• Skip to main content
  • Skip to secondary menu
  • Skip to primary sidebar
  • Skip to footer
Best Free Web Resources — local AI tools and hardware guides

Best Free Web Resources

Local AI Tools & Digital Engineering Resources

What can I run?
  • AI Tools
    • Start: Local LLMs
    • Ollama Models & VRAM
    • LM Studio Models
    • GPU Comparison
    • AI Hardware Matrix
    • Local LLM Directory
    • VRAM Calculator
    • Hardware Compare
    • Local Model Recommender
    • API Pricing & ROI
  • Local AI Guides
  • Web Resources
    • Inspiration
    • Photography
    • Games
    • SEO
    • Tutorial
      • Coding
      • WordPress
    • Resume/CV
    • Graphics
      • Logo
    • Freebies
Cover illustration: MiniMax H3 Image in ComfyUI

MiniMax H3 Image in ComfyUI: Files, VRAM, Settings and the Licence

November 17, 2026 by BFWR Editorial

MiniMax H3 Image runs in ComfyUI 0.45.0 or later on core nodes, from four files that total 56.63 GB in the stock templates, or 29.41 GB in the smallest combination Comfy-Org publishes. Neither MiniMax nor ComfyUI gives a VRAM figure for it; the templates start at 4 megapixels so smaller cards do not run out of memory. It does not use flux2-vae.safetensors. And its licence is not the H3 video licence: the same four territories are excluded, but written permission is needed above US$1 million in yearly revenue, not US$20 million, and for any hosted service.

Below: the exact files, folders and node names, the settings, the smallest download, reported errors and the licence. BFWR has not run MiniMax H3 Image. Everything here comes from MiniMax’s model card and licence, ComfyUI’s documentation, templates and source code, the published file listings and GitHub issue reports. GB means a billion bytes throughout. For the video model, see our MiniMax H3 video guide.

Quick answer

  • Needs: ComfyUI 0.45.0+ (released 10 November 2026) and template pack 0.11.104+.
  • Files: minimax_h3_image_pruned_int8_convrot.safetensors in models/diffusion_models/, qwen3vl_32b_minimax_h3_int8_convrot.safetensors and qwen3.5_9b_w4a8_gptq.safetensors in models/text_encoders/, minimax_h3_image_vae_fp16.safetensors in models/vae/.
  • Settings: euler, simple, 6 steps, CFG 1, denoise 1, no negative prompt. Best results around 16 MP (4096×4096).
  • VAE: its own 24-channel VAE. flux2-vae.safetensors belongs to FLUX.2 and FLUX.3 Klein templates, not this one.
  • Most common fix: pure colour noise from a checkpoint you converted yourself. Use the Comfy-Org pruned files.
  • Licence: no rights in the US, EU, UK or South Korea; written authorisation above US$1M annualised revenue or for any hosted service.

What MiniMax H3 Image is

MiniMax’s card describes one checkpoint for text-to-image and editing: “Generation and editing are unified in one checkpoint”. It targets about 16 MP, from 4096×4096 to 6272×2688, and “The checkpoint is distilled to 6 sampling steps (6 NFEs) with no classifier-free guidance.” The diffusion transformer is 33.1B parameters with “the H3 architecture without the audio layers, and it uses the MiniMax-H3 text encoder unchanged.” Edits accept up to 16 reference images, written as <Picture 1>, <Picture 2> and so on in the prompt.

Do not confuse it with the older MiniMax H3: Text to Image template, which makes stills with the video model. The new ones are MiniMax-H3-Image: Text to Image and MiniMax-H3-Image: Image Edit. Two more with (API) in the title send your prompt to MiniMax’s paid service and use none of the files below; our templates guide shows how to tell local and cloud templates apart.

The files, and where they go

These are the files each template’s loader nodes name, with sizes from the Hugging Face file listings on 15 November.

