Wan2.2 TI2V-5B on a single RTX 4090 at stock settings: ~8.5 min per 5 s clip, and faces broke in 2 of 3 (Qwen-Image-2.1 stills)
retained text
Wanted to see what the small 5B Wan2.2 does on a single 24 GB card, so I animated three Qwen-Image-2.1 stills with it on an RTX 4090 (and on a DGX Spark, out of curiosity) and wrote down what worked and what broke. The video is the three clips back to back: jellyfish, train, flooded city. ■ **The pipeline** • Stills: Qwen-Image-2.1 (nf4, 40 steps, CFG off) on the 4090. Each one started as a first version from a prompt (stills and prompts are in the comments) and then went through editing in Qwen-Image-2.1: about six passes for the jellyfish, one or two for the city and train, because repeated edits kept adding noise. So the prompts won’t reproduce the exact stills. The city and train scenes used an earlier still as a reference to keep the girl and the art style consistent. • Motion: Wan2.2-TI2V-5B, 50 steps, guidance scale 5.0, 121 frames at 24fps (5.04 s). Each prompt describes the motion and ends with “static camera”. ■ **Jellyfish** (518 s on the 4090) Motion: “A girl in a school uniform stands still before a giant aquarium window, looking up as a huge glowing jellyfish drifts slowly with its long tentacles trailing, small fish swim past, sunlight ripples down through the water, static camera.” ■ **Train and turtle** (518 s) Motion: “A girl sits on a train bench gazing out the window as a giant turtle glides past over a ruined, overgrown city, its eye blinking slowly, warm sunbeams shift across the carriage as the train sways gently, static camera.” ■ **Flooded city** (514 s) Motion: “A girl stands in shallow clear water on a flooded crosswalk between towering scaffolded buildings, looking up at a golden sky, clouds drift slowly, sheets of paper flutter in the wind, gentle ripples spread across the water, static camera.” ■ **Numbers (Wan step only)** • Output: 5.04 s, 24fps, about 720p, aspect ratio follows the input picture • RTX 4090 (24 GB): 518 s for the jellyfish clip, 24,008 MiB peak VRAM. It uses nearly the whole card • DGX Spark: 1,338–1,420 s per clip, 30,135 MiB median peak • Needed PYTORCH\_CUDA\_ALLOC\_CONF=expandable\_segments:True on the 4090; without it the VAE decode ran out of memory ■ **What held up, what didn’t** • In all three, the clip started from the picture and moved the way the prompt asked: the jellyfish drifted, fish swam, water rippled, the turtle blinked, and the camera stayed fixed. • In two of the three, the girl’s face came out distorted. The jellyfish clip had no defect, but there her back is to the camera, so it doesn’t prove faces are fine. In the train clip the light also grew too strong. ■ **Caveats** • Stock settings: 50 steps, no Lightning/distill LoRA, no TeaCache, no extra quantization. I didn’t try any of the speed-ups. • I ran it with model-compose, not ComfyUI (disclosure: I work with the model-compose team). The whole setup is one YAML file, and model-compose up builds the venv, downloads the weights and gives you an HTTP API plus a Gradio page. Config’s in the repo. • Small sample: 3 stills, 2 runs each with different seeds, and each clip here is the one picked from its pair. Don’t read too much into it. • Didn’t compare against other video models (Wan 14B, LTX, MiniMax H3). Setup, frames and full numbers: [MindrLabs/image-to-video-wan on Github](https://github.com/MindrLabs/image-to-video-wan) Anyone gotten TI2V-5B to keep faces intact? Was it the prompt, the resolution, or just switching to 14B?
document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xholms/wan22_ti2v5b_on_a_single_rtx_4090_at_stock