New AI models in Picsart AI Playground, from video to songs

The newest models in Picsart AI Playground cover every step of a project: MiniMax H3 for 2K video with sound, GPT Image 2.5 for precise image edits, and Lyria 3.5 for full songs.

All of them sit behind one prompt bar, next to more than 220 AI models from 28 providers. Here is what each new model does best, and how to combine them.

What's new in Picsart AI Playground

Picsart AI Playground keeps raising the bar for what creators can produce in a single workflow, and the latest additions reshape how creators approach full-scale content production. From native 2K video with the sound already on it to precise image edits and full songs with vocals, everything now lives in one place.

That means no juggling tools, no switching tabs, and no piecing things together across platforms. These new AI Playground models span video, image, music and voice, making it possible to build complete projects from scratch.

The newest additions include MiniMax H3, WAN 3.0, Seedance 2.5, GPT Image 2.5, Seedream 5.0 Pro, Ideogram 4.5 and Lyria 3.5, and together they expand the Playground into a complete creative pipeline. You can move from idea to image, from image to video, and from video to soundtrack without ever leaving the same interface. What once required multiple subscriptions, tools and exports now happens inside a single prompt bar.

Below is what is new, how each model works, and how to combine them into fully AI-generated content, from short-form social clips to full music videos and branded campaigns.

The new models at a glance

New video models

The latest video models bring a major leap in resolution, length and finished sound.

MiniMax H3 and H3 Max

MiniMax H3 generates native 2K video and writes the audio in the same pass. Dialogue, ambient sound and effects land against the action on their own, so a clip arrives finished instead of silent and waiting for post.

• Native 2K at 24 fps, up to 15 seconds
• Takes reference images, video and audio
• Multi-shot storytelling from text or an image

Its sibling, MiniMax H3 Max, is the speed-tuned version. It drops references, renders at 480p or 768p for 5 to 15 seconds, and returns a clip in seconds rather than minutes, which turns generation into something you iterate with.

WAN 3.0

WAN 3.0 generates native 30-second video in a single continuous take. The more useful part is that you no longer have to pick the number: describe the action and smart duration reads the pacing you implied, then sets the length to match.

Because the clip runs unbroken, light, motion and pacing hold all the way through instead of resetting at a splice.

• Inputs: text, image, audio and video, plus a document or a web page
• Extend any finished clip and keep building from a take that already works

Seedance 2.5

Seedance 2.5 changes the economics of iteration rather than the output. Draft Mode splits generation into a fast 480p preview and a connected 1080p final, so you only pay full price for the take you keep.

Test composition, motion, dialogue and camera moves in low-cost drafts, then render your favourite in 1080p from the same draft ID. Seedance reuses its prompt, references and seed, so the final keeps the take you approved.

Also new in video

Three more arrived alongside them:

• Flux 3: Black Forest Labs' first fully multimodal model makes a clip of up to 20 seconds and its synchronised soundtrack from one prompt, and accepts up to 10 image references.
• LTX 2.5: two tunings, both with sound in the same pass. Fast renders up to 4K and 20 seconds; Pro stops at 1080p and 10 seconds and puts the difference into per-frame quality.
• Gemini Omni 1.1 Flash: takes Google's Omni from 720p to 4K, with audio generated alongside the frames.

New image models

The new image models focus on control, consistency and range.

GPT Image 2.5

GPT Image 2.5 was tuned to change only the element you name and hold everything around it steady. Editing is where it makes its biggest leap, and the stability holds across a conversation: an earlier edit survives the next one instead of the picture degrading over a long back and forth.

• Up to 4K, with a 3840px maximum edge
• Up to 16 reference images
• Two variants: Flare tuned for speed, Sunburst tuned for precision

Seedream 5.0 Pro

Seedream 5.0 Pro wins three jobs outright. It writes legible text into an image instead of letter-shaped marks, it holds the same character across a whole set of images, and it edits a photo you already have from a plain description.

Holding one character across a set is what keeps a cast recognisable across a whole project.

• 1K up to 4K at standard output, and as far as 8K

Ideogram 4.5

Ideogram 4.5 pairs Ideogram's signature text rendering with drift-free editing: it changes only what you ask and leaves the rest of the image pixel-intact, edit after edit.

Its companion, Ideogram 4.5 Precise Edit, is built for targeted fixes, from swapping an element to correcting a line of text.

Also new in images

The rest of the new image line-up covers very different jobs:

• Qwen Image 3 Pro: Alibaba's flagship has no low-quality setting. Every resolution runs at roughly 4.2 megapixels, so text on a sign or panels in a storyboard survive.
• Muse Image 1.0: Meta's agentic image model plans the composition, looks up references it does not already know and checks its own work before handing you a result.
• Recraft V4.1 Flash: returns an image in about 1.3 seconds from a prompt alone, while holding the Recraft look.
• Grok Imagine 2: makes one precise frame from one reference image, with Grok Imagine Quality for images that have to survive a close look.

New music and voice models

Lyria 3.5

Lyria 3.5, Google DeepMind's latest music model, completes the pipeline. Describe the genre, mood, instruments, vocal style, tempo or duration you want and it turns those details into a cohesive piece of music.

