Picsart CLI & MCP

Appearance

Changelog ​

Notable changes to the Picsart model catalog and these docs. Newest first. The catalog is served by @picsart/ai-sdk; the same models are reachable from the CLI, the MCP server, and Skills.

2026-11-03 ​

SDK 7.0.0 catalog refresh — 247 production models across 32 providers. Compared with the previous catalog: 3 models added.

New ​

  • MiniMax H3 Highspeed (video, t2v) — MiniMax H3 Highspeed — balanced quality and speed: 768p video from text, start/last frame, or image/video/audio references. See MiniMax.
  • MiniMax H3 Lite (video, t2v) — MiniMax H3 Lite — fastest and cheapest H3: 480p/768p video from text, start/last frame, or image/video/audio references. See MiniMax.
  • MiniMax H3 Max Recast (video, v2v) — Recast the people in a video with MiniMax H3 Max — one reference photo per new person, replacing the main people left to right while keeping the source motion, camera, cuts, and sound. Sources of 5-30s with no single shot over 15s, output up to 1080p. See MiniMax.

2026-11-02 ​

SDK 7.0.0 catalog refresh — 244 production models across 32 providers. Compared with the previous catalog: 1 model added.

New ​

  • Claude Opus 5.5 (text, i2t) — Anthropic’s latest Opus model — top-tier reasoning for complex, long-form tasks. See Anthropic.

2026-10-30 ​

SDK 7.0.0 catalog refresh — 243 production models across 32 providers. Compared with the previous catalog: 10 models added, 5 removed.

New ​

  • Flux 3 Max (video, t2v) — Text-to-video with synchronized audio, or image-to-video from a start (and optional end) frame. Up to 20s at 1080p. See Flux (Black Forest Labs).
  • Flux 3 Max Keyframes (video, i2v) — Animate a video between up to 10 positioned keyframe images, with synchronized audio. Up to 20s at 1080p. See Flux (Black Forest Labs).
  • Ideogram Vector (image, t2i) — Ideogram Vector — SVG logos, icons and illustrations from a prompt, optionally guided by up to 5 reference images. See Ideogram.
  • Ideogram Vectorize (image, i2i) — Ideogram Vectorize — convert a raster image (JPEG, PNG or WebP) into an SVG. See Ideogram.
  • Kimi K3.1 Pro (text, i2t) — Moonshot’s most capable Kimi model — post-trained K3 with deep reasoning and a 1M-token context. See Kimi.
  • Kimi K3.1 Pro Highspeed (text, i2t) — Kimi K3.1 Pro at 2–3× the speed — same capabilities, faster responses. See Kimi.
  • MiniMax M3 (text, i2t) — MiniMax’s flagship model — adaptive reasoning with a 1M-token context. See MiniMax.
  • MiniMax M3.1 Flash (text, i2t) — Fast, low-cost MiniMax model — always-on reasoning with a 1M-token context. See MiniMax.
  • GPT-6.1 Sol (text, i2t) — OpenAI’s most capable GPT-6.1 model for deep reasoning and complex tasks, with a 1M-token context. See OpenAI.
  • GPT-6.1 Terra (text, i2t) — Balanced GPT-6.1 model — strong general-purpose quality at half the price of Sol, with a 1M-token context. See OpenAI.

Removed ​

2026-10-16 ​

SDK 6.48.1 catalog refresh — 238 production models across 32 providers. Compared with the last published catalog (6.29.0): 16 models added, none removed; Kimi, Depth Anything and Mirelo join the provider list.

