H3 Image in ComfyUI: Five Files, 21 GB of VRAM, and Eight Sampling Steps

The H3 image model is not a separate release from MiniMax. It is P-H3-Image, an open-weights model Pruna AI built from the MiniMax-H3 video model, and it runs in ComfyUI with five files and eight sampling steps. This guide covers the downloads, the real memory limit, and where 8 megapixels actually matters.

Most people who search for h3 image land on pages about a video model. That confusion is worth clearing up first, because the thing you can actually download today is an image model derived from it, and the two behave very differently in a workflow.

What is the H3 image model, exactly?

P-H3-Image takes MiniMax-H3, a video generation model, and retunes every component for still images. Pruna AI replaced the decoder, distilled the text encoder, shrank the transformer backbone, and cut sampling from 50 steps to 8. The result keeps H3’s photographic look while fitting on a single consumer GPU.

The release landed on November 3, 2026, with open weights on Hugging Face under the MiniMax H3 Community License. Pruna published a technical writeup covering each component they changed.

Property	Value
Base model	MiniMax-H3, a video model
Tasks	Text to image and image editing, one checkpoint
Sampling	8 steps, no CFG, no negative prompt
Resolution range	2 to 16 megapixels
Model size	20 GB smallest quantized set, down from 143 GB
Speed at 4 MP	2.6 s on one B200, down from 12.4 s
Editing input	Up to 9 images total, one source plus eight references
ComfyUI minimum	Version 0.44.0

The compression is not only file size. The text encoder went from 32B parameters to 4B through distillation, which drops it from 67 GB to 9 GB without changing output quality. The transformer backbone went from 33B to 20B by replacing the timestep block, which accounted for 40 percent of the original. Neither change required retraining from scratch.

Which files do you need to download?

The ComfyUI template pulls five files, about 35 GB total. One of them is the same diffusion model the MiniMax H3 video workflows already use, so if you have run those, you only need the other four.

File	Folder
minimax_h3_fl2va_pruned_int8_convrot.safetensors	models/diffusion_models/
p-h3-image_text_encoder_int8_convrot.safetensors	models/text_encoders/
p-h3-image_prompt_upsampler_w4a8.safetensors	models/text_encoders/
p-h3-image_vae_int8_convrot.safetensors	models/vae/
p-h3-image_lora_bf16.safetensors	models/loras/

You also need ComfyUI 0.44.0 or later, because both templates depend on the PH3ImageConditioning node added in that release. If the node shows as missing after you load a workflow, that is the cause, not a broken download.

Why three different file sizes get quoted

You will see 20 GB, 30 GB, 35 GB, and 55 GB cited for this h3 image checkpoint, and none of them are wrong. They measure different things, and mixing them up is the most common source of confusion when planning disk space.

The 20 GB figure counts the smallest quantized diffusion set alone, while 30 GB is int8 quantization of the full model. The 35 GB figure is every file the ComfyUI template downloads, and 55 GB is full BF16 weights, uncompressed.

Plan for 35 GB if you are following the ComfyUI route, because the template needs all five files present before it will run. Pruna’s own release notes quote the 20 GB number, which is why guides written from the announcement alone tend to understate the disk requirement by roughly 15 GB.

How do you install the h3 image files in eight steps?

The template is unforgiving about order: it checks for every file before it will run, so a missing encoder fails later than a missing diffusion model. Work through these steps once, in sequence, and most setup problems disappear before the first render.

1. Update ComfyUI to 0.44.0 or later. The PH3ImageConditioning node was added in that release, and both templates depend on it.
2. Start the h3 image download from the open-weight checkpoint on Hugging Face. Check the MiniMax H3 Community License before any commercial use, because the restriction sits on the base model rather than on the compression work.
3. Fetch the five files listed above: the pruned int8 diffusion model, two text encoders, the VAE, and the LoRA.
4. Place each file in its matching folder under models/. The text encoder and the prompt upsampler share models/text_encoders/.
5. Restart ComfyUI so the new nodes register. A node that still shows as missing at this point means the version check in step 1 was skipped, not that a file is corrupt.
6. Load P-H3-Image: Text to Image from the workflow library. The Pruna H3 image ComfyUI templates ship with the same eight-step sampler, so there is no sampling schedule to tune.
7. Write the request after Request:, keep Aspect ratio: in step with the Resolution Selector, and read the Preview Any node to see what the Prompt Upsampler actually sent to the sampler.
8. Render once at 8 MP, then drop to 4 MP for composition drafts. Sampling stays at eight steps throughout.

