None of these are API nodes. All three run open weights locally on Floyo's H100 GPUs and bill FloTime GPU credits, so cost tracks generation time rather than clip count. Every model file is pre-loaded — no downloads, no VRAM ceiling to plan around.

Benchmark 1 · Full-body dance

Fast, continuous motion with arm and hand detail

Benchmark 2 · Character replacement

Swapping a subject into existing footage

Benchmark 3 · Stylised character

Illustrated or 3D-rendered subject, not a photo

Benchmark 4 · Long take

Ten seconds or more of continuous motion

These are capability reads from the workflow architecture and shipped defaults, not scores from a matched render test. The benchmark clips in section 03 are where measured results come from once each row has been reviewed.

Wan Animate 2 is the cheapest pass by a clear margin — six LCM steps at 480p. SCAIL-2 matches it on sampling but adds a segmentation stage the others don't have, plus the longest cold start from loading three models and two LoRAs. MiniMax H3 is the expensive one on every axis: more steps, higher resolution, more frames, and a second VAE. The Turbo LoRA is the main lever if you want H3's output without H3's bill.

Pose in, character out. The lowest-friction route from a photo to a moving character on Floyo.

Wan Animate 2 released on August 7, 2026 from Alibaba Tongyi Lab as part of the Wan 2.2 video series. It is a 14B model that takes a single character image and a driving video with no skeleton extraction step in between — the pose is read directly from the footage. A distilled lightx2v LoRA ships with the weights, which is what brings sampling down to six LCM steps at cfg 1.

On Floyo, the workflow runs the generation subgraph twice. The first pass hands its last frames and a video frame offset to the second through continue_motion, so the clip extends past the 81-frame single-pass default without an obvious seam. Context windows are set to 21 with 8 overlap and pyramid fusion, the cache runs int8 on GPU, and pose_strength sits at 1.0 with a full 0–100% window. Audio is lifted straight off the driving video and muxed back in, and both the character clip and a side-by-side stitch are saved.

The only workflow here that knows who it is replacing. Segmentation first, generation second.

SCAIL-2 is a Wan 2.1 14B fine-tune that takes explicit masks as a conditioning input. The Floyo workflow runs SAM 3.1 twice — once over every frame of the driving video, once over your reference image — then feeds both tracks into SCAIL2ColoredMask, which produces the colored mask pair the video model consumes. Two LoRAs stack on top: a DPO LoRA at full strength and the lightx2v distill at 0.8, which is what keeps it to six euler steps.

One boolean changes the whole job. With replacement_mode on, your character steps into the original footage and the background comes through untouched. With it off, the same graph does plain motion transfer. The workflow runs at 512 × 896 with pose_strength 0.75 and a 5-frame carry between passes; the note node lists 512 × 896, 704 × 1280, 896 × 512 and 1280 × 704 as the tested resolutions, and anything else floors to a multiple of 32. Both mask passes preview on canvas, so you can confirm it tracked the right subject before committing.

Omni-modal reference-to-video. The only one that builds a scene and scores it at the same time.

MiniMax H3, also known as Hailuo 3.0, released on July 31, 2026 as an omni-modal video model: it generates picture and synchronised audio in one pass from text, images, or reference video. Resolution goes to 2K at 24 fps with clips from 4 to 15 seconds. The Floyo workflow uses the reference-to-video pipeline, driven by a Qwen3-VL 32B text encoder and decoding through two VAEs — one for video, one for audio.

The prompt is doing real work here, not decorating. It uses tagged references — <Subject 1> for the character, <Picture 1> for the image, <Video 1> for the motion — which is how the model binds appearance to movement while ignoring the person in the driving clip. Defaults are 1280 × 720 at 20 res_multistep steps with a beta scheduler and no negative conditioning. The driving video is forced to 24 fps and capped at 360 frames, and clip length snaps to a 17-frame latent boundary, so output can run slightly past the input. A Turbo LoRA ships with the workflow but is off by default; switching it on drops 20 steps to 4.