MiniMax H3 Image ComfyUI workflows: what is in the repo, what it needs, and how to stay safe

MiniMax H3 Image ComfyUI workflows are three community files on Hugging Face (Apache-2.0) for 8, 12 or 16 GB NVIDIA cards, needing about 38 GB of downloads; they are not from MiniMax, and this site does not run them: its editor changes one photo in about 37 seconds for 10 credits. Free with sign-up: 10 credits = 1 image, then from $0.13 per image on a monthly plan.

Your edited photo appears here (this site’s editor)
Studio background edit made on this site’s editor earlier: the same woman on a plain light-grey studio backdrop
Input photo (AI-generated, fictional): a smiling woman with curly brown hair in a green sweater, in a parkYour photo
Type your own change, or leave this empty and pick a one-tap edit below.0/400

Style changes (cartoon, 3D, anime) are not reliable yet. If the result looks unchanged, try again or describe the change more specifically.

One-tap edits

  • Not affiliated with MiniMax, Comfy Org or the repo’s author
  • The editor here is a different model
  • Free with sign-up: 10 credits = 1 image
workflow files: for 8, 12 and 16 GB of VRAM
3
of model downloads (our sum: 21 + 15.7 + 1.2 GB)
≈ 38 GB
author’s own time for one 16 MP image on a 16 GB card
66 s
licence of the workflow repo (not the weights)
Apache-2.0

What are the MiniMax H3 Image ComfyUI workflows?

A model repo on Hugging Face called vyrekai/minimax-h3-image-comfyui-workflows. It holds three ComfyUI workflow files, one each for 8, 12 and 16 GB of VRAM, a prompt guide written for AI chatbots, and a folder of example images. The README says all three make the same 16 megapixel image from the same seed, and differ only in how the picture is decoded, so smaller cards are slower rather than worse. The workflows include a switch between a 4 megapixel draft and the full size, an aspect-ratio picker, an optional prompt-enhance step for text-to-image, and ten reference-picture slots that are switched off by default. These are the author’s descriptions; we did not run them.

  • Three JSON workflow files: 8 GB (tiled decode 512), 12 GB (tiled 768), 16 GB and up (normal decode)
  • A prompt guide and eight example images, the largest at 16 MP
  • Six sampling steps, CFG 1.0, no negative prompt
Input photo (AI-generated, fictional): a smiling woman with curly brown hair in a green sweater, in a parkStudio background edit made on this site’s editor earlier: the same woman on a plain light-grey studio backdrop

What do the workflows need to run?

The README asks for an NVIDIA card with at least 8 GB of VRAM (the author saw at most 3.7 GB used for a 16 MP image), 24 GB of system memory (32 GB recommended), about 38 GB of free disk, and a recent ComfyUI: the author tested 0.45.3 and warns that older versions may lack the H3 Image nodes. It says no custom nodes are needed and mentions no API key. The models are separate downloads from the Comfy-Org repo: a 21 GB diffusion model, a 15.7 GB text encoder and a 1.2 GB VAE, plus an optional 7.3 GB prompt upsampler. Our sum is 37.9 GB, or 45.2 GB with the upsampler.

  • NVIDIA GPU, 8 GB VRAM minimum; RAM 24 GB (32 GB advised); about 38 GB of disk
  • No custom nodes and no API key mentioned
  • Four model files from Comfy-Org, listed in the table below
Input photo (AI-generated, fictional): a smiling woman with curly brown hair in a green sweater, in a park

Who made them, and how far should you trust them?

The repo belongs to a Hugging Face user named vyrekai, shown as VYREK, with no bio, no organisation and five followers; the profile links to an X and a GitHub account and lists two public repos, this one and a video workflow. It is not a MiniMax or Comfy Org account, and we do not know who is behind it. The page shows no date, but the profile lists the repo as updated about four days before our read. The official route is different: ComfyUI’s own docs describe H3 Image as a cloud partner node that runs on MiniMax’s servers with no local files. We found no report of a problem with this repo and did not open its files, which is not proof of safety.