If the first output looks nothing like your request, go back to step 7 before touching the model files. The upsampler rewrites the prompt, and tuning your own wording against a rewritten prompt costs a full iteration for no reason.

How much VRAM does the H3 image model need?

The int8 files fit in about 21 GB of VRAM at 8 megapixels. That figure is the honest gate on this model, and it excludes a large share of gaming cards. Official documentation states the requirement and stops there, with no fallback path for smaller cards.

The 4-bit files are the way through. Pruna’s own benchmarks show a 4 MP image running in 6.3 seconds on an L40S with 48 GB using 4-bit weights, against 19 seconds for BF16 with offloading. The 4-bit path costs quality, mostly in fine detail, but it turns a 21 GB requirement into something a 16 GB card can attempt with offloading.

With 12 GB or less, expect to swap blocks between system memory and the GPU on every step, which pushes a single h3 image render past a minute. That is workable for a one-off test and painful for iterating on a prompt.

If you have neither the card nor the patience, running it in the cloud is the practical answer. ComfyUI’s hosted service carries the templates already, and renting a single B200 or L40S for an afternoon costs less than the electricity bill for tuning a local setup. The same logic applies if you only need a handful of images, which is the case for most people evaluating a new model.

How do you run the eight-step text-to-image workflow?

Load the template named P-H3-Image: Text to Image from the workflow library, or download the JSON from Comfy’s template repository. Write your request after the Request: label, and keep the Aspect ratio: line consistent with the Resolution Selector above it.

Sampling is fixed at 8 steps. Adding more does not add detail, because the distilled schedule is tuned for exactly that count. There is no CFG value and no negative prompt to balance, which removes two of the usual dials and makes this h3 image workflow noticeably easier to operate than a standard diffusion pipeline.

The resolution selector runs at 8 megapixels with a multiple of 32, which produces these canvases:

Ratio	Size at 8 MP
1:1 Square	2912 x 2912
4:3 Standard	3360 x 2496
3:2 Photo	3552 x 2368
16:9 Widescreen	3872 x 2176
9:16 Portrait	2176 x 3872

A Prompt Upsampler based on Qwen3.5-9B rewrites your one-line request into the detailed prompt the model renders best. The rewritten prompt is exposed through a Preview Any node, so you can read what actually reached the sampler. Check it before you start tuning your own wording, because a bad upsampler output will make you chase a problem that lives in the rewrite rather than in your input.

The upsampler is also the reason short prompts work here. A single sentence naming the artifact and the subject is enough for the h3 image model, provided you let the rewrite step expand it rather than writing the expansion yourself.

Should you render at 4 MP or 8 MP?

Eight megapixels is the quality target and four is the speed compromise. The difference is not uniform across subjects, so the choice depends on what you are generating rather than on a blanket rule.

Setting	Best for	Watch out for
8 MP	Portraits, skin texture, film stills, small background figures	Slower, needs the full 21 GB
4 MP	Drafts, composition tests, batch iteration	Faces, hands, and small details break more often
2 MP	Quick thumbnails only	Visible artifacts across the frame

Pruna states plainly that hands, small background figures, and fine details distort more below 8 megapixels. Text follows the same curve: short quoted strings render well at full resolution, while dense body copy or code turns into pseudo-text at any setting. If your image depends on legible small text, the h3 image model is the wrong tool rather than a tool you need to tune harder.

The practical workflow is to draft at 4 MP until the composition holds, then rerender the winner at 8. That keeps iteration fast without paying the full cost on images you were going to discard anyway.

How do you edit an image with reference pictures?

The edit template works from a source image plus up to eight references, nine images in total. Load the image you want changed into the Load Image node, then describe the result after Request:. The output inherits the shape of your source, which is why the width and height fields on the conditioning node stay at zero.

Additional reference images become Picture 2, Picture 3 and so on inside the prompt. You address them by that token, which is how you tell the model to place a specific person, object, or style from a specific image. The scene composition ability comes directly from H3’s training on building video from reference stills.

One habit matters more than any parameter: name what must not change. Without an explicit list of things to preserve, the model tends to redraw the entire scene, and you lose the identity or setting you were trying to keep. Write the preservation list as its own sentence rather than folding it into the change request.

What prompting habits improve the output?

Four rules from the official documentation cover most of the gap between mediocre and usable results. They are worth following literally, because this model is unusually sensitive to prompt construction.

