P-H3-Image ComfyUI Workflow Guide

P-H3-Image runs natively in ComfyUI. Pruna AI publishes a separate repo, PrunaAI/P-H3-Image-Comfy, with int8 and 4-bit files that bring the model down to about 21 GB of VRAM, and Comfy ships two ready-made workflow templates for it. This guide walks through what you need, where each file goes and which settings matter. It follows the official Comfy tutorial for P-H3-Image; check that page for anything that changes after a new ComfyUI release.

If you only want to see what the model does before downloading 35 GB of files, you can try P-H3-Image online on this site with the welcome credits you get when you sign in.

Requirements

• ComfyUI 0.44.0 or later. Both templates use the PH3ImageConditioning node, which older versions don't have. If you run an older build, update first.
• Disk space. About 35 GB of model files in total.
• VRAM. About 21 GB with the int8 files.

For comparison, the full bf16 weights from the main repo need about 83 GB of GPU memory at 8 MP without the upsampler and about 101 GB with it, so the ComfyUI files are the practical route on a single workstation card.

Files and folders

Download these five files and place each one in the folder shown. The folder names are inside your ComfyUI install.

1. minimax_h3_fl2va_pruned_int8_convrot.safetensors goes in ComfyUI/models/diffusion_models/. This is the same diffusion model the MiniMax H3 workflows use, so if you already have it, you don't need a second copy.
2. p-h3-image_text_encoder_int8_convrot.safetensors goes in ComfyUI/models/text_encoders/.
3. p-h3-image_prompt_upsampler_w4a8.safetensors also goes in ComfyUI/models/text_encoders/.
4. p-h3-image_vae_int8_convrot.safetensors goes in ComfyUI/models/vae/.
5. p-h3-image_lora_bf16.safetensors goes in ComfyUI/models/loras/.

After copying the files, restart ComfyUI or refresh the model lists so the loader nodes can see them. If a loader shows a red "missing" box, the file name or folder doesn't match the list above.

The two P-H3-Image ComfyUI templates

Open the Template Library in ComfyUI and look for:

• P-H3-Image: Text to Image (file image_p_h3_t2i.json)
• P-H3-Image: Image Edit (file image_p_h3_image_edit.json)

If you don't see them, your ComfyUI is older than the release that added them. Update ComfyUI and open the library again. On Comfy Cloud, the templates appear with the next stable release.

Text-to-image settings

The text-to-image template is short, and most nodes are already set the way the model expects.

1. Write the request. Type what you want after the "Request:" label in the prompt node. One or two plain sentences are enough, because the prompt upsampler expands them into a full prompt.
2. Keep the aspect ratio line in sync. The "Aspect ratio:" line in the prompt should match the size set in the Resolution Selector. The selector is set to 8 MP with sizes in multiples of 32. If you change one, change the other.
3. Leave the sampler alone. The schedule is a fixed 8 steps. There is no negative prompt and no guidance value to tune.
4. Choose the seed mode. Set the seed to randomize to get a new image on each run, or fixed to reproduce the example that ships with the template.
5. Read the upsampled prompt. A Preview Any node shows the full prompt the upsampler wrote. If the image misses something, check whether the upsampler dropped or changed it.

The 8 MP sizes from the model card, if you set the selector by hand:

• 1:1 — 2912 × 2912
• 4:3 — 3360 × 2496, and 3:4 — 2496 × 3360
• 3:2 — 3552 × 2368, and 2:3 — 2368 × 3552
• 16:9 — 3872 × 2176, and 9:16 — 2176 × 3872

Image edit settings

The edit template works with the image you want to change plus optional reference images.

1. Load the image to edit first. The first Load Image node is always the picture being edited.
2. Add references if you need them. The template takes up to nine images in total. Each extra image is named <Picture 2>, <Picture 3> and so on, and you refer to them by those names in the request.
3. Write the edit as a finished result. For example: "Edit Picture 1. The jacket is now the green coat from Picture 2. Keep her face, pose and the background exactly as they are." The prompt guide explains this form in detail.
4. Don't set a size. Width and height on the conditioning node stay at 0. The output keeps the shape of the first image and is scaled up to at least 4 MP.

If you have less VRAM

The int8 files already target about 21 GB. If you still run out of memory:

• Run text-to-image and edits at a lower size for drafts, then render the final at 8 MP. Expect more broken hands and faces at 2 and 4 MP; the model is trained for 8 MP.
• Check that you downloaded the int8 files listed above and not the bf16 weights from the main repo, which need about 83 GB of GPU memory at 8 MP.
• PrunaAI/P-H3-Image-Comfy also has 4-bit files. Compare a few outputs against the int8 version before you rely on them.
• Close other programs that hold GPU memory, such as a second ComfyUI instance or a game.
• If your card is far below this, the model is also available online, without a local GPU: open the generator.

Limits to expect

• Prompting matters. Hype words such as "stunning" or "8k" and very long prompts make results worse. Very short prompts are weak if you turn the upsampler off.
• Hands and crowds. Hands can break, and small people in the background can be distorted.
• Small text. Short quoted strings render well; body copy and long code turn into pseudo-text.
• Images only. The workflows produce still images, no video or audio.

License note

The model is released under the MiniMax H3 Community License. It excludes the European Union, the United Kingdom, the Republic of Korea and the United States unless you have a separate license from MiniMax, and products built on it must show "Powered by MiniMax H3". Comfy's page says commercial use requires a license obtained through Comfy. This is a summary, not legal advice; read the license on the Hugging Face model card before you use the model or its images in a product.

Try it without installing

Setting up ComfyUI takes time and a large GPU. To test prompts first, try P-H3-Image online: sign in with Google, use the welcome credits and pick a size from 1K to 4K.