Node File Folder Repository Size
UNETLoader minimax_h3_image_pruned_int8_convrot.safetensors diffusion_models/ Comfy-Org/MiniMax-H3-Image 20.98 GB
CLIPLoader, type minimax qwen3vl_32b_minimax_h3_int8_convrot.safetensors text_encoders/ Comfy-Org/MiniMax-H3 27.14 GB
CLIPLoader, type minimax (prompt expander) qwen3.5_9b_w4a8_gptq.safetensors text_encoders/ Comfy-Org/Qwen3.5 7.34 GB
VAELoader minimax_h3_image_vae_fp16.safetensors vae/ Comfy-Org/MiniMax-H3-Image 1.17 GB
ComfyUI/models/
├── diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors
├── text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors
├── text_encoders/qwen3.5_9b_w4a8_gptq.safetensors
└── vae/minimax_h3_image_vae_fp16.safetensors

ComfyUI’s tutorial lists the encoder as 25.28 GB and the template notes list the expander as 6.84 GB in one place; those are the same files counted in GiB. One template note also shows a file called qwen3.8_27b_w4a8.safetensors in its folder diagram. No loader uses that name, and no such file is in Comfy-Org’s repositories, so do not go looking for it.

Is it the FLUX.2 VAE?

No. Google shows people searching for flux2-vae.safetensors alongside H3 Image, but neither H3 Image template loads it. MiniMax’s card describes the image VAE as “a CNN encoder and a 2D ViT decoder (24 layers), with 16x spatial downsampling and 24 latent channels”, and its vae/config.json says latent_channels 24. The FLUX.2 VAE config in Black Forest Labs’ klein repository says 32. On GitHub, flux2-vae.safetensors appears in Comfy-Org’s FLUX.2, FLUX.3 Klein, Ernie, Ideogram 4 and other templates; if that is what you were setting up, see our FLUX.3 Klein guide. Abiray’s GGUF card puts it plainly: “Do not use standard 4-channel SD/Flux VAEs”.

The real download size, and the smallest set

Combination Transformer Encoder Expander VAE Total
Stock templates pruned_int8_convrot 20.98 int8_convrot 27.14 7.34 1.17 56.63 GB
Stock, no expander 20.98 27.14 none 1.17 49.29 GB
Smaller encoder 20.98 nvfp4_awq 15.69 none 1.17 37.83 GB
Smallest from Comfy-Org pruned_w4a8 12.55 nvfp4_awq 15.69 none 1.17 29.41 GB
Pruned BF16 pruned_bf16 40.23 bf16 51.51 none 1.17 92.91 GB

ComfyUI’s template catalogue lists 66.57 GB for each H3 Image template, but the four files the templates load add up to 56.63 GB.

Comfy-Org’s H3 Image card lists both qwen3vl_32b_minimax_h3_int8_convrot.safetensors and qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors for text_encoders/; the templates pick the int8 one. Pick the other in CLIPLoader to save 11.45 GB. The H3 Image copies are byte-identical to the ones in Comfy-Org/MiniMax-H3 (same file hashes), so if you already run the H3 video templates you have the nvfp4 encoder and need only the transformer and VAE, 22.15 GB. The w4a8 transformer is published but no template uses it; select it in UNETLoader.

The text-to-image template “ships with the expander switched off”, so the 7.34 GB expander is optional there; the edit template ships with it on. Comfy chose it over MiniMax’s own 17.9 GB prompt expander “so the expander fits beside the 21 GB diffusion model on one GPU.”

Outside ComfyUI, a plain hf download MiniMaxAI/MiniMax-H3-Image fetches the original checkpoint, a diffusers copy and the text encoder, “about 204 GB in total” by MiniMax’s GitHub README; use its include patterns instead.

How much VRAM it needs

None of MiniMax’s card, ComfyUI’s tutorial or Comfy-Org’s README states a VRAM figure, so this is file arithmetic and other projects’ guidance, not a measurement. ComfyUI loads the encoder, the transformer and the VAE in stages, so the 56.63 GB total is not a VRAM requirement. File sizes also exclude the memory used while sampling, which grows with the canvas.