New ​

  • Kling 4.0 and Kling 4.0 Omni (video, t2v) — up to 30s at 720p or 1080p; 4.0 takes start, middle and end keyframes, Omni adds reference images, a reference or base video, and elements. See Kling.
  • ByteDance OmniHuman 2.0 (video, i2v) — animates one or more subjects with realistic body movement driven by audio. See ByteDance.
  • Grok Imagine 1.5 Lite (video, t2v) — faster, lower-cost video with audio, up to 15s. Grok 4.7 (text, i2t) — xAI's flagship reasoning model with a 500K-token context. See Grok.
  • Flux 3 Image Klein (image, t2i) — a distilled, faster Flux 3 Image with up to 10 references, a steps control and native 4K. See Flux.
  • MiniMax H3 Max Relight (video, v2v) — relights an existing clip from a sphere-render reference image. MiniMax H3 Context IR (text, v2t) — enhances a video prompt from text plus optional frames, references, videos and audio; returns a prompt, not a video. See MiniMax.
  • Claude Haiku 5.5 (text, i2t) — fast, lightweight Claude for high-volume text tasks. See Anthropic.
  • Gemini 4 Argon (text, v2t) — Google's flagship Gemini 4 reasoning model with image and video input and a 1M-token context. See Google.
  • Kimi K3 (text, i2t) — Moonshot's flagship reasoning model with a 1M-token context. New provider: Kimi.
  • SAM 3 Video (video, v2v) — segments and tracks named objects through a clip from a text prompt. See Meta.
  • Video Depth Anything (video, v2v) — temporally consistent per-frame depth maps, grayscale or color-mapped, optionally side by side. New provider: Depth Anything.
  • Mirelo SFX 2.0 (Text to Audio t2a, Video to Audio v2a, Video to Video v2v) — sound effects from a prompt, or a soundtrack synced to your video returned as audio or muxed onto the clip. New provider: Mirelo. Not selectable in the AI Playground app.

Changed ​

  • Flux Video Upscale is now displayed as Flux 3 Video Upscale. The id flux-video-upscale is unchanged.

Availability ​

The published @picsart/gen-ai CLI (2.82.0) and the hosted MCP server still run an older catalog. Of the models above, only MiniMax H3 Max Relight, Claude Haiku 5.5 and Video Depth Anything are callable from them today; the rest become reachable when the CLI and MCP server move to SDK 6.48.1. The SDK itself exposes every model listed.

2026-10-07 ​

SDK 6.29.0 catalog refresh — 222 production models across 29 providers. Adds Nano Banana 2.1, HeyGen Video, and TypeSafe Jev structured evaluation. Four older Luma Ray 2 / Flash 2 entries leave the production catalog. The catalog and navigation now include JSON output; Jev uses the SDK’s ai.run() API. See the complete update and availability notes.

2026-10-02 ​

SDK 6.23.0 catalog refresh — 223 production models across 28 providers. The comparison with the last published catalog adds 22 models and removes none. Additions include Flux 3 Image, Ideogram 4.5, GPT-6, LTX 2.5, Eleven v4, and Gemini 3.8 speech models. Seedance, Seedream, and Seed Audio now sit under ByteDance; their model IDs remain unchanged. Provider tables include current parameters. Preview models remain excluded. See the complete model list and availability notes.

2026-09-29 ​

MCP docs corrected to match the live server. The Picsart MCP server is hosted at https://api.picsart.com/gen-ai/mcp. Agents connect to it over HTTP and sign in to Picsart in the browser, so it needs no gen-ai CLI, no gen-ai login and no API key. Earlier pages described a local gen-ai-mcp command, required the CLI as a prerequisite, or pointed some integrations at a different address with an API key. All of those instructions were wrong and have been replaced on the MCP Quickstart, Installation, Authentication and every integration guide. The tool catalog also drops picsart_music_studio (not available), adds the viewing and progress tools, and fixes the picsart_remove_bg example (it takes one image URL). AnythingLLM Desktop now connects through the mcp-remote bridge, and the integrations index lists every supported agent plus the ones that cannot connect yet (NemoClaw, AnythingLLM Docker).

CLI docs checked against @picsart/gen-ai 2.78.0. Removed flags and commands the CLI does not have (--script, --dry-run and --no-download on generate, short flags on pricing, gen-ai extend), documented that generations save to Drive by default, rewrote the batch guide for the real JSON manifest format (jobs[] with id) and output folder, and added CI authentication via PICSART_ACCESS_TOKEN / PICSART_USER_ID.

2026-09-19 ​

SDK 6.2.3 catalog refresh — 201 models across 31 providers. Compared with the previously published catalog: 26 additions, 6 removals; Meta joins and Reve leaves the active provider list. The SDK upgrade from 6.0.0 adds 11 text models and changes result metadata and Seedance color-depth handling. See the complete update and migration guide.

2026-09-01 ​

Catalog refresh — now 181 models across 31 providers (@picsart/ai-sdk 5.24.0). These docs were last generated from 5.19.0, so this refresh spans 5.20–5.24: three video/image models added, one renamed, none retired.

