MiniMax H3 Max, board comments & Recraft styles
What's new?
MiniMax H3 Max — top-ranked video, with the sound already in it: MiniMax's post-trained H3 is now in the catalog, and it is currently at the top of the blind-preference video leaderboards (Design Arena and Artificial Analysis). What it changes for a game team:
- Audio is generated with the picture, not dubbed on after. Footsteps, impacts, ambience, and dialogue land on the frames they belong to, in native stereo — so a trailer beat or a cutscene is watchable straight out of the generation.
- It hits the beats in the order you wrote them. Prompt adherence is the model's headline improvement over base H3: a multi-action shot list plays back as a shot list, not as a vibe.
- Camera moves and character look hold across a shot. Directed pushes, tracks, and pans, with a character and an art direction that stay themselves from one cut to the next — the part that usually forces a re-roll.
- Ship-ready range. Text-to-video and image-to-video, 5–15 seconds, 480p, 768p, 2K, or 4K, six aspect ratios from 21:9 down to 9:16, and optional first-to-last-frame control for transitions you actually specify. Draft at 480p, then run the keeper at 4K.
Comments on the session board: Press C or use the toolbar to pin a thread on a tile or on the board itself. Reply, @mention teammates, resolve, and filter resolved threads — pins follow their tile as you rearrange the canvas.
Live presence and cursors: See who's in the session from the avatar stack in the header, and follow teammates' cursors on the board in real time.
Recraft V4 custom styles: Train a Recraft style from your reference images and generate with it in both raster and vector, on Standard or Pro. One training run covers both output kinds of that tier.
Background exports: Queuing a download no longer holds the dialog open. Exports run in the transfers panel alongside uploads, keep going if you close the dialog, and can be downloaded again from the finished row.
Auto-reload Creative Units: Opt in per workspace to top up automatically when the balance drops below a threshold you set, so a long session doesn't stall mid-generation. Enterprise and child workspaces stay on their existing billing.
A bigger asset viewer, without leaving the session: Opening an asset expands it into the session canvas instead of covering the whole page — header and chat stay put. In the viewer, both rails resize, the pager sits on the stage, and you can zoom to pixels.
Command palette finds everything: ⌘K searches projects, sessions, files, reference sets, workflows, and models in one ranked list. Hits on files, sets, and workflows open as shareable filtered URLs.
Video editor, faster to assemble: A CapCut-style rail for Image, Video, Audio, and Transition; right-click the timeline to add a clip at that frame; pick assets from browse and they land on a new track at the playhead. After a render, the board focuses the output.
Workflows: Segment Mesh is a node (the last 3D editor op that was missing one). And when a draft is better than the published original, Publish as new workflow forks it instead of overwriting.
Reference sets: A strength meter on the set editor shows how ready the set is to train. Artists can now create and manage sets without the training permission. Project tiles for workflows and sets offer Edit in place.
Auto aspect ratio: Models that can take their output size from the source image or first frame now offer Auto in the size picker, so image-to-video follows the source instead of snapping to a preset.
Home composer follows the file you attach: Drop a source video, mesh, or audio onto the matching tab and the composer switches to a model that can actually read it — with a one-click undo if you wanted the original model.
Bug fixes & improvements
Board layouts stick. Grouping, moving, resizing, and hiding tiles now survive leaving and returning to a session — including text-only groups and ungroup.
Stop means stop. A wedged workflow run terminates when you press Stop, instead of sitting in the session.
Downloads are named for the asset, not a random export key, and GIF originals export instead of failing the metadata path.
3D canvases size from their layout box, so posed meshes aren't clipped by a transformed rect.
Opening the viewer stops a playing thumbnail underneath it.
The composer shows the picked source video's preview frame, not a placeholder glyph.
Hidden models are no longer blocked by an opt-in workspace allowlist.
Content-policy chips on the board update live as verdicts land, and a project policy page is reachable with reference images as attachments.
Earn Free Units can show a Rewards tab for program grants, and the achievement Claim link lands on that dialog.
Skill pills in the agent drawer show which user skills are applied, and changing skill-loading mode no longer dead-ends.
Right-click in workflow node text fields opens the browser menu again; tile context menus follow the current selection.
New models9
VideoGoogle's multimodal video model — 720p, 1080p, or 4K clips with synchronized native audio from text or a still image.
VideoReference-guided video generation with up to 10 images and 3 short clips for consistent characters and style.
ImageRecraft's pro V4 tier, with higher fidelity and finer detail than the standard tier.
ImageRecraft's pro V4 text-to-vector tier, with higher fidelity than the standard vector tier.
ImageRecraft's V4 text-to-vector model. Generates scalable vector art (SVG) directly from a prompt.
ImageMeta's Muse Image model with faithful instruction-following and precise multi-reference editing.
ImageGoogle Virtual Try-On: dress a person photo with a clothing product image.
ImageMicrosoft's flagship image model for high-fidelity generation and native editing.
VideoMiniMax H3 Max post-trained video with native stereo audio at 480p, 768p, 2K, or 4K.
3D
Audio
