Implementation PR: #5752
?? The feature, motivation and pitch
vLLM-Omni MiniMax-H3 ????????
???????HF ??? / MiniMax API ?? / SGLang ???? minimax_h3_spec_research.md?? vllm-omni ?????feature/minimax-h3 ????????
?????pipeline_minimax_h3.py?reference_video.py?serving_video.py?api_server.py
?????????????
| ?? |
?? |
| t2va ????? |
? |
| fl2va ??????????? 0 ?? |
? |
| ref2va ?? + ????? |
? |
| ref2va ???????????????????????? |
?????? + multipart input_references ??? |
?????duration / num_frames?17n+5 ?????????flow_shift / audio_flow_shift |
? |
????? / ????????????????
FL2VA ??
Ref2VA ????????? ?9??? ?3??? ?3??? ?12???????/???
??
????????????????
| ?? |
?? |
vllm-omni ?? |
vllm-omni ?? API |
| ???t2va? |
? |
? |
? |
| ?? + ???fl2va? |
? |
? |
? |
| ?? + ???fl2va? |
? |
? |
? |
| ?? + ?? + ???fl2va? |
? |
? |
? |
| ?? + ???9?ref2va? |
? |
?? ? 1 ??????? |
?? ?? |
| ?? + ????3?ref2va/V2V? |
? |
?????????? |
?? ? multipart ?????URL ????????? |
| ?? + ? + ?? |
? |
? |
? |
| ?? + ? + ????3? |
? |
?? ??? 1 ? |
?? ?? |
| ?? + ?? + ?? |
? |
? |
? |
| ?? + ? + ?? + ?? |
? |
? |
? |
Implementation PR: #5752
?? The feature, motivation and pitch
vLLM-Omni MiniMax-H3 ????????
?????????????
input_references???duration/num_frames?17n+5 ?????????flow_shift / audio_flow_shift????? / ????????????????
FL2VA ??
frame_index: -1?diffuse()???keyframe_frame_indices=[0]????frame_index: [0, -1]?_load_image??????? "MiniMax H3 currently supports exactly one image"Ref2VA ????????? ?9??? ?3??? ?3??? ?12???????/???
_load_image???target_frame_count=num_frames?????????? N ?????? 2?15s/?????start_time_seconds?video_reference.video_url?????? PIL ?????????prepare_reference_videos??????? ? TypeError?multipartinput_references??????????
num_outputs_per_prompt?1?10?seed ?seed + output_index???forward?????????num_outputs_per_prompt?AudioX pipeline ?????H3 ???extra_args["duration"]???????????????????aspect_ratio/short_edge?????t2va ?????fl2va ??????ref2va ?? 16:9?width/height?t2va/ref2va ???? 768?1344????????????????