New ​

  • Flux Video Upscale (video, v2v — added in SDK 5.20.0) — upscales an existing clip toward 4K at 1.5x–3x, in either precise (source-faithful) or creative (detail-enhancing) mode. Source clips up to 20 seconds and 2K. See Flux.
  • MiniMax H3 Max (video, t2v — added in SDK 5.22.0) — the top tier of the H3 family: text-to-video or start/end-frame interpolation with a promptExpansionMode pass, up to 15s at 768p. Image-to-video is available as an edit workflow. See MiniMax.
  • Qwen 3.0 Pro (image, t2i — added in SDK 5.22.1) — the GA flagship Qwen-Image 3.0 Pro, covering text-to-image and image editing with prompt-rewrite modes (promptExtendMode) and an optional enableThinking pass. See Qwen.

Changed ​

  • Hailuo 03 is now MiniMax H3 (SDK 5.21.1) — the catalog id changed from hailuo-03 to minimax-h3. This moves a public identifier: ?model=hailuo-03 deeplinks and any saved reference to the old id no longer resolve. The workflow (minimax/v2/video-generation) and pricing key (minimax-h3) are unchanged, and no parameters moved.
  • Gemini Omni 1.2 Flash is now displayed as Gemini Omni 1.1 Flash (SDK 5.21.2) — this resolves the id/name mismatch flagged in the 2026-08-27 entry above. The id was always gemini-omni-1.1-flash-preview; the display name has been corrected to agree with it rather than the id being changed, so nothing matching on the id is affected.

2026-08-27 ​

Catalog refresh — now 178 models across 31 providers (@picsart/ai-sdk 5.19.0). These docs were last generated from 5.16.0, so this refresh spans 5.17–5.19: two video models added, none retired.

New ​

  • Wan 3.0 Prime (video, t2v — added in SDK 5.18.0) — the same all-in-one model as Wan 3.0, up to 7x faster. Takes image, video and audio references plus start/end frames; 5/10/15/30s at up to 1080P, adaptive aspect ratio, native audio, and an optional --enable-thinking pass. See Wan.
  • Gemini Omni 1.2 Flash (video, t2v — added in SDK 5.19.0) — the more capable of the two Gemini Omni entries. Over the original it adds a resolution param at all (360p · 720p · 1080p · 4k, default 720p), start/end frame interpolation, up to 5 reference images and 3 reference videos, clip extension from a source video under 30s, and every integer duration 3–10s rather than a fixed set. See Google.

NOTE

Gemini Omni 1.2 Flash's id is gemini-omni-1.1-flash-preview — the id says 1.1, the display name says 1.2. That mismatch comes from the upstream catalog, so anything matching on the id must use the 1.1 string. -preview in the id is likewise historical; the model is a production release.

Changed ​

  • Grok Imagine 1.5 is now classified t2v (was i2v), as of SDK 5.18.0 — its image input is optional, so it generates from a prompt alone. The CLI and MCP call is unchanged; only the input-type label moved. See Grok.

Docs ​

  • The provider-page generator now updates the frontmatter model count on multi-mode provider pages. Its regex matched only single-word mode lists, so slash-joined descriptions ("17 image/video/audio/text model(s)") silently kept a stale count — Google had been carrying one across refreshes.
  • The same generator no longer appends a blank line at EOF on every run for pages whose ## Parameters is the last section, so repeat refreshes are now byte-idempotent.

2026-08-25 ​

Catalog refresh — now 176 models across 31 providers (@picsart/ai-sdk 5.16.0). Eleven models added, nine retired.

New ​

  • Wan 3.0 — an all-in-one video model taking text, image/video/audio references, and start/end frames, with adaptive ratio, intelligent duration, and audio. See Wan.
  • Picsart Effects (image + video) — picsart-flow and picsart-flow-video apply curated Picsart presets as multi-step Magic Flow pipelines in one call. Preset ids come from a runtime catalog; list them with gen-ai models info picsart-flow --json. See Picsart.
  • Recraft V4 Styles — four style-reference models (Styles, Styles Vector, Styles Pro, Styles Pro Vector) with 10K-character prompts. See Recraft.
  • Grok Imagine 2.0 — sharper detail with a low/medium quality tier. See Grok.
  • MiniMax Music v3 — text-to-music with vocals or instrumentals from a style prompt and optional lyrics, with configurable audio encoding. See MiniMax.
  • Gemini 3.7 Flash — low-latency multimodal text generation for gen-ai describe. See Google.
  • ByteDance Video Enhance — now live, replacing the Video Upscaler: denoise, colour-correct and super-resolve footage up to 8K, with frame-rate conversion. See ByteDance.

