Tenos Resize to ~1 M Pixels

A zero‑nonsense ComfyUI node that resizes every incoming image (or batch) so the final resolution hovers around one million pixels—while respecting the original aspect ratio and snapping both dimensions to multiples of 64. Perfect for pipelines that need a predictable VRAM footprint without destroying composition.

---

Why you’d use it

• Stable VRAM budgeting – 1 Mpx keeps you under GPU‑meltdown territory for SDXL & friends.
• Clean, divisible dims – Multiples of 64 mean you never hit those “size must be divisible by 8/16/64” errors again.
• No potato quality – Pick from nearest, bilinear, bicubic, or area; the node auto‑anti‑aliases when down‑sampling with the smooth modes.
• Batch‑safe – Works on full tensors shaped (B, H, W, C); each frame is resized independently.
• Tenos‑grade math – Two rounding strategies duel it out under the hood; whichever lands closer to 1 Mpx wins. (It’s nerdy—see the code.)

---

Quick install

1. Drop it Save tenos_image_resize_target_pixels.py into:

   ComfyUI/custom_nodes/
   

   (Create the folder if it doesn’t exist.)

2. Restart ComfyUI – The node shows up under TenosNodes ➜ Image Processing ➜ Tenos Resize to ~1M Pixels.

3. Wire it in – Feed any image tensor; enjoy harmony.

---

Node spec

Field	Type / Options	Notes
Input – image	IMAGE	4‑D tensor (B,H,W,C)
Input – interpolation	One of area, bicubic, bilinear, nearest	Defaults to bicubic if you pass something weird
Output – image	IMAGE	Resized tensor, same batch size

Constants (tweak in code if you dare)

• TARGET_TOTAL_PIXELS = 1_000_000 – The sweet spot.
• DIMENSION_MULTIPLE = 64 – Keeps UNet / VAE happy.

---

How it works — the TL;DR math

1. Compute the ideal height and width for exactly 1 Mpx at the original aspect ratio.
2. Try rounding height first → recalc width → snap both to /64.
3. Try rounding width first → recalc height → snap.
4. Whichever candidate lands closer to 1 Mpx wins; ties go to option 1.

Edge‑cases handled:

• Zero / negative dims fall back to a 64 × 64 stub.
• If the image already matches the target dims, it passes through untouched (zero extra ops).
• Anti‑aliasing toggles itself only when you’re down‑scaling with bilinear or bicubic.

---

Example workflow (pseudo‑graph)

Loader → Tenos Resize to ~1M Pixels → Txt2Img → Save PNG

Need a weird upscale? Just chain it before a separate upscale node—this keeps your latent generation predictable.

---

Performance tips

• Area and nearest are cheapest but can look crunchy.
• Bicubic is slowest but usually prettiest—especially for photo content.
• On big batches, the node loops per‑image, so your VRAM stays flat but it scales CPU‑side.

---

Created with love from Tenos.ai