MiniMax H3 Image Studio V2.2 — workflow and setup guide

By Latent Jack

H3 Image Studio brings together a positive prompt, image size controls, optional reference images, experimental detailers, and a before-and-after detailer comparison. V2.2 supports text-to-image generation, editing with reference images, and optional depth guidance made from a photograph. The starting settings are six-step LCM at up to approximately 20 MP, with the person, face and eye detailers off.

If this is your first installation, start with section 8 below for setup instructions and model downloads. Once the dependencies and models are in place, the first six sections follow the workflow from prompt to saved PNG.

Start with the person, face and eye detailers off. Review the generated image first; section 5 explains how to test detailing.

Whole workflow
Current V2.2 UI capture · 11 November 2026

The V2.2 workflow. The close-up screenshots below show the controls for each step.

1. Enter your positive prompt

Write your scene description or edit instruction in the green Positive prompt node. For text-to-image, leave all reference images and the depth guide disabled. Choose the image size and aspect ratio, keep the three detailers off, and generate an image. This is the base image: the generated result before any detailers.

Positive prompt
Current V2.2 UI capture · 11 November 2026

The green Positive prompt node contains the neutral example prompt. Use this field for a new scene description or an instruction to edit the reference images.

The V2.2 generation settings are:

• Sampler: SamplerLCM

• s_noise: 1.1

• s_noise_end: 0.9

• noise_clip_std: 0

• Scheduler: simple

• Steps: 6

• CFG: 1

• Denoise: 1

Use ComfyUI's separate SamplerLCM node with the noise settings above. Selecting an lcm sampler in a different node may not apply the same noise controls.

The 20 MP images and editing examples below use the BF16 image model, BF16 text/vision encoder and image VAE. They were generated at that resolution with the detailers off and no upscaling. Their visible detail comes from H3 generation or editing with references.

Keep a fixed image seed when comparing changes. This is my tested starting point, not a guarantee of the best result for every subject or edit. The community 20 MP workflow linked below prompted the LCM comparison; its original examples used INT8 model files, while my local comparisons and release examples use BF16. Credit and source: https://www.reddit.com/r/StableDiffusion/comments/1xt9ocv/minimax_h3_image_20mp_workflow/ and https://huggingface.co/buckets/n8te0/diffusion/tree/Minimax-H3-Image-20MP

2. Choose the image size

Use megapixels mode and multiple = 32. The V2.2 WAS resolution selector uses one million target pixels per MP unit, rather than the older selector's 1024 × 1024 convention.

• Portrait, 20 MP, 9:16: 3360 × 5952 pixels

• Landscape, 20 MP, 16:9: 5952 × 3360 pixels

• Actual area of either: 19,998,720 pixels

Dimensions round to multiples of 32, so the output is approximately the selected aspect ratio and megapixel count. I use 4–20 MP for this workflow. The selector offers larger values, but those are not tested H3 limits or guaranteed to run on your hardware. In megapixel mode, its edge-size field is ignored.

Higher image resolutions and additional reference images use more memory. These examples were made locally on an RTX 5090 with 32 GB VRAM; they do not establish memory requirements or performance on other hardware. Start at a lower resolution when checking a composition or investigating an out-of-memory error.

Fix Image seed for comparisons. Randomize it for a new variation. Detailer seed is a separate control, explained below.

Image size
Current V2.2 UI capture · 11 November 2026

Image orientation and megapixel target, with the resulting pixel dimensions displayed. Image seed and Detailer seed are explained separately below.

LCM sampler settings
Current V2.2 UI capture · 11 November 2026

The separate LCM noise controls and six-step simple schedule. The Person, Face and Eyes switches are off in the final-output screenshots below.

3. Add reference images

Open the reference-image subgraph. Its sixteen image inputs are arranged in two rows: fifteen accept reference images, and the final input accepts a photograph to convert into a depth map.

Select a reference image and turn on its Include picture switch. Turning the switch off keeps the file selected for later. Clear picture removes the selection and preview, and turns the input off without deleting the source file. Empty inputs are skipped.

Picture numbers in the prompt follow the enabled reference images, not the input labels. If only inputs 2 and 7 are enabled, H3 receives them as <Picture 1> and <Picture 2>. Recheck your positive prompt whenever you change the enabled images.

In the Positive prompt, explain what to use from each reference image. For a clothing edit, identify the picture of the person and the picture of the garment. For a two-person image, specify which picture provides each person and outfit, and where each person should stand. Avoid unnecessary references that contradict the intended face, clothing or setting.

