This workflow is free for a limited time! Creating the classic Matrix 360 orbital bullet-time effect used to require high-speed multi-camera rig. Not anymore. By chaining 3 MiniMax H3 videos inside a single ComfyUI workflow, you can start from one still photo, launch your subject into mid-air,…
Annons
Annons
This workflow is free for a limited time! Creating the classic Matrix 360 orbital bullet-time effect used to require high-speed multi-camera rig. Not anymore. By chaining 3 MiniMax H3 videos inside a single ComfyUI workflow, you can start from one still photo, launch your subject into mid-air, freeze time for a full 360-degree camera orbit, and resume the motion to land the shot. This workflow: Chain three MiniMax H3 videos from a single starting image in ComfyUI. Extract the final mid-air frame of Clip 1 and feed it into the MiniMax H3 360 Orbit LoRA to freeze time. Pass the ending frame of the 360-degree orbit into Clip 3 so that the motion continues naturally. Combine all three video segments and their stereo soundtracks automatically with the Concatenate Video node. Download All Assets video_minimax_h3_bullet_time_audio_guide.jsonDownload video_minimax_h3_bullet_time_samurai_audio_guide.jsonDownload video_minimax_h3_bullet_time_city_audio_guide.jsonDownload Software needed We will use ComfyUI, a free AI image and video generator that you can run on Windows, Mac, or Google Colab. Think Diffusion provides an online ComfyUI service if you do not have a capable local GPU. (SDA readers get 20% extra credit.) How the 3-Clip Bullet-Time Workflow Works Standard MiniMax H3 Image to Video is designed to move subjects forward through time. The 360 Orbit LoRA is trained to lock every object in place while the camera circles 360 degrees between two identical keyframes. The final video is created by: Clip 1 (Action Video): Takes your single input photo as first_frame with the Orbit LoRA turned off, animating the subject from the ground up to the peak of a jump. Clip 2 (360 Orbit Video): Grab the final mid-air frame of Clip 1, connects that frame to both first_frame and last_frame, and enables the 360 Orbit LoRA so that the camera rotates 360 degrees while the subject hangs motionless in the air. Clip 3 (Continue the Motion): Continue the motion by generating a video that uses the last frame of Clip 2 as the first_frame with Orbital LoRA off. Automatic Video Stitching: The built-in Concatenate Video node joins Clip 1, Clip 2, and Clip 3 together in order, preserving the stereo audio from every segment. MiniMax H3 3-Clip Workflow Step-by-Step Step 0: Update ComfyUI This workflow uses native ComfyUI video concatenation and subgraphs, which require a recent ComfyUI build. Update your installation first through ComfyUI Manager. Click the Manager button on the top toolbar. Select Update ComfyUI. Restart ComfyUI and refresh your browser tab. Step 1: Download the models Download these five model files and place them in their matching ComfyUI folders. If you already completed the MiniMax H3 360 Orbit LoRA tutorial, you already have every required file installed: Download minimax_h3_fl2va_pruned_int8_convrot.safetensors (19.5 GB) and put it in ComfyUI > models > diffusion_models. Download qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (14.6 GB) and put it in ComfyUI > models > text_encoders. Download minimax_h3_video_vae_int8_convrot.safetensors (2.62 GB) and put it in ComfyUI > models > vae. Download minimax_h3_audio_vae_fp32.safetensors (577 MB) and put it in ComfyUI > models > vae. Download minimax_h3_flf2v_lora_v1.safetensors (148 MB) and put it in ComfyUI > models > loras. Optional speed boost: If you want Clip 1 and Clip 3 to sample in 8 steps instead of 20 steps, also download minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors (1.8 GB) and place it in ComfyUI > models > loras. Even though the workflow contains three Image to Video (MiniMax H3) subgraph nodes, ComfyUI caches the diffusion model, text encoder, and VAE weights in memory. The 37 GB model stack loads from disk only once during the run. Step 2: Load the workflow Download the workflow JSON file below and drag it onto your ComfyUI canvas: Download The workflow is arranged from left to right across three generation columns (Clip 1: Generate Action Video, Clip 2: 360 Orbit Video, and Clip 3: Continue Motion), followed by Concatenate 3 Video Clips on the far right. Step 3: Install missing nodes If any nodes appear with red borders after loading the workflow, install the missing custom nodes: Click Manager > Install Missing Custom Nodes. Install any packages that are listed. Restart ComfyUI. Step 4: Upload the starting image Upload your single starting photograph to the Load Image node on the far left. You can save and use this sample skateboarder image to test the exact setup shown in this guide: The single input image loaded into the workflow. All four skateboard wheels start flat on the concrete. Leave the ResolutionSelector node set to 1:1 (Square) and 0.59 megapixels (768x768). Its width and height outputs connect to all three video subgraphs so that every clip renders at the exact same