  • ComfyUI: the text-to-image template “opens at 4 MP on a square canvas, deliberately below the model’s native resolution so low-VRAM machines do not run out of memory.” The template note warns that it “doesn’t work well in low resolution”, so raise the size if the card has room. For edits: “If an edit runs out of memory, enable the Resize Image/Mask node in the template and lower its megapixels before sampling.”
  • diffusers, BF16 with automatic offload: “A 4736×3552 image then peaks at about 65 GB.”
  • SGLang: for “RTX 6000 Ada, RTX 5090 / 4090 and listed 12–48 GB cards” it recommends layerwise offload, which streams weights from system RAM. Those are derived recipes with “no hardware latency measurement or capacity guarantee”, and “Offload recipes assume a 256 GB-class host and local NVMe.”

The levers are the canvas size, the smaller encoder and the w4a8 transformer. Our ComfyUI models page lists memory for other image models.

Settings and nodes

Values from the template JSON and ComfyUI’s tutorial:

Node Value
KSampler 6 steps, CFG 1, euler, simple, denoise 1
ConditioningZeroOut feeds the negative input; there is no negative prompt
MiniMaxH3ImageConditioning (display name MiniMax H3 Image Conditioning) takes clip, the optional vae, prompt, width, height and image_1 onwards; outputs the conditioning and the empty latent
ResolutionSelector (text-to-image) 1:1, 4 MP, multiple of 32
ResizeImageMaskNode (edit) scales the source to 4 MP, then to a multiple of 32
TextGenerateV2 + JsonExtractString the prompt expander; returns expanded_caption
ComfySwitchNode expander off in text-to-image, on in edit

MiniMaxH3ImageConditioning is the node people search for by name. In ComfyUI’s source, a width or height of 0 “uses the width of image_1 after its reference scaling, or 2048 without images”, and the VAE input matters for edits: “With it the pictures are also anchored as reference latents, without it they only condition the text encoder.” To add references, the edit template’s note says to “add extra Load Image nodes” and plug each into both the Generate Text node and the conditioning node.

The recommended 16 MP sizes are 4096×4096 (1:1), 4736×3552 (4:3), 5024×3360 (3:2), 5472×3072 (16:9) and 6272×2688 (21:9), with portrait sizes swapped. ComfyUI says “Write everything you want, including things to avoid, into the positive prompt.”

ComfyUI listens only on 127.0.0.1 by default. Keep it that way: under this licence, letting other people use your instance may count as a hosted service (see below).

Common failures

These are from ComfyUI’s GitHub issues on 15 November. Most fixes are users’ reports, so treat them as things to try.

  • Pure colour noise with no error. On #17692, a user merged MiniMax’s own transformer shards into one file; it loaded and produced noise. ComfyUI implements only the pruned layout Comfy-Org ships. A pull request adding a warning (#17705) was not merged on 15 November. Use the Comfy-Org pruned files.
  • Noise with the int8 files. For the same int8_convrot format on the video model, a ComfyUI collaborator wrote “Update to cu130 minimum to use the int8 models.” (#15419), and Comfy-Org’s video card says to prefer it “if you are able to use pytorch with cu130”. RTX 50-series noise with comfy-kitchen kernels is tracked on #17098.
  • The expander returns commas. Generate Text needs the Qwen3.5 9B file, not the Qwen3-VL 32B encoder. #16558 reports “Every output is ~500 commas” when the 32B encoder is wired into it, because that file has no output head.
  • linear(): argument 'weight' must be Tensor, not NoneType in Generate Text. Reported on Qwen3-VL 8B and Gemma 4 since v0.44 (#17610). If it hits the expander, switch the expander off and write a full prompt yourself.
  • Missing nodes or templates. Update to 0.45.0 or later.

GGUF files

Abiray/MiniMax-H3-Image-GGUF converts Comfy-Org’s pruned BF16 file to Q2_K (6.88 GB) through Q5_K_M (14.08 GB), loaded with the Unet Loader (GGUF) node from city96’s ComfyUI-GGUF and a workflow in the repository, image_minimax_h3_image_t2i_gguf.json. Its VRAM tiers are the uploader’s own recommendations, and ComfyUI-GGUF had no commit in the 90 days to 15 November.

