Creator Digest
AAI Search· 774K subscribers

New top local AI image editor is here!

20 min videoNov 12, 2026Worth watching in full

MiniMax H3 Image is a fast local image generator and editor; run it in ComfyUI, using quantized GGUF models if VRAM is limited.

  • For text-to-image, use a resolution of at least 2 megapixels; the presenter recommends 4 megapixels or more.
  • In ComfyUI, download the diffusion model, text encoder, and VAE, then select them in the provided workflow.
  • For lower VRAM, load a quantized GGUF model with the ComfyUI-GGUF extension, accepting some quality loss.

The ComfyUI setup and workflow walkthroughs add practical value beyond the comparisons. · Best for: Local-AI users who want to generate or edit images in ComfyUI.

The video reviews MiniMax H3 Image, compares its image generation and editing with other models, and demonstrates local ComfyUI workflows, including a quantized-model setup for lower VRAM.

Capabilities and model comparisons0:00

  • The presenter reports fast generation and editing, with support for up to 16 reference images and a stated minimum of 6 GB VRAM.
  • Compared with Qwen Image 2.1, both models handle character sheets and blurry-image upscaling well; Qwen is more faithful in photo restoration, while MiniMax adds detail.
  • Against Korea 2, MiniMax does better in some pose and design examples, while Korea 2 looks more realistic in some prompts. Neither model consistently renders text correctly.

Install ComfyUI models5:29

  • Update ComfyUI, then find the MiniMax Image text-to-image and editing workflows under Templates.
  • Download the diffusion model, text encoder, and VAE, placing them in ComfyUI's diffusion_models, text_encoders, and vae folders.
  • The presenter uses a 12.5 GB W4A8 diffusion model and omits the optional prompt enhancer to save storage and compute.

Text-to-image workflow10:04

  • Refresh the model list and select the diffusion model, text encoder, and VAE in the workflow.
  • Enter a prompt, choose the aspect ratio and resolution, and leave the sampler and scheduler at their defaults to start.
  • The presenter advises using at least 2 megapixels, ideally 4 or more; their example took about 11 seconds for a 2K image on a 16 GB VRAM GPU.

Image editing and reference images12:51

  • The editing workflow resizes the input image to 4 megapixels before image conditioning; it can use the resized image dimensions for the output.
  • For multiple references, add Load Image nodes, connect them to image slots, and enter the output width and height manually if bypassing the original resize nodes.
  • Describe how each reference should be used in the prompt. The demonstrated two-reference edit took about 40 seconds.

Run quantized models with less VRAM16:40

  • Choose a compressed GGUF model that fits your hardware; the presenter notes that more compression can reduce quality.
  • Install or update ComfyUI-GGUF by City96, then replace the standard diffusion-model loader with a GGUF loader in the workflow.
  • Select the downloaded GGUF model in the workflow. The same loader replacement applies to the image-edit workflow.

Overall assessment19:35

  • The presenter considers Korea 2 or Moss Image better in some text-to-image respects, but highlights MiniMax's speed at higher resolutions.
  • For editing, the presenter describes MiniMax as close to Qwen Image 2.1 and among the strongest open image editors.

Step by step

  1. Update ComfyUI to the latest version and open the MiniMax Image workflow from Templates.
  2. Download the diffusion model, a compatible text encoder, and the VAE; place each in its corresponding ComfyUI model folder.
  3. Refresh the model list, select the downloaded models, and disable the optional prompt enhancer if you do not want its extra model and compute requirements.
  4. Enter a prompt and choose an aspect ratio and resolution of at least 2 megapixels; run the workflow.
  5. For editing, load an image and describe the change. For multiple references, add Load Image nodes, connect them to image slots, and set output dimensions if needed.
  6. For a lower-VRAM setup, install ComfyUI-GGUF, replace the standard diffusion loader with a GGUF loader, and select a quantized model.

Tools & resources mentioned

  • ComfyUI — Local interface and workflows for running MiniMax H3 Image.
  • MiniMax Image text-to-image and image-edit workflows — Available through ComfyUI Templates; the presenter also mentions downloadable workflow files.
  • MiniMax H3 Image model files — Diffusion model, text encoder, and VAE used by the workflows.
  • ComfyUI-GGUF by City96 — Extension needed to load quantized GGUF diffusion models.
  • Qwen Image 2.1 and Korea 2 — Models used in the image-editing and text-to-image comparisons.
Watch on YouTube

Notes are AI-generated from the video's transcript and may miss nuance.