Deprecated / retired ​

  • Pika is retired. pika-2.2, pika-2.2-scenes and pika-2.2-frames are gone and Pika is no longer a provider. The provider page is kept as a pointer to replacements — Seedance 2.5 covers multi-reference composition and start/end-frame morphs.
  • Google Imagen 4.0 family — imagen-4.0, imagen-4.0-ultra and imagen-4.0-fast removed; use the Nano Banana (Gemini Image) models instead.
  • Also removed — bytedance-video-upscaler (superseded by Video Enhance), kling-v2-new-image, and qwen-image-edit-plus (Qwen 2 Pro now covers reference-guided editing from up to 3 images).

Docs ​

  • Provider pages now render catalog-kind parameters (voice, avatar, and effect-preset pickers) instead of leaving the Values cell blank. These lists are fetched at runtime, so the docs point at gen-ai models info <id> --json rather than snapshotting a list that would rot.
  • The count drift guard now also covers reference/providers/index.md and README.md, and recognises the "N AI model providers" phrasing.
  • The wiki generator keeps pages for retired providers, so a tombstoned page like Pika exists on both the Pages site and the Wiki instead of silently vanishing from one.

2026-08-07 ​

Catalog refresh — now 174 production models across 32 providers (@picsart/ai-sdk 3.35.6).

New ​

  • Seedance 2.5 — ByteDance's next-generation video model, with companion Seedance 2.5 Video Edit (edit an existing clip from a prompt) and Seedance 2.5 Video Extend (continue a clip past its original end). See Seedance.
  • Seedream 4.7 — text-to-image up to 4K with multi-image reference input and batch counts up to 10. See Seedream.
  • Flux 3 Video — Black Forest Labs' first video model on Picsart: text-to-video and image-to-video up to 20 seconds at FHD, with native audio generation and a fast --draft mode. This makes Flux a multi-mode provider. See Flux.
  • HeyGen Video Avatar — pick a HeyGen avatar and voice, supply a script, and get a lip-synced presenter video up to 4K. Unlike Talking Photo it needs no input image. See HeyGen.

2026-07-31 ​

AI Playground · July 31 Release — catalog refreshed to 168 production models across 32 providers (@picsart/ai-sdk 3.30.0).

New ​

  • Hailuo 03 — MiniMax text-to-video and image-to-video with start/end frames, multimodal references, 5–15 second clips, and output up to 2K. See MiniMax.
  • Video Enhance — ByteDance video enhancement can denoise, color-correct, super-resolve existing footage up to 8K, and convert frame rate. The SDK definition is gated until its worker rollout completes.
  • Ideogram P-Image — out of preview and generally available, with four speed/quality tiers and output up to 2K. See Ideogram.
  • HEIC support — iPhone and Windows HEIC/HEIF photos now upload with automatic conversion.
  • Catalog additions — Seed Audio and Seed Audio Multilingual add named voices and voice cloning, while Gemini 3.6 Flash and Gemini 3.5 Flash Lite expand text/image analysis. See Seed Audio and Google.

Improvements ​

  • Empty boards — redesigned with a card-stack hero and a responsive “Need a spark?” inspiration strip.
  • Explore — the Picsart Effects strip now serves the creator and business catalog: about 950 presets across 16 categories.
  • Topaz catalog — Topaz Video Upscale is available with Proteus, Artemis, Nyx, Gaia, and Starlight enhancement families; eight legacy image-enhancement IDs are consolidated as presets on topaz-upscale-image.

2026-07-17 ​

Catalog refresh — now 169 models across 31 providers (@picsart/ai-sdk 3.17.0).

