fix: resolve RAM OOM during video segmentation with batch processing

[farazshaikh]

Problem

Video segmentation with the SAM3 video model causes RAM Out-Of-Memory errors on longer videos (800+ frames) because:

• init_state() loads ALL video frames into RAM at once via load_resource_as_video_frames
• For 800 frames at 1024x1024, this pre-allocates ~5GB+ of RAM in a single tensor
• Memory grows unbounded as frame count increases

Solution: Mega-Batch Processing

Instead of loading all frames at once, process the video in mega-batches of 200 frames:

Total: 800 frames
Mega-batch 1: frames 0-199   -> init_state loads 200 frames -> process -> cleanup
Mega-batch 2: frames 200-399 -> init_state loads 200 frames -> process -> cleanup
Mega-batch 3: frames 400-599 -> init_state loads 200 frames -> process -> cleanup
Mega-batch 4: frames 600-799 -> init_state loads 200 frames -> process -> cleanup

Key Features

1. Bounded RAM usage: Peak RAM reduced from O(total_frames) to O(200 frames) = ~1.2GB instead of ~5GB
2. Tracking continuity: First batch uses text prompt, subsequent batches use box prompts from the last frame to re-seed tracking
3. Per-batch cleanup: Inference state reset + temp directory deleted + gc.collect() after each mega-batch

Memory Profile

Frames	Before (RAM)	After (RAM)	Reduction
200	~1.2GB	~1.2GB	1x
400	~2.4GB	~1.2GB	2x
800	~5GB	~1.2GB	4x
1600	~10GB (OOM)	~1.2GB	8x

Technical Details

• MEGA_BATCH_SIZE = 200: Number of frames per SAM3 initialization
• FRAME_BATCH_SIZE = 32: Batch size for processing outputs (unchanged)
• Frame indices are remapped: global -> local for SAM3, local -> global for results
• last_frame_boxes/last_frame_labels passed between mega-batches for tracking

Trade-offs

• Object IDs may change between mega-batches since each batch is a new SAM3 session
• Box-prompt seeding maintains spatial tracking but not ID continuity
• For strict ID continuity, users should use shorter video segments

Testing

Tested with 800-frame video tracking 3 objects:

• Before: OOM crash during init_state
• After: Completes successfully with stable ~1.2GB RAM usage

Changes

• src/comfyui_sam3/nodes.py: Refactored _segment_with_video_model method (229 insertions, 229 deletions)