P-H3-Image is Pruna’s open image model built on MiniMax H3, and its stock workflow runs 8 steps. A new 2-step LoRA cuts that to just two denoising steps, and the sample grids on the model card show sharp, finished images at 2912 x 2912 in about 2.9 seconds each on a workstation GPU.
The LoRA comes from the pDMD research team, and a ComfyUI-ready copy was published on Hugging Face by akhaliq.
This guide covers the file, the folder, the four node changes you need in ComfyUI, the right step count for each kind of image, the optional 2K refine pass, VRAM by file set, and fixes for washed-out or garbled results.
- What: a pDMD 2-NFE distillation LoRA for P-H3-Image, repacked for ComfyUI at akhaliq/p-h3-image-pdmd-2step. Apache-2.0.
- File:
minimax_h3_pdmd_2nfe_comfyui.safetensorsgoes inComfyUI/models/loras/. - Node changes: add Load LoRA (Bypass, Model Only) at 1.0, swap ManualSigmas for BasicScheduler (simple, 2 steps, denoise 1.0), keep euler and Basic Guider with no CFG, set flow shift to 12.
- Steps: 2 for speed, 3 as the safe middle, 4 for text and flat graphics.
- VRAM: the same as P-H3-Image itself: about 21 GB peak with the w4a8 set, 31 GB int8, 56 GB bf16 (at 8 MP).
- Licence catch: images still fall under the MiniMax H3 Community License, which needs “Powered by MiniMax H3” attribution and has regional limits.
This post is part of our MiniMax H3 ComfyUI guide, which lists every H3 workflow, file and fix.
What the 2-Step LoRA Is
Diffusion models make an image by removing noise over many steps. Each step is one full pass through the network, which is called an NFE (number of function evaluations). P-H3-Image’s stock workflow uses 8 steps.
pDMD (Projected Distribution Matching Distillation) is a training method that teaches a model to land on a good image in far fewer passes. The team behind it describe it in the paper PDMD: Projected Distribution Matching Distillation for Video Diffusion Models (arXiv 2609.35768). Their pdmd_2NFE_lora release is the 2-pass version.
The akhaliq repo holds two files:
lora_model_0.safetensors: the original bf16 LoRA, unchanged from the pDMD release.minimax_h3_pdmd_2nfe_comfyui.safetensors: the same weights with ComfyUI key names. This is the one you want.
The model card is blunt about why the LoRA matters: stock P-H3-Image at 2 steps gives washed-out, half-formed images. With the LoRA, the same 2 steps give full composition and detail.
If you have not met P-H3-Image yet, our P-H3-Image launch post explains what Pruna changed from MiniMax H3.
Files, Folders and VRAM
You need a working P-H3-Image setup first. The files come from PrunaAI/P-H3-Image-Comfy, and you pick one file per component. You can mix precisions.
| Component | Folder | bf16 | int8 (convrot) | w4a8 |
|---|---|---|---|---|
Diffusion model (e.g. p-h3-image_w4a8.safetensors) |
ComfyUI/models/diffusion_models/ |
40.5 GB | 21.2 GB | 13.4 GB |
Text encoder (e.g. p-h3-image_text_encoder_w4a8.safetensors) |
ComfyUI/models/text_encoders/ |
9.1 GB | 5.1 GB | 3.6 GB |
| Prompt upsampler (optional) | ComfyUI/models/text_encoders/ |
19.3 GB | 9.7 GB | 6.7 GB |
VAE (e.g. p-h3-image_vae_int8_convrot.safetensors) |
ComfyUI/models/vae/ |
5.6 GB | 3.1 GB | use int8 |
2-step LoRA minimax_h3_pdmd_2nfe_comfyui.safetensors |
ComfyUI/models/loras/ |
one file (bf16) | ||
Pruna lists these peak VRAM figures at 8 MP for each full set:
| Set | Download | Peak VRAM (8 MP) |
|---|---|---|
| bf16 | 55 GB | 56 GB |
| int8 | 30 GB | 31 GB |
| w4a8 with int8 VAE | 20 GB | 21 GB |
The optional prompt upsampler peaks at about 21 GB (bf16), 12 GB (int8) or 9 GB (w4a8) on its own. The LoRA adds very little on top.