Editing with references can change fine detail throughout the image. Asking H3 to preserve a face or background does not lock those pixels. Compare the result with the source references, checking faces, hands, texture, props and colour balance even after a small edit.

Reference images
Current V2.2 UI capture · 11 November 2026

Fifteen reference-image inputs and a separate photograph-to-depth input, arranged in two rows.

Before Clear picture
Current V2.2 UI capture · 11 November 2026

A reference image is selected. Include picture is already off in this screenshot.

After Clear picture
Current V2.2 UI capture · 11 November 2026

After Clear picture, the selection and preview are empty, and Include picture remains off. These screenshots show clearing the selection, not an on-to-off switch change. The source file is not deleted.

4. Make a depth guide from a photograph

The final image input accepts a normal photograph; it does not require a prepared depth map. Turn on Enable depth guide to convert that photograph with Depth Anything V2 Large at the saved preprocessing resolution of 1024. Nearer areas in the resulting map are light and more distant areas are dark.

H3 receives the converted relative depth map after the enabled reference images. This connection does not also send the source photograph as an appearance reference. Leaving the depth input empty or disabled skips both conversion and the added depth instruction. Clear picture also disables the depth guide.

The map gives H3 a visual reference for composition and perspective. This is not a dedicated H3 ControlNet connection and does not guarantee an exact pose or layout. Inspect the generated result. The screenshots below show the controls; this guide does not include a controlled V2.2 comparison with and without depth guidance.

Photograph for the depth guide
Current V2.2 UI capture · 11 November 2026

Select a normal photograph in the depth input and turn on Enable depth guide. This screenshot shows an empty, disabled input, not a generated depth map.

Depth map conversion
Current V2.2 UI capture · 11 November 2026

Depth Anything V2 Large converts the selected photograph at preprocessing resolution 1024. This screenshot shows the settings, not a new depth-guided result.

5. Test the optional detailers

Person, Face and Eyes are OFF by default. Testing has not yet shown that these detailers reliably improve the image. They can change a face, expression or texture that looked better in the generated image before detailing, so these settings still need tuning.

Detailing runs in this order: Person → Face → Eyes. Each enabled detailer receives the image from the previous stage. A disabled detailer, or one with no accepted region to process, passes the image through unchanged. SAM 3 finds the regions, the helper filters out small fragments, and Impact Pack processes the accepted crops.

The saved V2.2 starting settings are:

• Person: 6 steps, simple, CFG 1, denoise 0.10; guide 1024 / maximum crop 1536

• Face: 6 steps, simple, CFG 1, denoise 0.10; guide 1024 / maximum crop 1536

• Eyes: 6 steps, simple, CFG 1, denoise 0.20; guide 768 / maximum crop 1024

A separate SamplerLCM inside the detailer subgraph supplies noise strength 1.1 → 0.9 and clipping 0 through Impact Pack's custom-sampler hook. These low-denoise crop passes use a different sigma range from full-image generation. Good results from the main sampler do not prove that the detailer settings will improve a crop.

Detailer seed is independent of Image seed. Each pass receives a distinct integer offset. A collision guard keeps the image seed out of the reserved first 4096 region/cycle offsets; the saved region caps are 3/6/12 with one cycle. Fix the detailer seed for a repeatable experiment, or randomize it without changing the base image seed.

Use the Enable person detailer, Enable face detailer and Enable eye detailer switches to test one at a time. Check the result in Image Comparer before keeping it.

Person, Face and Eyes detailers
Current V2.2 UI capture · 11 November 2026

The Person, Face and Eyes detailers. Each enabled stage receives the image from the previous stage; the settings are enlarged below.

Detailer crop settings
Current V2.2 UI capture · 11 November 2026

The saved crop sizes and denoise values. These settings do not enable the detailers: the main Person, Face and Eyes switches are off by default.

Separate image and detailer seeds
Current V2.2 UI capture · 11 November 2026

Two labelled screenshots: Image seed and the independent Detailer seed. They are separate control views, not one continuous workflow view.

LCM sampler and detailer seeds
Current V2.2 UI capture · 11 November 2026

The LCM sampler and separate detailer seeds. Some badges come from a wrapper installed on the test machine; that wrapper is not an extra dependency. Use the custom-node list in section 8.

6. Compare and save the correct result

Image Comparer shows the generated image before detailing on A, and the final image after any enabled Person → Face → Eyes detailers on B. Save final PNG saves the same image shown on B. With all three detailers off, A and B match and the saved PNG is the generated image without detailing.

