wizard / examples /README.md
yuzhewu207's picture
Add example cases with outputs
7facd34 verified
|
Raw History Blame Contribute Delete
2.11 kB

Example cases

Three ready-to-run cases with our outputs. Each folder is a case as the inference code expects it:

file content
first_frame.png the first frame
cond_v7.mkv proxy video (R height, B depth, G geometry tier / character facing, skeletons on top)
cond.mp4 the same proxy in the earlier colour coding; only the character layer is read from it
tier7.npy, gbuf.npz geometry tier per cell; depth and cameras
slots.npz image region of each character per frame
ref_a.png, ref_b.png reference images of the two entering characters (slots 2 and 3)
prompt.txt text prompt (Chinese, as in training)
output.mp4 our result (bidirectional model, 30 steps, guidance 3, seed 0)
case what happens weights used for output.mp4
fight_character_strikes_player/ two characters enter from the sides; at 1.6 s character 1 (purple panther) strikes the player, who is knocked down and stays down weights/bidirectional_v7co2/dit_step35000.safetensors
fight_player_attacks_character/ two characters enter; at 1.6 s the player attacks character 2 (clockwork titan), which flinches from the hit weights/bidirectional_v7co2/dit_step35000.safetensors
entry_two_characters/ two characters with the same skeleton and different reference images enter one after the other; only the references tell them apart weights/bidirectional/dit_step35000.safetensors

Run one (from code/, see code/README.md for the setup):

C=../examples/fight_character_strikes_player
INF_UNK_NOCOORD=1 python -m wizard.inference.bidirectional \
    --dit ../weights/bidirectional_v7co2/dit_step35000.safetensors --case $C --frames 121 \
    --ref1 $C/ref_a.png --ref2 $C/ref_b.png --ref-slots 2,3 --seed 0 --out out.mp4

This writes out.mp4 (result next to the proxy) and out_gen.mp4 (result only); out_gen.mp4 should match output.mp4. One case takes about 4 minutes on one H200.

The first frames come from the WBench benchmark; the reference images were generated with Qwen-Image.