Tokenization guide

Documentation Index

Fetch the complete documentation index at: https://docs.perceptron.inc/llms.txt

Use this file to discover all available pages before exploring further.

Understand how Perceptron’s vision-language models count tokens for images and video so you can estimate costs and optimize preprocessing pipelines.

​

Isaac 0.1 and 0.2 Token Counting

Isaac 0.1, Isaac 0.2 1B, and Isaac 0.2 2B (Preview) all use the same patch-based image tokenizer:

• Native resolution: Processes images at their original resolution; supports a wide range of aspect ratios.
• Patch size: 16 × 16 pixels.
• Spatial merge size: 2 × 2 (4 patches into a single token).
• Token formula: ⌈width / 32⌉ × ⌈height / 32⌉.

​

Constraints

• Minimum: 256 patches → 64 tokens.
• Maximum: 6,144 patches → 1,536 tokens.
• Images outside these bounds are automatically resized while maintaining aspect ratio.
• Due to the resize algorithm (dimensions must be divisible by 32), the practical maximum is typically around 1,508 tokens for common aspect ratios.

​

Calculation Examples

Example 1: 640×480 (VGA) — No Resize Needed

1. Round dimensions to nearest multiple of 32: 640×480 (already divisible).
2. Calculate patches: (640 ÷ 16) × (480 ÷ 16) = 40 × 30 = 1,200 patches.
3. Calculate tokens: 1,200 patches ÷ 4 = 300 tokens.
4. Check constraints: 256 ≤ 1,200 ≤ 6,144 ✓ (no resize needed).
5. Cost (Isaac 0.x at $0.15/M input): 300 × ($0.15 / 1,000,000) = $0.000045.

Example 2: 1920×1080 (Full HD) — Requires Resize

1. Calculate original patches: (1920 ÷ 16) × (1080 ÷ 16) = 120 × 68 = 8,160 patches.
2. Check constraints: 8,160 > 6,144 (exceeds maximum, resize needed).
3. Resize to 1664×928 (maintains ~16:9 aspect ratio, divisible by 32).
4. Calculate new patches: (1664 ÷ 16) × (928 ÷ 16) = 104 × 58 = 6,032 patches.
5. Calculate tokens: 6,032 patches ÷ 4 = 1,508 tokens.
6. Cost (Isaac 0.x at $0.15/M input): 1,508 × ($0.15 / 1,000,000) = $0.000226.

​

Perceptron Mk1 Token Counting

​

Image tokenization

• Native resolution: Processes images at their original resolution; supports a wide range of aspect ratios.
• Patch size: 16 × 16 pixels.
• Spatial merge size: 2 × 2 (4 patches into a single token).
• Token formula: ⌈width / 32⌉ × ⌈height / 32⌉.

​

Video tokenization

Dynamic resolution and frame rate: samples video at target_fps = 2, with the sampled frame count clamped to min_frames = 2 and max_frames = 256. Frames are smart-resized at the original aspect ratio so each frame fits within the total video patch budget max_patches_per_video = 131072.

• Patch size: 16 × 16 pixels.
• Spatial merge size: 2 × 2 = 4 spatial patches per token.
• Temporal patch size: 2 frames.
• Effective token cell: 32 × 32 pixels across 2 frames.
• Sampled frames: clamp(duration_seconds × 2, 2, 256), rounded to a multiple of 2.
• Token formula: ceil(sampled_frames / 2) × ceil(width / 32) × ceil(height / 32).
• Max video tokens: 131072 is the per-video patch budget; the effective cap is about 16K video tokens per video.

​

Constraints

• Context window: 32K tokens (image + video + text + reasoning + answer all share the same budget).
• Output tokens: 8K maximum.
• Supported MIME types: image/png, image/jpeg, image/webp, video/mp4, video/webm.

​

Common Image Sizes (Isaac 0.x)

Token counts and costs for common image resolutions on the 0.x family. Pricing: $0.15 per million input tokens.

​

Optimization Guidance

​

Recommended Resolutions

We recommend passing in the original resolution of the image. If the resolution is greater than the maximum supported, we recommend client-side preprocessing. Lower resolution can erode quality but may improve latency and reduce token counts.

​

Client-Side Preprocessing

You can resize images before sending them to reduce token usage and costs: When to Resize:

• Below minimum (Isaac 0.x): If your images are smaller than 256 patches for Isaac 0.x, resize them yourself to avoid automatic upscaling.
• Above maximum (Isaac 0.x): If your images exceed 6,144 patches for Isaac 0.x, resize them yourself to maintain control over quality.

Recommendations:

1. Resize to multiples of 32: When resizing, aim for dimensions divisible by 32 (e.g., 1280×720, 1024×1024, 1920×1088) to avoid additional processing overhead.
2. Maintain aspect ratio: Preserve original proportions to avoid distortion.
3. Faster uploads: Pre-resized images reduce bandwidth usage.

​

Video preprocessing (Perceptron Mk1)

Practical implications of the Video tokenization spec above:

• Sampler caps clip length to ~128 seconds: at target_fps = 2 and max_frames = 256, anything longer is truncated.
• Frames are smart-resized at the original aspect ratio so the total patch count across all sampled frames fits within max_patches_per_video = 131072. The more frames sampled, the lower each frame’s effective resolution.