The saver is connected directly to the final image path. Its relative prefix is images/%year%-%month%-%day%/H3_Image_Studio_V2_2, so results go under your ComfyUI output/images/YYYY-MM-DD folder. Keep the PNG and its embedded workflow/prompt when you want to revisit a result.

Before and after detailers
Current V2.2 UI capture · 11 November 2026

Image Comparer A shows the generated image before detailing. B shows the final image after any enabled detailers.

Save final PNG
Current V2.2 UI capture · 11 November 2026

Save final PNG saves the same image as comparison B: the final result after any enabled detailers. The output path uses a dated folder inside images.

Reference images are inputs to H3; they are not the A side of this comparer. The comparer checks before and after detailing. Its connections and the final save connection have been checked, but you still need to judge whether a detailer makes the image look better.

7. Three practical editing examples

These editing examples use the six-step LCM settings at approximately 20 MP, with no upscaling and all detailers off.

Change hair colour. One source portrait supplies the woman, pose and setting. The prompt changes blonde hair to chestnut brown. The identity and large hairstyle remain recognisable, but the edited skin, iris and lip detail is visibly softer, with some tonal changes. This is a useful targeted edit, not proof that everything outside the hair remained identical.

Transfer a jacket. Picture 1 supplies the black-bobbed woman, pose and scene; Picture 2 supplies the burgundy-and-cream jacket. The result follows the jacket's panels, stripes and cuffs, and closely preserves the person and scene from Picture 1. Fine facial, fabric and camera details are regenerated, so this is not exact product reproduction.

Combine two people. Picture 1 supplies the blonde woman and night-time car scene; Picture 2 supplies the black-bobbed woman. The landscape result keeps their identities and outfits distinct. The wider car/street scene is regenerated, and a small blurred pedestrian appears in the background despite the request for exactly two women.

Hair-colour edit
Reference edit · six-step LCM · approximately 20 MP

Source portrait, exact hair-colour edit prompt and generated result. Fine detail is softer in the edited image. Detailers are off.

Clothing edit
Reference edit · six-step LCM · approximately 20 MP

The first reference provides the person and pose; the second provides the jacket. The generated edit follows the jacket design while changing some fine detail. Detailers are off.

Two people in one image
Reference edit · six-step LCM · approximately 20 MP

Two source reference images and the generated two-person scene. The wider background and car are regenerated. Detailers are off.

Each editing comparison shows the source reference images, exact edit prompt and generated result. These are separate from the workflow's before-and-after detailer comparison. The accompanying seven-second reel presents the still images with wipes and zooms; it contains no generated character motion or audio.

8. Install the workflow and models

The workflow uses ComfyUI's MiniMax H3 image nodes, subgraphs, SamplerLCM and BasicScheduler. Install these additional custom-node packs:

• WAS Suite — WASResolutionSelector
  https://github.com/WASasquatch/was-node-suite-comfyui

• rgthree — seed controls and Image Comparer
  https://github.com/rgthree/rgthree-comfy

• Custom Scripts — integer math expressions for detailer seeds
  https://github.com/pythongosssss/ComfyUI-Custom-Scripts

• Impact Pack — detailers and CustomSamplerDetailerHookProvider
  https://github.com/ltdrdata/ComfyUI-Impact-Pack

• TBG SAM3 — segmentation nodes
  https://github.com/Ltamann/ComfyUI-TBG-SAM3

• ControlNet Aux — Depth Anything V2 preprocessing
  https://github.com/Fannovel16/comfyui_controlnet_aux

The ZIP contains one workflow JSON and two helper-node folders: ComfyUI-H3-Image-Picture-Slots and ComfyUI-EF-Detailer-Guard. The helpers provide clearable reference-image inputs and filtering for detailer regions. Only their required Python/JavaScript files and MIT LICENSE files are included. Use the instructions and links in this guide to install model weights and third-party node packs separately.

For installation:

1. Use ComfyUI with native H3 Image, SamplerLCM and subgraph support.

2. Install the six external node packs above, following each project's instructions.

3. Copy both supplied helper folders directly into ComfyUI/custom_nodes, using the paths below.

4. Place the three base model files in the folders listed below.

5. Restart ComfyUI after installing nodes, refresh the browser and open the supplied JSON.

Both helper folders belong directly inside the existing custom_nodes directory:

• ComfyUI/custom_nodes/ComfyUI-H3-Image-Picture-Slots/

• ComfyUI/custom_nodes/ComfyUI-EF-Detailer-Guard/

For the Windows portable layout, the same locations are:

• ComfyUI_windows_portable/ComfyUI/custom_nodes/ComfyUI-H3-Image-Picture-Slots/

• ComfyUI_windows_portable/ComfyUI/custom_nodes/ComfyUI-EF-Detailer-Guard/

Each named helper folder should contain its own __init__.py directly inside it. Copy the two helper folders themselves, rather than adding another custom_nodes folder inside custom_nodes. Once they are in place, restart ComfyUI and refresh the browser so the Python nodes and picture-clearing interface are loaded.