  • Community account, not MiniMax and not Comfy Org
  • Official alternative: ComfyUI’s “Text to Image (API)” and “Image Edit (API)” partner nodes
  • Licence: Apache-2.0 on the workflow files; MiniMax’s own licence on the weights
Input photo (AI-generated, fictional): a smiling woman with curly brown hair in a green sweater, in a parkStudio background edit made on this site’s editor earlier: the same woman on a plain light-grey studio backdrop

Is there a ComfyUI workflow on this site?

No. This site does not run ComfyUI or H3 Image: its editor takes one photo and a written request of up to 400 characters, runs a different model, and makes one 1024 × 1024 image in about 37 seconds for 10 credits. If you want to edit a photo without setting up a 38 GB install, that is the nearest thing. For a related model people run on their own card, see our LongCat Video page; for a hosted image model, our Seedream 5.5 page; and for the character-swap model that uses the same base, our H3 Image character swap page.

  • Photo in, photo out; no workflow to install
  • Not H3 Image and not ComfyUI
  • Still images only: no video
Edit a photo here
Input photo (AI-generated, fictional): a smiling woman with curly brown hair in a green sweater, in a park

What we read, and what we could not confirm

The repo card (the author’s own claims, not tested by us)

  • Three workflow files make a 16 MP final image and a 4 MP draft. The 8 GB file decodes in tiles of 512, the 12 GB file in tiles of 768 and the 16 GB file normally. An “Enhance prompt” switch uses a Qwen3.5 9B upsampler for text-to-image only and is off in the 8 GB file. Sampler: SamplerLCM with the simple scheduler, six steps, CFG 1.0.Source: Hugging Face repo card ·
  • Timings on an RTX 5070 Ti (16 GB) with 24 GB of RAM, the 12 GB and 8 GB runs simulated: a 16 MP text-to-image takes 66 s, 63 s and 60 s for the 16, 12 and 8 GB files, at 4.4, 4.8 and 5.7 s per step. Four reference pictures take 228 s at 16 MP or 156 s at 8 MP on the 8 GB file. The author says real 8 and 12 GB cards are slower chips, so expect longer.Source: Hugging Face repo card ·
  • Tips: if decoding runs out of memory, lower the tile size to 384 or 256 and close other GPU apps; switching between 4 MP and 16 MP with the same seed gives a different picture; write long, photo-like prompts, because short ones tend to give a glossy, doll-like face.Source: Hugging Face repo card ·
  • The repo has no discussion threads and shows no download count. The profile shows 85 likes on the repo and a second public repo, a video workflow.Source: Hugging Face repo and profile ·

Other pages we read

  • The model files come from the Comfy-Org repo, which repackages MiniMax’s H3 Image for ComfyUI under MiniMax’s community licence and links official Text to Image and Image Edit templates. It states no ComfyUI version and no hardware needs.Source: Comfy-Org on Hugging Face ·
  • The licence text applies worldwide excluding the EU, the UK, the Republic of Korea and the United States, is dated November 10, 2026, and needs MiniMax’s prior written authorization above 1 million USD of annualized revenue or for hosted services. Its acceptable use policy bans impersonating a person without consent. Not legal advice.Source: MiniMaxAI LICENSE ·
  • ComfyUI’s own docs describe H3 Image as a cloud partner node with two templates, Text to Image (API) and Image Edit (API). It runs on MiniMax’s servers through a Comfy API account, needs no model download or local GPU, and needs ComfyUI 0.45.1 or newer. The page states no price.Source: ComfyUI docs ·
  • A second community repo, the H3 Image character swap, ships a 40 GB checkpoint and its own ComfyUI workflow, and quotes about 140 GB of VRAM for a full diffusers run.Source: Hugging Face model card (akhaliq) ·

Staying safe with downloaded workflows and models

  • Loading a pickle file can run code. Hugging Face advises loading models only from users and organisations you trust, and says its scanner is not foolproof and that checking is the user’s responsibility.Source: Hugging Face docs (pickle) ·
  • Safetensors is described as a format for storing tensors safely, as opposed to pickle. The Comfy-Org files named above are .safetensors files.Source: Hugging Face docs (safetensors) ·
  • Custom nodes are code. Comfy Org’s January 2025 post names two past custom-node incidents, ComfyUI_LLMVISION and ultralytics, and says it plans a way to remotely disable bad nodes. This repo says it needs no custom nodes.Source: Comfy Org blog ·

