MiniMax H3 Image in Two Steps: What the LoRA Buys, and What It Costs
A community LoRA cuts MiniMax H3 Image from 6 denoising steps to 2. At 16 MP that is 1.4x faster, not 3x, and the finishing pass can cost more than it saves.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

A Japanese daily digest put the new pDMD adapter at number eight in its roundup, under a headline about cutting production costs. The claim attached to it was the usual kind: high-resolution output supported, interactive previews refreshing in about a second, image work getting faster. All of that is roughly true, and all of it is the least interesting part.
Here is the number that should have led. At 4096 by 4096 on an H200, stock MiniMax H3 Image takes about 25.1 seconds per picture on its native schedule. With the adapter merged in, it takes about 17.4. That is 1.4 times faster, not six steps divided by two.
The adapter is a drafting win rather than a render-time one, and the difference matters when you decide whether to install 40 GB of weights.
What the adapter actually is
MiniMax H3 Image is a 33.1B image transformer that renders at a native 4096 by 4096, which is 16 megapixels. The stock checkpoint is step-distilled for a fixed six-step schedule, and the practical consequence is that you cannot run it at two steps at all. At two steps it is still mid-denoise: the composition is legible, because the model is far from noise, but the picture is soft, flat and visibly unfinished.
The pDMD LoRA, from the Projected Distribution Matching Distillation paper, is rank 128 applied over attention and feed-forward layers of all 50 blocks. Merged into the official transformer, it produces a finished image in two denoising evaluations on a plain flow-matching grid, with no classifier-free guidance, at the model's native resolution.
The card ships it four ways, and the choice matters more than it looks:
- A 2 GB ComfyUI-layout LoRA, to sit on top of the Comfy-Org repack you may already have.
- A single 40 GB checkpoint with the LoRA pre-merged in fp32, for people who do not want a second node in the graph.
- A diffusers-layout copy that merges into a
MiniMaxH3Transformer3DModel, either the official checkpoint or Pruna's P-H3-Image. - A pre-converted transformer folder, so the diffusers pipeline loads without an on-the-fly conversion step.
Two constraints come with it. The LoRA strength has to be exactly 1.0, because any other value scales the distillation delta that carries the whole effect. And quantized checkpoints, the rotated int8 and w4a8 builds, cannot take a pre-merge, so on those you run the LoRA path or nothing. If you skipped that lesson on another recent adapter, the MOSS-Image two-forward release had the same trap.
The graph itself has one detail worth knowing before you open ComfyUI. The H3 image latent is a one-frame video latent, so the empty latent node is the video one, not an image node. From there: sigma shift at 12, Euler, a simple scheduler, two steps, denoise 1.0, positive prompt only. The card was verified against ComfyUI master on 10 November with zero unmatched keys, which is more than most community merges can claim.
Where the speed lands
Two-step generation at 16 MP saves eight seconds on a datacenter card. Two-step generation at preview size changes what you do while thinking, and that is the honest case for it.
The demo build, a slim FP8 pipeline on a shared GPU slice, renders 2048 by 2048 in about 4.6 seconds with its detail pass included, against roughly 7 seconds for the official six-step demo at the same size. A two-step frame at around a megapixel lands in one to one and a half seconds. At that speed, typing a prompt and watching the layout answer back becomes a real loop instead of a queue with a progress bar.
The catch is the one every preview mode has. A one-megapixel preview is not a small version of your four-megapixel picture. The noise is tied to canvas size, the same reason a seed does not survive a change of resolution, so the preview is a composition sketch. Judge framing there. Do not judge a face, a hand or a texture there.
The detail pass is the actual decision
Once you have a fast draft, the question becomes how to finish it, and the card offers two answers with very different bills.
The cheap one is a same-model regenerate. Render the draft at half size with two steps, then regenerate at full size from (1 - s0) * latent(upscaled draft) + s0 * noise with three steps on the refiner grid, where s0 sits between 0.5 and 0.7. That costs one to two seconds at four megapixels. Composition and style hold, and for most prompts this is the whole finishing step.
The expensive one is an in-context refiner: a second copy of the transformer carrying an eight-step HyperFlow schedule, a sharpening schedule and the in-context LoRA, regenerating the draft at its native 16 MP while the draft sits in the transformer's context as reference rows. The documented recipe is three steps at s0 = 0.835, about 35 seconds per 16 MP image and 95 GB of peak memory. Measured end to end at 4096, that pipeline runs about 50 seconds per image, which is twice as slow as the stock six-step model it was supposed to replace.
Read that again, because it is the part the roundups leave out. If you always polish, the two-step merge makes your pipeline slower, not faster. The saving is in the middle of the process, where you throw away nine drafts, not at the end, where you keep one.
s0 is the dial that decides polish versus repaint
The sigma value is not a quality slider. It is a permission slip: how much of the draft the second pass is allowed to ignore.
On a fresh image, 0.5 to 0.7 is a polish. On an edit, the useful range is 0.4 to 0.55, where the source holds and the texture sharpens. Push an edit to 0.7 and the model redraws more boldly: in the card's own sweep of a glasses edit on the Mona Lisa, the face drifts and flat areas pick up speckle. That is the same behaviour we described for Pruna's image-only build, where the keep-list in the prompt is what protects the rest of the frame, and the sampler is a second chance to lose it.
The heavy refiner has its own failure at 0.835. On flat, clean areas, studio backdrops and smooth gradients, it can invent speckle-like micro-texture. If your product work lives on a seamless grey sweep, that is exactly the region a client zooms into, so either drop to the in-context LoRA alone at 0.5 to 0.7 or skip the finishing pass for those frames.
What I would do with it
- Install the 2 GB LoRA, not the merged checkpoint, unless you have no appetite for graph edits. The LoRA keeps the option of dropping back to six steps for a final delivery.
- Build two graphs, not one: a two-step draft at 1 to 2 MP, and a three-step finish at the size you actually deliver. Switching between them is cheaper than arguing about settings.
- Keep s0 in the graph as a numbered input with two presets, 0.6 for text-to-image and 0.5 for edits. Values above 0.7 belong in a separate graph that you open on purpose.
- Check three crops at 100 percent before you accept anything: a face, a hand, and a large flat area. The flat area is where distillation artifacts hide, because nothing else in the picture is competing for the model's attention there.
- Write long prompts. The model wants a full scene description rather than a keyword stack, and MiniMax publishes its own prompt-shape notes plus a local prompt expander if you would rather write short.
Two things to keep straight about provenance. This is a community merge of a research LoRA, not a MiniMax release, and every timing above was measured on one H200 or one shared GPU slice by the person who packaged it. Your card will be slower, and your prompts will not be their twelve. The LoRA itself is Apache-2.0. What you render with it is still governed by the MiniMax H3 community licence, with the territory and attribution terms that model carries, which we went through in the P-H3-Image guide.
The release earns its disk space on the loop it makes possible: you can throw away drafts without paying a queue, then finish the one that survived. That trade, judgment for compute, is what an adapter like this actually offers. Whether it pays depends on how many drafts you throw away, and only you know that number.