AI video creation platform PixTV has added MiniMax H3 Image, the image model MiniMax recently open-sourced. On PixTV's infinite canvas, users can call it directly to generate 4K stills, then pass the images to a video model on the same canvas to keep going.
MiniMax released H3 Image and opened its weights on November 10. According to the company, it can produce a 4096×4096 image of roughly 16 megapixels in about 10 seconds, accepts up to 16 reference images, and handles text-to-image, image reference and image editing in a single model. PixTV had already listed the MiniMax H3 video model; this update adds the image half, so the H3 family's image and video models are now connected on one platform.
Lock the image first, then make it move, on one canvas
Most people making AI shorts already work in two stages: first fix the stills for characters, scenes and props and confirm the look is right, then feed those stills to a video model. The pain of doing the two steps separately is the constant exporting and importing, plus the style break that comes with switching models.
PixTV's approach is to put image, video and audio models on one infinite canvas, with nodes wired directly together. A 4K character reference made by H3 Image can be dragged straight into the H3 video model as a first frame or reference, or routed to other video models on the platform, such as Seedance and Kling, for comparison. The canvas also carries tools like AI storyboarding, white-model video and video replication, so the whole pipeline stays in the browser.
The two models also share a technical link. According to the Hugging Face model card, H3 Image reuses the H3 video model's Qwen3-VL text encoder as is, with no changes. The same prompt is therefore “read” into something closer on both sides, which in theory should reduce style drift when moving from image to video. Real results will depend on the shot, and no third-party systematic comparison exists yet.
What a 4K still does for video
Video models usually output below 4K; the H3 video model's API tops out at 2K. Generating stills in 4K may look like overkill, but it has three practical uses:
- Room to crop: from a single 4096×4096 image you can cut several 16:9 sections as first frames for different shots, with enough detail left in both characters and scenery;
- Ultrawide framing: H3 Image supports 21:9 output directly at 6272×2688, suited to openers, banners and cinematic establishing shots;
- Multi-reference consistency: it reads up to 16 reference images at once, so character turnarounds, costume details and scene-mood images can all be supplied together to keep reference stills consistent across shots.
Pricing and usage depend on the platform
MiniMax's published API reference price is about US$0.069 per standard 4K image, with a 50% discount on requests at 4K and above during the first week. How PixTV bills for it, and how many H3 Image calls are included in membership plans, is not listed on the platform's public pages for now; what you see after logging in is what applies.
For creators already making short dramas or ad spots on PixTV, the most direct change is that they no longer need to go elsewhere for character reference images. For anyone who has not used an AI video tool yet, a good start is generating one 4K image with H3 Image, then connecting it to a video node on the PixTV canvas to watch a still become a few seconds of footage. For e-commerce product imagery, PixPix has also launched H3 Image.
Sources: MiniMax official blog, CocoLoop, Hugging Face model card, PixTV platform page; used to verify H3 Image's resolution, reference-image limit and text-encoder lineage. API pricing follows the figures in the official announcement.