Our own checks

  • We added up the author’s file sizes: 21 + 15.7 + 1.2 = 37.9 GB, or 45.2 GB with the optional 7.3 GB upsampler. That matches the author’s “about 38 GB” and “45 GB”.Source: Our own arithmetic ·
  • We ran the three prompts in the cards below through this site’s own request filter, in English and Chinese: all pass, at 279, 256 and 218 characters of the 400 allowed. We did not generate them, so no result is shown.Source: This site’s request filter, tested ·

Not confirmed

  • Whether the workflows work as described: we ran nothing and downloaded nothing, and every speed and VRAM figure above is the author’s. We also did not open the JSON files or the prompt guide.Source: Hugging Face repo card ·
  • The repo’s creation or update date (relative dates only), who vyrekai is, and the lowest ComfyUI version that works: the author tested 0.45.3, while ComfyUI’s docs say 0.45.1 for the API nodes.Source: Hugging Face profile ·
  • Whether an 8 GB or 12 GB card really reaches these speeds, and what a hosted route costs: ComfyUI’s docs state no price.Source: ComfyUI docs ·

Downloads

The model files the workflows need

File names, sizes and folders as listed on the repo card, read 2026-11-19; the two total rows are our own sums. Models are downloaded from the Comfy-Org repo.
RoleFileSizeFolder
Diffusion modelminimax_h3_image_pruned_int8_convrot.safetensors21 GBmodels/diffusion_models/
Text encoderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors15.7 GBmodels/text_encoders/
VAEminimax_h3_image_vae_fp16.safetensors1.2 GBmodels/vae/
Prompt upsampler (optional)qwen3.5_9b_w4a8_gptq.safetensors7.3 GBmodels/text_encoders/
Total without the upsampler (our sum)21 + 15.7 + 1.237.9 GBabout 38 GB of free disk
Total with the upsampler (our sum)37.9 + 7.345.2 GBabout 45 GB of free disk

Set-up, as the README describes it

How to set up the workflows (we did not run them)

  1. Check your hardware and the licence first

    You need an NVIDIA card with at least 8 GB of VRAM, 24 GB of system memory and about 38 GB of free disk. Read MiniMax’s licence for your country and use: the EU, the UK, South Korea and the US are outside its territory, and hosted services or revenue above 1 million USD need written permission.

  2. Update ComfyUI and download the models

    Use a recent ComfyUI (the author tested 0.45.3). Get the diffusion model, text encoder and VAE from the Comfy-Org repo, which is the page the README links to, and put them in the three folders in the table above. The upsampler is optional.

  3. Pick the workflow file for your card

    Choose the 8, 12 or 16 GB file, open it in ComfyUI, and leave the ten reference slots switched off unless you want to edit pictures. Start with the 4 MP draft, then switch to the full 16 MP size; the same seed gives a different picture at the other size.

  4. Write a long, photo-like prompt

    The README says short prompts tend to give a glossy, doll-like face, so describe the lens, light, clothes and setting. If decoding runs out of memory, lower the tile size to 384 or 256 and close other GPU apps.

  5. Check files before you trust them

    Open a workflow in a text editor before you load it, add custom nodes only from sources you trust, and keep to .safetensors model files. Hugging Face’s scanner is not foolproof, and we have not opened this repo’s files.

Prompts

Photo-edit prompts for this site’s editor (not run)

These three prompts are our own wording for this site’s photo editor, which runs a different model from H3 Image and has no workflows to install. We did not generate them for this page, so there is no result to show; we only checked on 2026-11-19 that they pass the editor’s request filter. They keep the face and name no brand or person. We have not tested them in H3 Image or ComfyUI. Results vary with the photo.

A long, photo-like description (not run)

MiniMax H3 Image: what it is and what it costs

Not run: we spent no credits on this page. The repo’s README advises long, photo-like prompts, so this one is written that way. Checked only against this site’s request filter on 2026-11-19.

  • Long, photo-like description

    Keep my face and hair exactly the same. Redraw the scene as a candid photo taken on a 35mm lens: soft overcast daylight, a green knit sweater with visible stitches, a muted park path behind me in gentle blur, natural skin texture, slight film grain. No text, logos or watermarks.

Relight as golden hour (not run)

The H3 Image character swap model

Not run. Keeps the person and the park, changes only the light.

