DreamX-World-5B: keyboard-controlled world video on Wan2.2-TI2V-5B (WASD/IJKL, up to ~1 min)
Image + prompt + action sequence -> camera-controlled video. Actions: w push in, s pull out, a/d move left/right, i/k tilt up/down, j/l pan left/right; combine as "wi", "wl", "dj". action_speed_list = relative segment durations. 704x1280, 16 fps; --num_output_frames in latent frames, divisible by 3 (123 -> 489 frames, ~30.6 s; 243 -> ~60.6 s). Needs Wan2.2-TI2V-5B components. MIT.
DreamX-World (AMAP-ML; arXiv 2606.16993) is a general-purpose interactive world model. DreamX-World-5B is its long-horizon autoregressive variant, built on Wan2.2-TI2V-5B: compared with the 5B-Cam variant it generates in causal chunks with KV-cache reuse, so rollouts of hundreds of frames are practical. It was trained on Unreal Engine renders, gameplay and real video with estimated cameras, then for action control, event response, RL-based action following and forcing/distillation for speed.
Inputs: a JSON list of items with image_path, caption (for example "Style: Photorealistic. <scene and desired world behaviour>"), action_seq and action_speed_list; the speed list is used as the relative duration of each action segment ([4, 6, 6] splits the rollout 4:6:6). Actions: w push in, s pull out, a move left, d move right, i tilt up, k tilt down, j pan left, l pan right; combine them in one string, e.g. wi (push in, tilt up), wj (push in, pan left), dj (move right, pan left). Actions become camera trajectories applied through PRoPE camera conditioning.
Running: clone github.com/AMAP-ML/DreamX-World and install its requirements (torch 2.5.1, torchvision 0.20.1, diffusers 0.30.1+, transformers 4.46.2+, xfuser 0.4.1, flash_attn 2.8.3, triton 3.1.0); download Wan-AI/Wan2.2-TI2V-5B for the text encoder (UMT5-XXL), tokenizer and VAE; point BASE_CHECKPOINT_PATH at the DreamX-World-5B .pt checkpoint; run inference_ar_forcing.sh or inference_ar_forcing.py (config causal_camera_forcing_5b.yaml, --fps 16, --seed, --color_correction_strength 1.0, --chunk_relative).
Length: --num_output_frames counts latent frames; pixel frames = (n - 1) x 4 + 1, and n should be divisible by 3 (the causal block size). 21 -> 81 frames (~5.1 s), 63 -> 249 (~15.6 s), 123 -> 489 (~30.6 s), 243 -> 969 (~60.6 s). Output is 704 x 1280 at 16 fps in BF16, using the Wan2.2 VAE (4x temporal, 16x spatial compression). Licence: MIT.
Based on the GD-ML/DreamX-World-5B card pinned to revision 67487c4a61466bb7166d30b7187dd465e0ac9f6c and observed on 2026-10-27.
The card states no GPU memory needs or speed; nothing was downloaded or run.
The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.
Structured fields
{
"citations": [
{
"url": "https://huggingface.co/GD-ML/DreamX-World-5B/raw/67487c4a61466bb7166d30b7187dd465e0ac9f6c/README.md",
"note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 1056:1446, 3996:4150, 4112:4150, 4331:4368, 3611:3826, 5243:5293, 5295:5408, 5930:6501, 5595:5617, 6684:6740; exact excerpt SHA256 c210605302a2f659024cb2d6402350a64d04f6d426106dd767f117c27c07724c,f2360669062ea65625704e7729f70db33aabeb714111b3c31bb6a67e995de36c,9c852213550815ee3"
}
]
}Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | permitted | unknown | Original attributed paraphrase and short functional values only. Source text, images, weights and other assets remain separately governed. |