  • New models — Seedream 5.0 Pro (text-to-image, see Seedream), Nano Banana 2 Lite (gemini-3.1-flash-lite-image, see Google), ElevenLabs Music v2 (full music tracks, see ElevenLabs), and Picsart Angle Change (picsart-qwen-image-edit-angle, image editing, see Picsart).
  • Recraft V4 & V4.1 image-to-image. All 12 V4/V4.1 models (base, Pro, Utility, Utility Pro, and their vector variants) now take an optional source image (-i) with an adjustable image weight (--weight, 0–100, default 80). See Recraft.
  • Happy Horse 1.1 — model ids renamed from happyhorse-1.5-* to happyhorse-1.1-* (same models; update any saved commands).
  • Parameter updates — Grok Imagine Video 1.5 adds 1080p; HeyGen Talking Photo prompt limit raised to 5000 chars; Ideogram Character now takes a required prompt; Seedance 2.0 reference images/videos must be at least 0.4MP; the ByteDance Video Upscaler now rejects sources at or above 1080p up front (it only upscales sub-1080p video).

2026-06-26 ​

Catalog refresh — now 165 models across 31 providers (@picsart/ai-sdk 3.10.0). A big lineup update since the 3.6.2 catalog:

  • Seedance 2.0 family — the full lineup. Seedance 2.0 and Seedance 2.0 Fast (text-to-video with keyframe control and native audio), each with a Video Edit and Video Extend variant, plus the new Seedance 2.0 Mini (T2V + edit + extend). See Seedance.
  • Text & analysis models (new mode) — analyze images and video with LLMs via gen-ai describe: Anthropic (new provider) Claude Opus 4.8 / Sonnet 4.6 / Haiku 4.5, OpenAI GPT-5.5, and Google Gemini 3 Pro (the only one that reads video). See Text & analysis.
  • Google — added the Imagen 4.0 family (4.0 / Ultra / Fast) and Gemini Omni.
  • Happy Horse — added Happy Horse 1.5 (text-to-video and ref-to-video).
  • Deprecated / retired — removed from the catalog: GPT Image 1, OpenAI TTS-1 / TTS-1 HD, the Kling image & multi-image family, LTX Pro / Fast / Retake, Wan 2.6 (T2V / R2V / Image), Seedream 4.0, Runway Gen-3 Turbo & Aleph, Recraft v2 (+ Vector), Qwen v1, and Seedance 1.5 Pro / I2V. These stay resolvable for historical jobs and pricing but are no longer offered for new generations.

New CLI command — gen-ai describe. Analyze an image or video with an LLM (Claude, GPT, or Gemini) and get a text answer — caption, OCR, classify, or summarize a clip. The prompt is optional, video auto-routes to a video-capable model, and output goes to stdout for easy piping. See the CLI Quickstart.

2026-06-19 ​

Try in Playground links. Every model now has a direct Try in Playground link — on each provider page's parameter section and on every card in the interactive Model Catalog — that opens the web Playground with the model preselected.

2026-06-18 ​

Catalog — now 176 models across 30 providers (@picsart/ai-sdk 3.6.2). 36 models added since the previous catalog:

  • Kling — Kling V3 Turbo, plus the Kling image & multi-image family (V2 / V2.1 / V1.5 Image, Multi-Image, Multi-Image V2.1).
  • PixVerse (new provider) — V6 and C1 lines: text-to-video, image-to-video, and Fusion reference-to-video.
  • LTX — LTX Pro, LTX Fast, LTX Retake.
  • Wan — Wan 2.6 (text-to-video), Wan 2.6 Ref-to-Video, Wan 2.6 Image.
  • Luma — Ray 3.2, Ray 3.2 Edit, Ray 3.2 Reframe.
  • OpenAI — GPT Image 1, plus TTS-1 and TTS-1 HD (text-to-speech).
  • Seedance — Seedance 1.5 Pro, Seedance I2V.
  • Runway — Gen-3 Alpha Turbo, Aleph 2.
  • Recraft — Recraft 20B and Recraft 20B Vector.
  • Ideogram — Ideogram 4.0. · Seedream — Seedream 4.0. · Qwen — Qwen. · Grok — Grok Imagine 1.5.
  • Async (new provider) — Async Flash v1.0 (text-to-speech, 100+ voices).
  • Picsart — Remove Background now runs on picsart-sod-v8-2 (replaces the previous picsart-remove-bg).

Docs — Public documentation site launched: CLI, MCP, and Skills guides, a searchable Model Catalog, and a reference page per provider with parameters and CLI + MCP examples.

Looking for the full catalog right now? Browse the Model Catalog or the Providers grid — both are generated from the live catalog.

Built on @picsart/ai-sdk · gen-ai CLI · Picsart MCP · Media Studio · Skills

© Picsart