i1_full_fp8.safetensors
├── 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

Smaller files: GGUF

city96 has GGUF builds of Full, Dev and Fast. They load through the ComfyUI-GGUF nodes, which also have a QuadrupleCLIPLoader (GGUF) for a GGUF Llama encoder.

ModelQ8_0Q6_KQ5_K_MQ4_K_M
Full, Dev or Fast18.7 GB14.7 GB13.0 GB11.5 GB

Each version comes in the same sizes. The four fp8 encoders add 15.9 GB on top.

What fits your computer.

HiDream I1 at 1024 × 1024 with the fp8 encoders. ComfyUI runs the encoders first and moves them aside, so the model file sets the limit on the card, but plan on plenty of system RAM for the rest.

  • 8 GBNo

    Even Q4_K_M is 11.5 GB. Flux.1 as GGUF is the better pick here.

  • 12 GBOffloads

    Q4_K_M GGUF with part of it in system RAM. Expect it to be slow.

  • 16 GBTight

    Comfy’s docs put fp8 at 16 GB and up. Q6_K (14.7 GB) with the encoders offloaded is the common pick.

  • 24 GBFits

    fp8 at 1024 stays just under 24 GB with enough system RAM. Q8_0 GGUF works too.

  • 32 GBFits

    RTX 5090: full precision. Comfy’s docs put it at 27 GB and up.

  • MacUntested

    We found no ComfyUI reports from Macs; people asking get pointed to Draw Things. fp8 doesn’t save memory on a Mac, so it would be GGUF.

Set it up.

  1. Update ComfyUI

    QuadrupleCLIPLoader is a core node. If you can’t find it, your ComfyUI is too old. In a manual install:

    Terminal, in the ComfyUI folder 
    git pull
pip install -r requirements.txt

  2. Download the six files

    One model (Full, Dev or Fast), the four encoders and the VAE, from the list above. About 33 GB with the fp8 model.

  3. Put them in their folders

    Model in diffusion_models, all four encoders in text_encoders, VAE in vae. Then restart ComfyUI so the loaders list them.

  4. Open the matching template

    In ComfyUI’s template browser, pick HiDream I1 Full, HiDream I1 Dev or HiDream I1 Fast. Each one is set up for its version.

  5. Check the loaders

    Load Diffusion Model gets the HiDream file. QuadrupleCLIPLoader gets, in order, clip_l_hidream, clip_g_hidream, the T5 and the Llama file. Load VAE gets ae.safetensors.

  6. Write a prompt and run

    Use the negative prompt only with Full. The first run loads four encoders and takes a while.

Settings that work.

Comfy’s templates match HiDream’s own inference script. The shift is set with a ModelSamplingSD3 node.

VersionStepsCFGSamplerSchedulerShiftNegative
Full505uni_pcsimple3yes
Dev281lcmnormal6no effect
Fast161lcmnormal3no effect

Dev and Fast run at CFG 1, where a negative prompt does nothing. Comfy’s Dev and Fast templates still carry a negative prompt node; leave it empty.

Size
1024 × 1024
Shift node
ModelSamplingSD3
Encoders
4, one loader
VAE
ae.safetensors

The latent is EmptySD3LatentImage. HiDream’s script lists these sizes: 1024 × 1024, 768 × 1360, 1360 × 768, 880 × 1168, 1168 × 880, 1248 × 832 and 832 × 1248.

When it goes wrong.

A black image, with RuntimeWarning: invalid value encountered in cast
The Flux clip_l is loaded. Use clip_l_hidream and clip_g_hidream from Comfy-Org.
A black image from Full at the default weight type
Set a different weight type in Load Diffusion Model, such as fp8_e4m3fn.
QuadrupleCLIPLoader is missing
ComfyUI is too old. Update it; the node is built in.
The negative prompt does nothing
Dev and Fast run at CFG 1. Use Full at CFG 5 if you need a negative prompt.
Out of memory on a 24 GB card at full precision
The full model needs about 27 GB. Use the fp8 file or Q8_0 GGUF.
Token indices sequence length is longer than the specified maximum sequence length
A warning from the CLIP encoders, which stop at 77 tokens. It doesn’t stop the run.

The example workflows only give me black images. What’s wrong with my setup?

GitHub, ComfyUI issue #7715

Random seeds keep giving the same face. Is the model censored, and does an abliterated Llama help?

Hugging Face, HiDream-I1-Full

Questions.

Which text encoders does HiDream I1 need?

Four: HiDream’s own CLIP-L and CLIP-G (clip_l_hidream and clip_g_hidream), T5-XXL and Llama 3.1 8B Instruct. They load together through one QuadrupleCLIPLoader node. In fp8 they add up to 15.9 GB.

Why does HiDream give me a black image?

Usually because the Flux clip_l is loaded instead of clip_l_hidream. Use HiDream’s own CLIP-L and CLIP-G files from Comfy-Org. If Full still gives black images, change the weight type in Load Diffusion Model.

Should I use Full, Dev or Fast?

Full for the best quality and a working negative prompt, at 50 steps. Dev at 28 steps for most work. Fast at 16 steps for quick drafts. All three need the same memory.

How much VRAM does HiDream I1 need?

Comfy’s docs put fp8 at 16 GB and up and full precision at 27 GB and up. On 16 GB, a Q6_K GGUF with the encoders offloaded is the common choice. On 24 GB, fp8 and Q8_0 GGUF both work.

Can I use HiDream I1 commercially?

The model is MIT licensed, and its card says generated images can be used commercially. The Llama 3.1 8B encoder has its own Llama 3.1 Community License.

Does HiDream I1 run on a Mac?

We found no ComfyUI reports from Macs. People who ask get pointed to Draw Things. In ComfyUI it would need a GGUF, since fp8 files don’t save memory on a Mac.

Sources: HiDream-I1 Full model card, HiDream inference script, ComfyUI HiDream tutorial, Comfy-Org HiDream files, ComfyUI issue #7715, ComfyUI issue #7684, memory thread, GGUF encoder thread.

HEISS UI

Four encoders, the right four.

HEISS UI runs HiDream I1 Full, Dev and Fast on the ComfyUI you already have, each with settings that work.

  • No black images from the wrong encoder. It uses HiDream’s own text encoders and fetches them when they’re missing.
  • Missing parts, shown first. Each one listed with its size and a button. Get all checks free space, and downloads resume and are verified.
  • Drop in the file and it runs. It knows the model from the file itself, even renamed, and picks settings that work.
  • Failures that explain themselves. When a run runs out of memory, it says so and offers the fix.
Download HEISS UI GitHub

Free and open source. macOS, Windows and Linux. Runs on your ComfyUI.

HEISS UI with a gallery of generated images and the prompt composer at the bottom.
Model · Image Flux.1 Dev, Schnell and Krea Dev: smaller, with the same VAE and T5. Model · Image Qwen-Image Another large model, with strong text in images and an edit version. Hardware 24 GB graphics cards What runs on a 3090 or 4090, and what needs system RAM. Compare HEISS UI vs InvokeAI Two ways to make images locally, and where each one is the better pick.