The licence: not the same as H3 video

Hugging Face shows license:other. The text is the MiniMax-H3-Image Community License Agreement, dated 10 November 2026; MiniMax changed only the application URL on 11 November. This is a reading of the published text, not legal advice.

  • Same four excluded territories. The grant is “Solely within the Applicable Territory”, which excludes “the European Union, the United Kingdom, the Republic of Korea and the United States of America”. Section V.4 covers outputs too: “You may not use, reproduce, modify, distribute, or display the MiniMax-H3-Image Works or any of their Outputs or results outside the Applicable Territory.” MiniMax invites people there to contact it.
  • A much lower revenue threshold. Section IV.1 requires prior written authorisation “if the Annualized Revenue of you and/or your affiliates exceeds 1 Million USD, or, you provide Hosted Service(s) to any Third Parties.” The H3 video licence asks for it only if “your commercial products and services generate more than 20 million US dollars”. The image test counts your whole company and its affiliates, not just the product, and the licence does not define Annualized Revenue.
  • Any hosted service. A Hosted Service is any service through which the model’s “functionality is made available to Third Parties”, including APIs and web access, at any revenue. Read literally, that includes a ComfyUI instance other people can reach.
  • Branding. Show “MiniMax-H3-Image” on the interface of a commercial product that uses it.
  • Outputs. “MiniMax claims no rights over the Outputs you generate”, but you may not use them to improve another AI model, and policy item 12 forbids publishing generated content “without clearly and prominently disclosing that such information and/or content is machine-generated”. Item 19 bans military use.
  • The prompt expander. MiniMax’s 17.9 GB PE model is derived from Apache 2.0 Qwen3.8-27B but “is part of MiniMax-H3-Image and is accordingly subject to this Agreement”. The Qwen3.5 9B file the templates use comes from Comfy-Org/Qwen3.5, tagged Apache 2.0.

The encoder file the templates download sits in Comfy-Org/MiniMax-H3, which carries the H3 video licence, while MiniMax’s own H3 Image repository holds an identical copy under the image licence. Both exclude the same territories; the image terms are the stricter on revenue and hosting. Comfy-Org’s own card flags the difference: the image licence “is different from the original MiniMax H3 license”. Neither repository is gated, so a download succeeding says nothing about whether you are licensed. Our licence checker covers other models.

Is it maintained

On 15 November, MiniMax-AI/MiniMax-H3-Image on GitHub (created 9 November) was not archived, had 5 commits, no releases and a last push on 11 November. ComfyUI added support in v0.45.0 (PR #17655). Check ComfyUI’s releases page before reporting a bug.

FAQ

Does MiniMax H3 Image use flux2-vae.safetensors?

No. It uses minimax_h3_image_vae_fp16.safetensors (1.17 GB) in models/vae/, a 24-channel VAE. The FLUX.2 VAE has 32 latent channels.

What is the smallest MiniMax H3 Image download for ComfyUI?

29.41 GB from Comfy-Org: the pruned_w4a8 transformer, the nvfp4_awq encoder and the VAE. The stock templates load 56.63 GB.

Can I use MiniMax H3 Image commercially?

Outside the US, EU, UK and South Korea, yes, but you need MiniMax’s written authorisation if you and your affiliates exceed US$1M in annualised revenue or you offer it as a hosted service.