dimensions. Step 5: Configure Clip 1 (Generate Action Video) In the Clip 1: Generate Action Video (MiniMax H3) node, only first_frame is connected from Load Image, while last_frame is left disconnected so that the model can animate freely. Check these settings on Clip 1: duration: Set to 2.0 seconds (56 frames at 24 fps). Prompt: Describe the starting pose in , then instruct the subject to launch upward into slow-motion suspension at the end of the shot. The last frame of this video gen will be fed into Clip 2. Step 6: Configure Clip 2 (360 Orbit Video) Now we feed that mid-air frame into the Clip 2: 360 Orbit Video (MiniMax H3) node. Notice that the IMAGE output from Get Last Frame (Clip 1) connects to both first_frame and last_frame on Clip 2. Verify these settings on Clip 2: duration: Set to 3.0 seconds, matching the requirement of the 360 orbit LoRA. turbo_mode: Set to true to activate the internal 360 LoRA. lora_name: Select minimax_h3_flf2v_lora_v1.safetensors with turbo_model_strength at 1.0 and turbo_steps at 28. Prompt: Keep the frozen-time orbit prompt intact: One frozen instant. Only the camera moves. In a continuous 360 orbit. Preserve every person and object in exactly the same world position, orientation, shape and pose throughout the shot. Airborne objects remain suspended at the captured height and angle: no wobbling, shaking, spinning, drifting, falling or continued action. Keep faces, hands, clothing, liquids and the background motionless while retaining their natural appearance. Camera parallax is the only source of apparent movement. No cuts, zoom, morphing or added objects. Because both ends of Clip 2 are pinned to the mid-air frame from Clip 1, time stops completely while the camera orbits 360 degrees around the suspended skateboarder and returns to the exact starting angle on frame 73: Clip 2: The 360 Orbit LoRA rotates the camera all the way around the airborne pose and returns to the front view. Step 7: Configure Clip 3 (Continue the Motion) Between Clip 2 and Clip 3, a second Get Video Components and Get Last Frame (Clip 2) (ImageFromBatch with batch_index = -1, length = 1) pair extracts the 73rd frame of the orbit and wires it into first_frame on Clip 3: Continue Motion (MiniMax H3). Configure Clip 3 to finish the action: last_frame: Leave disconnected so that the subject is free to move again. duration: Set to 2.0 seconds (56 frames). turbo_mode: Set back to false (20 steps, no Orbit LoRA) so that normal physics and gravity resume. Prompt: Anchor to the mid-air pose, then describe the subject descending smoothly from the jump, landing all four wheels on the concrete deck, and rolling forward. Clip 3: Gravity resumes from the final frame of the orbit as the skateboarder lands back on the deck. Step 8: Concatenate and save the combined video On the right side of the graph, the Concatenate 3 Video Clips (ConcatenateVideo) node receives the VIDEO outputs from Clip 1 (video0), Clip 2 (video1), and Clip 3 (video2). Click Queue Prompt (or Run) to execute the full pipeline. On an RTX 4090 GPU, rendering all three clips takes about 4 minutes in total. Each stage saves its own preview video at the bottom of its column, while the final Save Video (Combined 3-Clip Video) node outputs the complete 7.7-second (185-frame) sequence with continuous stereo audio. Tips Keep Clip 1 short (2.0 seconds): Video models naturally apply gravity after a jump. If you set Clip 1’s duration to 5 seconds, your subject will jump and land before Clip 1 finishes, leaving them back on the ground by the final frame. A 2.0-second duration (56 frames) cuts the clip right at the mid-air peak. Pick a specific peak frame with batch_index: If your subject reaches the best mid-air pose slightly before the end of Clip 1, inspect Clip 1 and change batch_index on Get Last Frame (Clip 1) from -1 to the exact frame number (such as 42 or -10). You can also insert a Trim Video (Video Slice) node before Concatenate Video if you want to trim any extra frames after that point. Write slow-motion apex prompts for Clip 1: Including phrases like rising in smooth slow motion and remaining suspended high in mid-air at the apex of the jump at the end of the shot in Clip 1’s prompt helps the model hold the airborne pose across the final few frames. Use the 8-step Turbo LoRA for faster drafts: If you want to test prompts quickly, set turbo_mode to true, select minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors, and set turbo_steps to 8 on Clip 1 and Clip 3. Keep Clip 2 on minimax_h3_flf2v_lora_v1.safetensors at 24 to 28 steps so that the 360-degree camera orbit stays smooth. Keep dimensions locked across all clips: The Concatenate Video node requires every input clip to share the exact same pixel dimensions and frame rate. Always drive width and height on all three subgraphs from the same ResolutionSelector node (768x768). What’s next Turn Any Still Photo Into a 360 Orbit Infinite Loop Video MiniMax H3 Image to Video MiniMax H3 Multiframe Reference MiniMax H3 Fun ControlNet Union MiniMax H3 Text to Video ComfyUI The post 360 Orbit Action Video first appeared on Stable Diffusion Art.