Run H3 Image
in ComfyUI

Choose P-H3 Image in ComfyUI before downloading files.

The tutorial’s starting point is Pruna’s native ComfyUI package. Its intended image baseline is about 8MP with an 8-step schedule. H3 Image Studio is a different third-party project, while MiniMax H3 Image is MiniMax’s own image checkpoint. Similar names do not make their loaders interchangeable.

Check the repository and workflow filename first. Then choose the component precision your machine can support. Our online samples were generated with MiniMax through fal; they illustrate that service’s output and are not local Pruna performance measurements.

H3 Image model files, folders and memory.

Build from the native eight-step workflow.

Open Pruna’s text-to-image workflow after updating ComfyUI. Keep its conditioning, latent, scheduler and decoder connected as supplied. Resolve missing components before modifying the prompt or increasing the canvas.

Pruna P-H3-Image — native ComfyUI checklist
1. Update ComfyUI to a revision with native P-H3 support.
2. Open p-h3-image_text_to_image.json from Pruna's workflows folder.
3. Choose one supported file per component: diffusion model, text encoder, prompt upsampler, VAE.
4. Resolve every missing loader filename before sampling.
5. Preserve the template's 8-step ManualSigmas grid, Basic Guider and Euler sampler.
6. Start with the supplied 8MP preset; do not copy MiniMax H3 Image's ~6-step configuration.
7. Save exact model revisions, prompt, expanded prompt, dimensions and seed.
8. Open p-h3-image_edit.json separately for editing.

This setup guide was checked against the linked developer documentation on November 11, 2026. We have not benchmarked these local workflows on our own hardware. Reported memory and speed figures retain their original test conditions.

Use the template’s sampling grid rather than changing an unrelated sampler to eight steps. A successful output should be saved with the exact input settings. When a test fails, change one variable at a time so the cause remains visible.

For editing, open Pruna’s editing workflow. Its input preprocessing and conditioning differ from a fresh-image graph. Keep a known working copy before adding references or optional nodes.

Single-Image VAE — Making H3 Output One Still Frame

A still-image graph needs a compatible latent and a decoder that returns the intended single image. Pruna already supplies its own image decoder. The two methods below are alternative routes for original H3 single-frame workflows; they are not additional requirements for the Pruna tutorial.

Method A: Mamad8’s image-specialized VAE

Mamad8/MiniMax-H3-Image-VAE provides a merged decoder for a single temporal latent, T=1, in H3’s 24-channel latent format. Place minimax_h3_t1_image_vae_step1597.safetensors in ComfyUI/models/vae/ and load it through the standard VAE loader and encode/decode nodes in a compatible original-H3 graph.

The author labels this route experimental and describes softness and loss of fine text or microtexture. Check the actual output before adopting it. Do not replace Pruna’s packaged VAE or MiniMax’s separate image VAE merely because all contain “H3” in the filename.

Method B: Fizgig H3 Still Latent + Still Decode

Fizgig H3 Still uses two custom nodes with the original H3 model and its existing VAE. Its latent node supplies one frame to the sampler; its decode node replaces the ordinary decoder. Keep conditioning and latent dimensions aligned, and follow the repository’s example graph.

Original H3 single-still route
Conditioning + Fizgig H3 Still Latent → sampler
Sampler output + original H3 VAE → Fizgig H3 Still Decode → Save Image
Use the author's example graph and its exact model files.
This is separate from Pruna's native eight-step pipeline.

This path does not require Mamad8’s VAE. Combining both changes without a validated graph makes it harder to identify decoding problems. Neither optional path has been benchmarked on this site’s local hardware.

Reference images: connect the input, then name its role.

Pruna’s current model card documents up to eight input images. In its editing workflow, the first image is the base to edit and later inputs supply reference details. Start with one unmistakable change, then add extra images only when they contribute a distinct object, background or style.

Follow the graph’s expected conditioning and numbering syntax. A reference displayed in a preview node may still be disconnected from the sampler. After removing an input, review image order and the written instruction together. The online H3 Image Editor uses MiniMax through fal and a separate limit of up to 16 references.

LoRA setup: match the adapter to the base model.

The Pruna workflow directory includes p-h3-image_lora_text_to_image.json and p-h3-image_lora_edit.json for original H3 weights. Start with the designated adapter and model-only loader. Keep this route distinct from the standalone Pruna checkpoint.

Do not assume an artistic LoRA, a Studio node option or a file for MiniMax’s newer image checkpoint is compatible. Record the base model revision, adapter filename and strength. Test the unmodified graph first, so any change in output can be traced to the adapter rather than an unresolved loader.

Common H3 Image ComfyUI errors.

Missing native nodes, mismatched components and a disconnected reference path need different checks. Choose the symptom below before installing another node pack.

Use H3 Image online without local model downloads.

If you want to inspect a complete image before maintaining a local environment, open the H3 Image Generator. Its prompt controls and original examples run in the browser. The online workspace targets MiniMax H3 Image, rather than silently substituting Pruna’s model.

Saved examples and prompts are free to explore. Live generation remains unconnected in this local preview. For a real input/output example, edit images with H3 Image and inspect the vase recoloring before-and-after.