So the realistic floor is a 24 GB card with the w4a8 set and no upsampler. On a 16 GB card, lower the resolution well below 8 MP and expect ComfyUI to offload to system RAM. Our VRAM guide explains how offloading trades speed for memory.
To grab the LoRA from the command line:
hf download akhaliq/p-h3-image-pdmd-2step minimax_h3_pdmd_2nfe_comfyui.safetensors --local-dir ComfyUI/models/lorasP-H3-Image support is on ComfyUI’s master branch. Until the next stable release, you need the latest ComfyUI from GitHub (nightly). The desktop app and Comfy Cloud get it with that release.
ComfyUI Setup in 4 Changes
Start from Pruna’s own p-h3-image_text_to_image.json workflow (in the workflows/ folder of P-H3-Image-Comfy). Run it once at the stock 8 steps to make sure your files load. Then make these changes:
- Add the LoRA. Put a Load LoRA (Bypass, Model Only) node between Load Diffusion Model and the sampler. Select
minimax_h3_pdmd_2nfe_comfyui.safetensorsand set strength to 1.0. - Swap the schedule. Delete the ManualSigmas node and add a BasicScheduler. Set scheduler simple, steps 2, denoise 1.0, and connect its SIGMAS output where ManualSigmas went.
- Set the shift to 12. The model card says flow shift 12. In Pruna’s workflow the shift lives in the ModelSamplingMiniMaxH3 node, which ships at 1.0.
- Leave the guidance alone. Keep the euler sampler, Basic Guider, no CFG and no negative prompt.
Save it as a new workflow so you keep the 8-step version for comparison. If you have not saved a variant before, our guide to saving ComfyUI workflows covers it.
2, 3 or 4 Steps?
The card gives clear advice on step count, and it matches how distilled models usually behave.
| Steps | Result | Use it for |
|---|---|---|
| 2 | Sharp and usable, but the layout drifts from the 8-step result and graphics can show fake text | Fast drafts, photos, ideas |
| 3 | The safer middle ground | Most everyday images |
| 4 | Closest to the 8-step baseline | Quoted text, logos, posters, flat graphics |
Keep every other setting the same when you change steps. Only the BasicScheduler steps value moves.
Optional 2K Refine Pass
For a final image you want to keep, the card describes a two-stage recipe:
- Render a 2 MP draft at 2 steps with the LoRA.
- Upscale it 2x.
- Re-generate at 8 MP with the Efficient-Large-Model Open-H3-Regenerate-2K refiner LoRA, 3 euler steps, starting at sigma 0.7.
The start point is a mix of the draft and fresh noise, (1 – 0.7) x draft + 0.7 x noise. In ComfyUI terms, this is the same as running the refine sampler at partial denoise on the upscaled latent.
The full run takes about 7 seconds per image on the card’s hardware, which it says is still about 1.7x faster than the plain 8-step baseline.
Two warnings from the card:
- Use sigma 0.5 for portraits. At 0.7, faces and people can drift.
- Use the refiner only in this second pass. It is a video-refinement LoRA, and merging it into the model for plain text-to-image collapses prompts.
Diffusers Route
If you prefer Python, the card loads P-H3-Image through a ModularPipeline, removes the prompt upsampler, and merges the LoRA with the merge_lora function from its run_p_h3_turbo.py script.
- Shift: 12.
- 2 denoising steps use 3 timestep points. 4 steps use 5.
- The sample grid uses 2912 x 2912 and seed 3.
The diffusers path is heavier than ComfyUI: about 83 GB in bf16 without the upsampler, or about 76 GB with ComponentsManager auto-offload. For a single consumer GPU, ComfyUI with the w4a8 set is the practical choice.
