SSyncDev

PAN 2

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

PAN 2 is IFM's open world model: from a scene and text actions it generates up to 20 s of 704x1248 video of the predicted outcome, on a 29B diffusion core.

Key specifications

Specification score
15.0
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Video output

Capabilities

VisionOpen weights

What is PAN 2?

Describe a kitchen, say "drain the pasta", and get a video of what happens next. PAN 2 is MBZUAI's Institute of Foundation Models' open world model, and its job is prediction rather than pretty footage: given a scene and a sequence of text actions, it renders the likely outcome.

Code and weights are Apache-2.0.

Two kinds of reasoning

The system splits thinking from drawing.

  1. Conceptual reasoning. An instruction-following language model rewrites your input into a structured world-state caption. IFM evaluated several, including Kimi K3, which it used for most studies.
  2. Perceptual reasoning. A 29B MMDiT latent diffusion model, trained from scratch and released as IFM/PAN2, turns that caption into video. It reads text through a frozen Qwen3.5 9B encoder and decodes through the Wan 2.2 VAE.

Free-text prompts work, but IFM says results are best when the reasoner rewrites them first.

Output and defaults

Item Value
Resolution 704x1248
Frame rate 24 fps
Length 241 frames (~10 s) by default, up to 481 (~20 s)
Inputs Optional first image, text description, text actions
Sampler 50 steps, flow shift 9, guidance 2.5, BF16 transformer
Engines Diffusers, vLLM recipe, local inference script

No GPU memory figure is given. Each worker holds the full 29B model, one worker per GPU, so plan for large datacentre cards.

Safety is a separate install

IFM strongly recommends its pan2-guardrail package. It screens requests with a 4B guard model before generation and checks frames afterwards. In IFM's tests it cut unsafe text-to-video outputs from 33.9% to 1.5% and produced none for image-to-video, at the cost of blocking 11.5% of benign text-to-video requests. That false-positive rate is high enough that you should tune it on your own prompts.

Honest limits

The list is long, and worth reading before you plan anything:

  • about two in three simulated long plans skip at least one step;
  • output shifts with prompt wording;
  • appearance drifts across clips in long sessions;
  • coarse actions work better than dexterous ones;
  • physics can be skipped, such as pasta in a colander before it is drained;
  • realistic output can be mistaken for real footage.

Benchmark results on physics and robotics suites are published as charts only, so we quote no figures.

Dating

Work on PAN 2 was public from 7 October 2026, but the public launch came on 10 November. This page dates the model from the launch, so expect earlier mentions to reference an October release.

How to evaluate it

Script ten multi-step tasks from your domain, such as a warehouse pick or a lab procedure, with three to five actions each. Generate each twice and score step completion, physical plausibility and consistency between clips. That tells you whether PAN 2 is a planning aid or still a research tool for your use. Related models are under image-text-to-video and text-to-video.

How to evaluate PAN 2

Compare the model on the work you need to complete.

Tasks and inputs

PAN 2 is listed for Image-text-to-video, Text-to-video. It accepts text, image and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Video at 704x1248 and 24 fps showing the predicted outcome of text actions applied to a described scene, about 10 seconds by default and up to 20.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

xAI

Grok Imagine Video 1.5 Lite

Grok Imagine Video 1.5 Lite generates video from text or an input image, with resolution-based pricing on the xAI API.

— context

MiniMax

MiniMax H3

MiniMax H3 is a multimodal generation system that produces video with synchronized stereo audio from supported text, image, video and audio references.

— context

ByteDance Seed

Seedance 2.5

Seedance 2.5 from ByteDance Seed: ByteDance multimodal audio-video model with up to 30-second single-pass generation, multi-round extension, reference inputs and timestamp-level editing.

— context

Google

Gemini Omni Flash Preview

Gemini Omni Flash Preview from Google: Video generation and editing model for fast, conversational text- and image-to-video workflows.

1.05M context

xAI

Grok Imagine Video 1.5

Grok Imagine Video 1.5 creates video from text, image references and supported voice inputs, with frame controls for planned shots.

1.02K context

Google

Veo 3.1 Lite Preview

Veo 3.1 Lite Preview from Google: Video model for prompt-guided generation, editing, and motion workflows.

1.02K context