• Richer arrangements and expressive vocals
• Structured lyrics
• Tracks up to three minutes long

Also new in audio

• MiniMax Music v3: the other new route to a full song, with vocals or instrumentals and your own lyrics if you have them.
• Gemini 3.8 Flash TTS: for the spoken part, it performs your script in a voice you describe, across 130+ languages, directed line by line.

Together, these models redefine what all-in-one creation actually means.

How to make a fully AI-generated video in Picsart AI Playground

Creating a complete AI-generated video now follows a simple three-step process. Each step builds on the previous one, turning a single idea into a finished piece of content.

Step 1. Build your visuals

Start with Seedream 5.0 Pro to set the look and your hero subject. A prompt like "Young musician, curly hair, leather jacket, electric guitar, city rooftop at sunset, warm golden light" gives you a character it can hold across a whole set of images, which is what keeps storytelling continuous.

When one frame is nearly right, take it to GPT Image 2.5 and change the single thing that is wrong while the rest of the frame holds.

Step 2. Bring it to life

Use WAN 3.0 for establishing shots and atmospheric visuals, where a single unbroken 30-second take holds its light and pacing all the way through. For moments that need sound and references, switch to MiniMax H3, which delivers 2K with the dialogue and effects already on it.

For long dialogue-driven scenes, Kling 3 Omni still runs to two minutes. If you are iterating a lot before committing, Seedance 2.5 is the cheaper way round: draft in 480p until the take is right, then render that exact take in 1080p.

Testing the same prompt across different models reveals subtle differences in tone, motion and realism, and the Playground is built to make that comparison easy.

Step 3. Score it

Finish with Lyria 3.5 or MiniMax Music v3. Describe the mood, the instruments and the tempo, and score the piece you just built rather than searching a stock library for something that nearly fits. A prompt like "Indie pop track, 90 BPM, melancholic but hopeful, layered guitars and synths" gives you a soundtrack shaped to your visuals.

If it needs a narrator, Gemini 3.8 Flash TTS reads the script in the voice you describe.

You can go one of two ways from here. Generate clips with the sound built in, using MiniMax H3, Flux 3 or LTX 2.5, or assemble separate clips and music inside the Picsart Video Editor for more control.

Power combos: the best model pairings in Picsart AI Playground

Each model in the AI Playground has a distinct strength, but the real creative power shows up when you combine them strategically.

Brand campaign or product launch

Start with Seedream 5.0 Pro to define your visual direction and a cast that stays consistent across the set. Polish the product images with GPT Image 2.5, changing one element at a time while the rest of the frame holds. Animate the reveal with WAN 3.0 in one unbroken take, then add a bed with Lyria 3.5.

This combination keeps your visuals, motion and audio aligned from start to finish.

Music video

Begin with Lyria 3.5 or MiniMax Music v3 to generate the full track. Build the performer with Seedream 5.0 Pro, then shoot the performance with MiniMax H3, which writes the sound in the same pass as the picture. Use WAN 3.0 for the long atmospheric shots in between.

This approach flips the traditional workflow: the music drives the visuals instead of the other way around.

Short film teaser or narrative scene

Create your lead character with Seedream 5.0 Pro, then bring them to life with MiniMax H3 for the character beats, dialogue and effects included. Use WAN 3.0 for environment and continuity, and Kling 3 Omni when a dialogue scene needs to run long. Finish with Lyria 3.5 for the score.

Short-form social content

Generate a strong visual hook with Recraft V4.1 Flash, then animate it with Seedance 2.5, drafting cheaply in 480p until the timing is right and rendering the keeper in 1080p. When speed matters more than references, MiniMax H3 Max returns a clip in seconds rather than minutes.

This combo is fast and efficient, built for the pace social content runs at.

Concept board for client review

Recraft V4.1 Flash shines here: an image in about 1.3 seconds means you can explore directions as fast as you can react to them. Use Qwen Image 3 Pro wherever the board carries real copy, since text survives at its 4.2 megapixels, and GPT Image 2.5 to adjust a single element without disturbing the rest.

This workflow helps you explore broadly before committing to a final direction.

What you get when every model lives in one Playground

The point is not the individual models. It is that the handoffs disappear.

An image made in one model becomes the reference for a video in another, and that video gets scored by a third, without an export, a re-upload or a second subscription. Comparing two models on the same prompt takes one more click rather than one more account.

When a new model lands, it lands in the same prompt bar as the rest, which is why a list like this one is worth rereading in a few months. Everything generated stays in one library, so a project is a set of related assets rather than a folder of downloads. And every model runs on one credit balance, so trying a second model costs a generation, not a subscription.

Start creating with the newest AI models

Picsart AI Playground keeps expanding, and with MiniMax H3, WAN 3.0, Seedance 2.5, GPT Image 2.5, Seedream 5.0 Pro, Ideogram 4.5 and Lyria 3.5, the entire creative pipeline now lives in one place. From first idea to final export, everything happens in a single workflow.

Open Picsart AI Playground, pick the model that matches the job, and run the same prompt through a second one to see what you would have missed. If the output is heading for a timeline, the Picsart Video Editor is where it gets cut.