2-Step vs 8-Step vs Other Fast Image Models
| Setup | Steps | Strength | Weak spot |
|---|---|---|---|
| P-H3-Image stock | 8 | Best fidelity and text | Slowest of these |
| P-H3-Image + pDMD 2-step LoRA | 2 to 4 | About 2.9 s at 2912 px on a pro GPU | Layout drift and fake text at 2 steps |
| 2-step draft + 2K refine | 2 + 3 | Detail close to baseline, about 7 s | People can change at sigma 0.7 |
| Qwen-Image 2.1 Lightning v2 | 4 | Smaller model, easier on VRAM | Different look and licence |
| Pruna P-Q-Image-2.1 | 6 | Runs on smaller cards | Not as fast as 2 steps |
The speed figures above are from the model card, rendered with a slim FP8 plus fused-Triton build on an RTX PRO 6000 or H200. A stock ComfyUI build on a consumer card will be slower, so treat 2.9 s as the best case.
Licence
The pDMD 2-NFE LoRA is Apache-2.0, so the LoRA file itself is free to use and share.
The images are a different story. Anything made with P-H3-Image falls under the MiniMax H3 Community License.
The akhaliq card says it excludes use in the EU, UK, Korea and the US without a separate MiniMax licence, and it requires a “Powered by MiniMax H3” attribution.
Read the full licence on the MiniMax-H3 repo before you use these images in client work or products.
Troubleshooting
Images look washed out or half-formed
The LoRA is not being applied. That is exactly what stock P-H3-Image looks like at 2 steps. Check that the Load LoRA node sits between Load Diffusion Model and the sampler, that strength is 1.0, and that you picked the ComfyUI file (minimax_h3_pdmd_2nfe_comfyui.safetensors), not lora_model_0.safetensors.
“lora key not loaded” warnings in the console
ComfyUI prints this when a LoRA’s key names do not match the model. The usual cause here is loading lora_model_0.safetensors, which keeps the original pDMD key names. Switch to the ComfyUI copy, minimax_h3_pdmd_2nfe_comfyui.safetensors.
Fake text or squiggles on posters and logos
The card warns that 2 steps can invent pseudo-text on graphic prompts. Go to 4 steps for anything with quoted text or flat design.
The layout looks different from my 8-step image
That is expected drift at 2 steps. Use 3 or 4 steps, or treat the 2-step image as a draft and run the refine pass.
Out of memory
Use the w4a8 diffusion model and text encoder with the int8 VAE, skip the prompt upsampler, and lower the resolution. Check what else is using the GPU with our VRAM usage guide.
P-H3-Image nodes or ModelSamplingMiniMaxH3 are missing
Your ComfyUI is too old. Update to the latest master from GitHub, because P-H3-Image support is not in a stable release yet.
Faces change in the refine pass
Lower the refine start sigma from 0.7 to 0.5, as the card suggests for portraits.
FAQ
Does this LoRA work with plain MiniMax H3 for video?
This repo is set up and tested for P-H3-Image text-to-image. Use it with the P-H3-Image files, not as a video speed LoRA. For H3 video speed-ups, see the Turbo LoRAs in our best MiniMax H3 LoRAs list.
Can I use CFG or a negative prompt?
No. The setup is guidance-free: Basic Guider, no CFG, no negative prompt.
Is 2.9 seconds realistic on my GPU?
That number is from an RTX PRO 6000 or H200 with a tuned FP8 build. Expect longer times on consumer cards and in stock ComfyUI. The relative speed-up over 8 steps still holds.
Does it work with the P-H3-Image edit workflow?
The card’s instructions and samples cover text-to-image. Try it on the edit workflow at 3 or 4 steps, and compare against the stock 8-step edit before relying on it.
Which shift value should I use?
12, as the card states, for 2, 3 and 4 steps.
Is there an online demo?
Yes, akhaliq published a demo Space at akhaliq/p-h3-image-2step on Hugging Face.
Sources
- akhaliq/p-h3-image-pdmd-2step (model card)
- PrunaAI/P-H3-Image-Comfy (files, VRAM, workflows)
- pdmd2026/pdmd_2NFE_lora
- PDMD: Projected Distribution Matching Distillation for Video Diffusion Models, arXiv 2609.35768
