E-commerce AI image platform PixPix has added MiniMax's H3 Image model. Sellers can pick H3 Image in the PixPix workspace and generate 4096×4096 4K product images directly, without deploying a model or wiring up an API.
H3 Image is an open-source image model that MiniMax released on November 10. The company says it produces a 4K image in about 10 seconds, accepts up to 16 reference images, and handles both generation and editing in a single model. A Highspeed variant is also available, which MiniMax says runs 3 to 4 times faster than mainstream tools. PixPix is built around e-commerce work: product image sets, A+ detail pages, product retouching, model try-on and best-seller replication are all ready-made workflows, and H3 Image is now one of the underlying models those workflows can use.
How to prepare reference images
The most common failure in product imagery is that the result simply does not look like the item: the color drifts, the logo shifts, an accessory is added or missing. H3 Image's cap of 16 reference images leaves plenty of room, and you can organize your material like this:
- 4 to 6 shots of the product itself: front, side, back and top-down, plus one or two close-ups of details such as ports, fabric or hardware;
- 2 to 3 brand elements: a color swatch, packaging, or your store's existing hero images, used to lock in the overall tone;
- 3 to 5 scene and model references: the staging you want, model poses or outfit ideas;
- Use the remaining slots for style samples, such as compositions that sell well for competitors.
MiniMax's own showcase example uses 10 reference images. The more references you add, the more constraints the model has to balance, so first-time users may want to start with 5 to 8 and add more once they see the results.
Where 4K actually helps
Many marketplaces do not require 4K for main product images, but a 4K original is useful in a few places:
- Zoomed-in detail on product pages: when shoppers pinch to inspect fabric texture or the facets of a piece of jewelry, there are enough pixels to keep the image sharp;
- One image, several crops: a single 4096×4096 square can be cropped into a main image, a landscape banner and a vertical short-video cover, saving repeated generation;
- Ultra-wide banners: H3 Image supports 21:9 output at 6272×2688, so store homepage banners and campaign headers no longer need to be stitched together.
MiniMax gives the effective resolution range for H3 Image as 4 to 20 megapixels, with results best around 16 megapixels, so the 4K tier lands right in the model's comfort zone.
You can write shorter prompts
Open-sourced alongside H3 Image is H3-Image-PE, a 27B prompt-expansion model. It turns a one-line request such as "white sneakers on a wooden surface, natural light" into a full description covering lighting, materials and composition, and it also recommends an aspect ratio. The official API and some platforms call this kind of expansion in the background, so sellers who are not comfortable with prompting do not need to write long inputs. Whether PixPix turns expansion on by default depends on the option shown in its workspace.
On cost, MiniMax lists an API reference price of about $0.069 per 4K standard-tier image, with a 50% discount on requests at 4K and above during the first week. Whether PixPix bills by credits or per image, and how many H3 Image calls each plan includes, is not listed on its public pages; check the pricing page on PixPix after signing in. Sellers who want to turn finished product images into video ads can bring them to the canvas on PixTV and connect a video model.
Sources: MiniMax official blog, Hugging Face model card, CocoLoop, PixPix platform pages. The reference-image cap, resolution range and prompt-expansion model specs were checked against these; API pricing follows MiniMax's published figures, and the split of reference images is an editorial suggestion.