This workflow was prepared with ComfyUI 0.45.0 and frontend 1.58.5. These are the tested local versions, not confirmed minimum requirements. Install Python dependencies into the environment that runs ComfyUI. If a model loader cannot find a file, check its name and folder before changing the workflow.

Model file locations, relative to your ComfyUI folder:

• models/diffusion_models/MiniMax-H3-Image/minimax_h3_image_pruned_bf16.safetensors
  https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/diffusion_models/minimax_h3_image_pruned_bf16.safetensors

• models/text_encoders/MiniMax-H3/qwen3vl_32b_minimax_h3_bf16.safetensors
  https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_bf16.safetensors

• models/vae/MiniMax-H3-Image/minimax_h3_image_vae_fp16.safetensors
  https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/vae/minimax_h3_image_vae_fp16.safetensors

• models/sam3/sam3.pt — used by enabled detailers
  https://huggingface.co/facebook/sam3

• custom_nodes/comfyui_controlnet_aux/ckpts/depth-anything/Depth-Anything-V2-Large/depth_anything_v2_vitl.pth
  https://huggingface.co/depth-anything/Depth-Anything-V2-Large

The diffusion model, text/vision encoder and image VAE are required to generate images. SAM 3 weights are needed when using the detailers; Depth Anything weights are needed when using the depth guide. Install the node packs even if you leave those features off, so ComfyUI can load the complete workflow. Check the model publishers' access requirements and terms before downloading.

The clean public copy starts with a neutral tree-frog prompt, all picture slots empty and disabled, all detailers off, and a relative dated output prefix. Static graph, privacy and dependency checks have passed, and the current UI captures show this copy. A clean-machine installation has not been tested.

9. Personal testing notes — why I am interested in this workflow

ComfyUI announced native MiniMax H3 Image support on 10 November 2026, describing six-step generation and 4096 × 4096 output. My 20 MP figures below come from the measured local runs on 11 November, rather than a claim that the announcement specifies that larger size. Official announcement: https://blog.comfy.org/p/minimax-h3-image-is-now-available

Generating around 20 million pixels without an upscaling pass has been one of the most enjoyable parts of this testing. I can inspect faces, hair and materials in the full-resolution PNG. Photorealistic portraits and glamour photography have become more promising as the compositions and prompts have improved; a detailed, lifelike frog from the earlier showcase remains a particular favourite. Those are my observations from the images I have made, not a claim about every subject the model can handle.

The measured native 20 MP text-to-image runs on my RTX 5090 took 108.448–119.894 seconds across ten images, with a median of 111.588 seconds: roughly two minutes. The three reference edits took 159.541–211.339 seconds, roughly 2 minutes 40 seconds to 3 minutes 31 seconds. These are ComfyUI execution-start to execution-success measurements, excluding queue waiting and including executed loading, encoding and saving stages. Caching and system load vary. Other machines and workflows will differ.

The hair change, jacket transfer and two-person image have been useful editing tests. I have not compared enough other editing models to rank them.

Earlier higher-step ER-SDE/beta57 tests could exaggerate skin lines, highlights and fabric texture. Reference sheets and portraits did not always favour the same settings. The six-step LCM setup gave me results I preferred, while the detailers still need tuning. More processing has not reliably meant a better image.

My main interest is making reference images for video. These examples test characters and clothing; props, spaceships and recurring locations are still ideas to explore. I am also curious what future community LoRAs or checkpoint variants might add. For now, these are early hands-on findings.

10. License and credit

The MiniMax H3 Image Studio V2.2 workflow file and the two included helper-node folders, ComfyUI-H3-Image-Picture-Slots and ComfyUI-EF-Detailer-Guard, are released under the MIT License. They carry the notice Copyright (c) 2026 Latent Jack. The MIT text is embedded in the workflow JSON and provided in a LICENSE file in each helper folder.

Free to use, adapt and share. If you share this workflow or build on it, a credit and link back to Latent Jack would be appreciated.

That public credit is voluntary; it is not an extra condition of the MIT License. When redistributing copies or substantial portions of the covered software, keep its copyright and MIT permission notices. Third-party node packs and model weights remain subject to their own licenses. MIT license text: https://opensource.org/license/mit