  • Relight as golden hour

    Keep my face, hair and clothes exactly the same. Relight the photo as golden hour: low warm sun from the left, a soft rim light on my hair, long gentle shadows on the path and slightly warmer colours. Keep the park background. No text, logos or watermarks.

Black-and-white film portrait (not run)

LongCat Video, another open model people run themselves

Not run. A style change that keeps face, hair and clothes.

  • Black-and-white film portrait

    Keep my face, hair and clothes the same. Turn the photo into a black-and-white film portrait: rich contrast, fine grain, soft window-style light from the side and a shallow depth of field. No text, logos or watermarks.

Pricing on this site

Credit packs or a monthly plan.

Credits come in one-time packs ($4.99 for 50 credits, $9.99 for 120 credits, $19.99 for 300 credits); each image costs 10 credits. Or pick a plan from $9.90 a month (500 credits), cancel anytime.

  • Basic · $9.90/month500 credits a month · about 50 images · $0.20 per image
  • Pro · $19.90/month1500 credits a month · about 150 images · $0.13 per image

Billed every month and renews automatically. Cancel anytime in Settings; you keep the plan until the end of the month you paid for. Each month's credits are for that month; unused ones expire.

Refund policy

FAQ

Questions, answered

What are the MiniMax H3 Image ComfyUI workflows?

Three ComfyUI workflow files, for 8, 12 and 16 GB of VRAM, in the Hugging Face repo vyrekai/minimax-h3-image-comfyui-workflows, with a prompt guide and example images. They make a 16 megapixel image with MiniMax’s open H3 Image model. Everything on this page about how they work is the author’s claim.

Can I run the MiniMax H3 Image ComfyUI workflow on 8 GB of VRAM?

The author says yes: the 8 GB file needed at most 3.7 GB of VRAM for a 16 MP image in their test, at 60 seconds on a simulated 8 GB limit. The author also says real 8 GB cards are slower chips, and you still need 24 GB of system memory and about 38 GB of disk. We did not test it.

Which files do I have to download for the workflow, and how big are they?

A 21 GB diffusion model, a 15.7 GB text encoder and a 1.2 GB VAE, about 38 GB in total by our sum, plus an optional 7.3 GB prompt upsampler that brings it to about 45 GB. They come from the Comfy-Org repo on Hugging Face.

Do the workflows need custom nodes or an API key?

The README says no custom nodes are required and mentions no API key. It does say you need a recent ComfyUI, tested on 0.45.3, because older versions may lack the H3 Image nodes. We did not open the workflow files to check.

Is the MiniMax H3 Image ComfyUI workflow safe to download?

We found no report of a problem and did not open the files, so we cannot vouch for them. Hugging Face advises loading models only from sources you trust and says its scanner is not foolproof; Comfy Org has named past incidents with custom nodes. Prefer safetensors files, skip unneeded custom nodes, and open the workflow in a text editor first.

Is this the official MiniMax H3 Image ComfyUI workflow?

No. It comes from a community account, not MiniMax or Comfy Org. The official route in ComfyUI’s own docs is a cloud partner node, “Text to Image (API)” and “Image Edit (API)”, which runs on MiniMax’s servers and needs ComfyUI 0.45.1 or newer.

Can I use the workflow commercially?

The workflow files are under Apache-2.0, but the model weights fall under MiniMax’s H3 Image community licence. That licence excludes the EU, the UK, South Korea and the United States from its territory, and needs MiniMax’s written permission above 1 million USD of yearly revenue or for hosted services. Read the licence; this is not legal advice.

Can I run these workflows here, and is the editor free?

No: this site does not run ComfyUI or H3 Image. Its own editor edits one photo with a different model. Partly: new accounts get 10 free credits after signing in, enough for 1 free image (once per account, valid for 7 days, while the day's new-account gifts last). Each image costs 10 credits, from $0.13 per image on a monthly plan.

Just need one photo edited, with no 38 GB install?

Upload it, write the change, and download the result in about 37 seconds.

Edit a photo here

Independent site. Not affiliated with MiniMax, Comfy Org, the repo’s author or Hugging Face, and this site does not run H3 Image or ComfyUI. Facts come from the pages linked under Sources, read on 2026-11-19; the workflows’ claims are the author’s and not tested by us. The licence summary is not legal advice.