ModelListedExternal

Phantom Wan 1.3B

Apple-MLX port of Phantom-Wan-1.3B (ByteDance, ICCV 2025) — multi-subject subject-to-video (S2V): compose up to 4 distinct characters from reference images into one coherent generated video. Runs natively on Apple...

Sourcehuggingface.co/mlx-community/Phantom-Wan-1.3B/tree/48c677afbd28494c78277c43e8aa2ae35ceed329

KindText modelBase modelbytedance-research/PhantomVersiont5_encoderBgrade

Model source

Kind: Text model
File size: 10.6 GB
Base model: bytedance-research/Phantom
Version: vt5_encoder
Updated: Oct 16, 2026
Downloads: 69
Source: Hugging Face

Source description

Phantom-Wan-1.3B — MLX

Apple-MLX port of Phantom-Wan-1.3B (ByteDance, ICCV 2025) — multi-subject subject-to-video (S2V): compose up to 4 distinct characters from reference images into one coherent generated video. Runs natively on Apple Silicon. Apache-2.0.

Phantom-Wan-1.3B is a fine-tune of the stock Wan2.1-T2V-1.3B DiT, so this port rides the mlx-video Wan2.1 substrate; the only net-new surface is the reference-injection path (refs encoded as trailing temporal latent frames + dual-scale chained CFG).

Contents (self-contained)

File	What
`transformer-bf16.safetensors`	DiT, bf16 (2.84 GB)
`transformer-4bit.safetensors`	DiT, int4 group-size 64 (0.98 GB) — embeds/time/head kept hi-precision, cosine 0.99633 vs bf16
`vae-encoder.safetensors` / `vae-decoder.safetensors`	Wan2.1 16-ch VAE (bf16)
`t5_encoder.safetensors`	umT5-XXL text encoder (bf16)
`config.json`	architecture + sampling defaults

Features

Multi-subject (≤4 refs) composition from reference sheets.
Lossless streaming VAE decode — peak memory flat ~20 GB at any video length; 81-frame output decodes fine (whole-sequence decode OOMs past ~49 frames).
Dual-scale chained CFG (guide_img=5.0, guide_text=7.5), FlowUniPC scheduler (shift 5, 50 steps).

Usage

from phantom_wan_mlx import pipeline_mlx as P   # github.com/xocialize/phantom-wan-mlx
P.s2v(
    "two friends walking together in a park",
    ["subjectA.png", "subjectB.png"],
    "out.mp4",
    size=(832, 480), frame_num=81,
)

Parity

DiT loads turnkey into the mlx-video WanModel (825/826 keys; the gap is the computed RoPE buffer); forward + VAE numerics inherited from the published mlx-video / Bernini-R Wan2.1 substrate. Streaming decode is bit-exact to whole-sequence decode on mx.cpu.

License

Apache-2.0 (mirrors upstream Phantom + Wan2.1). Port by MVS Collective (xocialize-code).

Sources

1 source

Hugging FacePrimaryVersion 48c677afbd28494c78277c43e8aa2ae35ceed329

huggingface.co/mlx-community/Phantom-Wan-1.3B/tree/48c677afbd28494c78277c43e8aa2ae35ceed329

Verified Aug 16

Model artifacts

1 artifact

t5_encoder.safetensors

safetensors · 10.6 GB · SHA-256 e86ee4199903…3787 · Hugging Face

Download

Source excerpts

2 excerpts

ReadmeAug 16huggingface.co/mlx-community/Phantom-Wan-1.3B

--- license: apache-2.0 library_name: mlx tags: - mlx - text-to-video - subject-to-video - video-generation - phantom - wan base_model: bytedance-research/Phantom pipeline_tag: text-to-video --- # Phantom-Wan-1.3B — MLX Apple-MLX port of **[Phantom-Wan-1.3B](https://github.com/Phantom-video/Phantom)** (ByteDance, ICCV 2025) — **multi-subject subject-to-video (S2V)**: compose up to 4 distinct characters from reference images into one coherent generated video. Runs natively on Apple Silicon. Apache-2.0. Phantom-Wan-1.3B is a fine-tune of the stock **Wan2.1-T2V-1.3B** DiT, so this port rides the [`mlx-video`](https://github.com/Blaizzy/mlx-video) Wan2.1 substrate; the only net-new surface is the reference-injection path (refs encoded as trailing temporal latent frames + dual-scale chained CFG). ## Contents (self-contained) | File | What | |------|------| | `transformer-bf16.safetensors` | DiT, bf16 (2.84 GB) | | `transformer-4bit.safetensors` | DiT, int4 group-size 64 (0.98 GB) — embeds/time/head kept hi-precision, cosine 0.99633 vs bf16 | | `vae-encoder.safetensors` / `vae-decoder.safetensors` | Wan2.1 16-ch VAE (bf16) | | `t5_encoder.safetensors` | umT5-XXL text encoder (bf16) | | `config.json` | architecture + sampling defaults | ## Features - **Multi-subject (≤4 refs)** composition from reference sheets. - **Lossless streaming VAE decode** — peak memory flat ~20 GB at any video length; 81-frame output decodes fine (whole-sequence decode OOMs past ~49 frames). - Dual-scale chained CFG (`guide_img=5.0`, `guide_text=7.5`), FlowUniPC scheduler (shift 5, 50 steps). ## Usage ```python from phantom_wan_mlx import pipeline_mlx as P # github.com/xocialize/phantom-wan-mlx P.s2v( "two friends walking together in a park", ["subjectA.png", "subjectB.png"], "out.mp4", size=(832, 480), frame_num=81, ) ``` ## Parity DiT loads turnkey into the mlx-video `WanModel` (825/826 keys; the gap is the computed RoPE buffer); forward + VAE numerics inherited from the published mlx-video / Bernini-R Wan2.1 substrate. Streaming decode is **bit-exact** to whole-sequence decode on `mx.cpu`. ## License Apache-2.0 (mirrors upstream Phantom + Wan2.1). Port by MVS Collective (xocialize-code).

NoteAug 16huggingface.co/mlx-community/Phantom-Wan-1.3B

Source context: 69 downloads · 0 likes · Pipeline text-to-video · Library mlx · Repo mlx-community/Phantom-Wan-1.3B