Related guides and tools

  • MiniMax H3 in ComfyUI: video files, VRAM and the licence
  • FLUX.3 Image Klein in ComfyUI
  • Qwen-Image 2.1 in ComfyUI
  • Best local image generation models in 2026
  • ComfyUI install in 2026: Desktop, Portable or Manual
  • ComfyUI templates: which run locally and which need the cloud
  • Image model directory
  • ComfyUI models and VRAM
  • Model licence checker

Sources

  • MiniMaxAI/MiniMax-H3-Image (model card, configs, file listing) and its LICENSE
  • Comfy-Org/MiniMax-H3-Image, Comfy-Org/MiniMax-H3 and Comfy-Org/Qwen3.5 (cards and file sizes)
  • ComfyUI tutorial: MiniMax-H3-Image
  • Comfy-Org/workflow_templates (index.json, image_minimax_h3_image_t2i.json, image_minimax_h3_image_edit.json) and the ComfyUI source (comfy_extras/nodes_minimax_h3.py)
  • MiniMax-AI/MiniMax-H3-Image on GitHub (README)
  • diffusers: MiniMax-H3-Image and SGLang cookbook
  • black-forest-labs/FLUX.2-klein-4B (VAE config, for the channel comparison)
  • MiniMax H3 licence (for the comparison)

Updated 15 November 2026. File sizes, file hashes, template contents, node source, licence texts, GitHub issues and repository activity were read on that date.

⚡ Free Interactive Tools

Work out what your hardware can actually run

Free calculators and directories, sized from real model weights — before you download gigabytes or order a card:

🎨 ComfyUI VRAM Peak memory per image model ⚡ VRAM & KV Calculator Weights, cache and overhead 🚀 Start: Local LLMs Hardware, runner and first model 📊 What Can I Run? Models by VRAM tier, 8GB–128GB

Filed Under: Technology, Tutorial

About BFWR Editorial

BFWR Editorial is the byline for guides researched and written by Best Free Web Resources. These guides are built from primary sources — model cards, configuration files, official documentation and vendor pricing pages — and every figure is dated and linked so you can check it yourself. Size and memory numbers are arithmetic computed from published model data or from our own calculators, never measurements: BFWR does not operate a hardware test lab and publishes no benchmarks of its own. Where a figure comes from a vendor, we say so.

Reader Interactions

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Primary Sidebar

⚡ FREE INTERACTIVE TOOLS

Local AI Hardware & VRAM Matrix

Benchmark your PC or Mac for local AI, calculate VRAM requirements, and compare cloud API costs vs local hardware ROI.

🚀 New here? Start with Local LLMs → 💻 Hardware Matrix (79 Devices) → 📚 Local LLM Directory (30 Models) → 🎮 AI GPU Comparison → ⚡ VRAM & KV Cache Calculator → 💰 API Pricing & Break-Even → 🍎 Mac VRAM Unlocker (90% Boost) →
📦 Ollama 🖥️ LM Studio ⚙️ llama.cpp 🚦 vLLM 💬 Open WebUI 🎨 ComfyUI 🔒 Self-Hosted 📊 VRAM Tiers 🎯 Recommender ⚖️ Comparisons

Search

  • Facebook
  • Instagram
  • RSS
  • Twitter
  • YouTube

Footer

Latest Guides

  • MiniMax H3 Image in ComfyUI: Files, VRAM, Settings and the Licence
  • LFM3 1.7B Locally: 256K Context in About 5GB, and the Licence Limit
  • Fooocus in 2026: Is It Still Worth Installing, and What to Use Instead
  • Gemma 4 31B Locally: VRAM, Context Length and Which GPU It Fits
  • 5 Ways You Can Make Business Operations Simple Without Upgrading Your Tech
  • LTX-2.5 in ComfyUI: Files, Settings and the $10M Licence Rule

AI Tools & Calculators

  • What can I run?
  • VRAM Calculator
  • GGUF Size Calculator
  • Fine-tuning VRAM
  • LoRA Training VRAM
  • Image Generation VRAM
  • Mac VRAM Optimizer

Browse & Compare

  • Models Directory (30 models)
  • Hardware Matrix (79 devices)
  • Image Models
  • Model Licences
  • Comparisons
  • Model Recommender
  • API Pricing & ROI

Guides & Resources

  • Technology
  • Coding
  • Tutorials
  • Freebies
  • Graphics

Copyright © 2026 · Best Free Web Resources · All Rights Reserved.  |  About · Privacy Policy · Contact · AI Tools Hub