1. Open by naming what the image is, such as “a photograph by an experienced street photographer” or “a two-color risograph print”. Naming the artifact sets the look more reliably than stacking style adjectives.
2. Describe one scene with a viewpoint, one main subject caught mid-moment, a few supporting elements, and light that has both a source and a direction.
3. Put every string of text in quotes, keep it short, state where it sits, and end with “no other text”. Unquoted text becomes pseudo-text.
4. Drop hype words like stunning, epic, cinematic, and 8k. They push the model toward a glossy, oversaturated look that reads as synthetic.

Very long prompts degrade output and very short prompts give weak results unless the upsampler rescues them. The sweet spot is a short request with concrete nouns, which the upsampler can then expand without inventing the scene for you.

What breaks first when you run the h3 image templates?

Most failures are configuration, not model quality. If you arrived here looking for a MiniMax H3 image generator, the useful mental model is a fixed pipeline with only two dials: prompt quality and resolution. Work down this list before assuming the checkpoint itself is broken.

• PH3ImageConditioning is missing after loading a workflow: ComfyUI is older than 0.44.0. Upgrade, then restart.
• The run stops or crawls at 8 MP: the int8 files need about 21 GB of VRAM. Use the 4-bit files with offloading, or draft at 4 MP.
• Hands, small background figures, or fine detail distort: render at 8 MP. Distortion below that is documented behaviour, not a prompt problem.
• Text in the image turns into pseudo-text: shorten the string, put it in quotes, state where it sits, and end with “no other text”. Dense body copy will not render at any setting.
• The output ignores your wording: read the Preview Any node. The Prompt Upsampler rewrote it, so fix the rewritten text rather than the original request.
• An edit redraws the whole scene: add an explicit sentence naming what must not change. Without it, identity and setting are not preserved.
• You plan to sell the output: the community license covers personal and research use, while commercial use needs a separate request through Comfy’s licensing page.

Each item maps to a setting already described above, which is why the fastest fix is almost always a checkbox in the workflow rather than a new download.

Why does searching for h3 image return video results?

Because the underlying model is a video model, and a large share of pages about MiniMax-H3 predate or ignore the image variant. The naming does not help: the official checkpoint is called P-H3-Image, the folder is p-h3-image-comfy, and the community nodes use MiniMax H3 in their titles.

When you are comparing options, the useful distinction is between running the image model locally and using a hosted generator. Local gives you unlimited iterations and full resolution control; hosted gives you a working result in a browser tab. Our comparison of commercial image models covers the second path, and the MOSS-Image ComfyUI setup covers a parallel local route with a different layout-control model. If you would rather skip installation entirely, Tarko AI is the lighter option, and deploying DeepSeek 4.1 Flash locally shows the same setup discipline applied to a text model.

What does the license allow you to do?

P-H3-Image ships under the MiniMax H3 Community License. That license excludes some regions unless you hold a separate license from MiniMax, and commercial use requires a request through Comfy’s licensing page. Read the terms before you build anything you intend to sell, because the restriction sits on the base model rather than on Pruna’s modifications.

Personal and research use is covered by the community license. The exclusion list is regional, so whether it affects you depends on where you are generating, not on what you are generating.

Frequently asked questions

Is the H3 image model free?

The weights are free to download and run under the MiniMax H3 Community License. Some regions are excluded from that license, and commercial use needs a separate license from MiniMax through Comfy’s licensing page.

Can I run it without 21 GB of VRAM?

Yes, using the 4-bit files with offloading. Pruna’s benchmarks show a 4 MP image rendering in 6.3 seconds on an L40S using 4-bit weights, and smaller cards can attempt the same with heavier offloading at the cost of speed and some fine detail.

Does it generate video?

No. P-H3-Image is image only. It handles text-to-image generation and still-image editing, and the temporal components of the original H3 model were removed during conversion.

Why do my hands look wrong?

Hands, small background figures, and fine details distort more at 2 MP and 4 MP than at 8 MP. Render at 8 megapixels when the subject includes hands or small people, and expect some residual artifacts regardless.

Can it render readable text in an image?

Short quoted strings render reliably at 8 MP. Dense text such as body copy, paragraphs, or code turns into pseudo-text at any resolution, so use a different model if legible long text is the requirement.

Do I need a negative prompt?

No. The 8-step distilled schedule has no classifier-free guidance and no negative prompt, so there is nothing to balance there. Prompt quality and resolution are the two levers that matter.

Related

• MOSS-Image in ComfyUI: Setup, Weights, and Layout Workflows
• GPT Images 2.6: What Changed and What It Costs
• Deploy DeepSeek 4.1 Flash Locally