Bobs Latent Optimizer

If you've ever typed "1344" into a width box, cross-checked it against a trained aspect-ratio table, then manually divided by 8 to feed a tiled upscaler - this node is that math, packaged. BobsLatentNode (the standard variant of Bobs_Latent_Optimizer) takes an aspect ratio and a megapixel target, hands you a correctly-shaped empty latent, and also computes the tile dimensions you'd otherwise guess for a downstream tiled upscaler. It doesn't call any API, needs no key, and downloads no models. It's just geometry, done right.

The name slightly oversells it: it doesn't "optimize" a latent in the neural sense. It allocates a fresh empty latent and picks dimensions that are legal for the model you're about to run. That's the part people burn time on, because the rules differ per family - FLUX wants multiples of 64, SDXL has a set of trained ratios around 1024×1024, and video models carry extra constraints. This node bakes those rules in.

How it works

You give it an aspect_ratio (it accepts 1:1, 16:9, 3:2, or decimal forms like 1.5:1) and an mp_size dropdown - 10 presets from 0.25 up to 4, where 1 is a 1024×1024 area and 4 is 2048×2048. It computes the pixel width and height that cover roughly that area at that ratio, then rounds them to the alignment step for your model_type. From there it allocates a zeroed latent tensor with the right channel count for that family - 16 channels for FLUX, 4 for SDXL-style models - and the right VAE downscale, so the pixel dimensions you set actually describe the tensor you get.

The trickier half is the tiling math. upscale_by is not an upscale - the node never touches your image. It's used to estimate the final pixel output size (base × upscale_by), then the node picks tile dimensions for the tiled upscaler you'll run after VAE decode. It aims for a 2×2 grid of tiles, and only adds tiles (3×2, 2×3, 3×3…) when a single tile would exceed the 2048×2048 limit. Tile sizes are rounded up to a multiple of 8, which the tiled upscaler nodes expect.

The inputs that matter

• aspect_ratio - the shape of the base image. Common values like 16:9 work as-is.
• mp_size - the "how big" dial. 1 (1024×1024) is the default and the one you'll reach for most; jump to 1.5–2 for higher-res first passes.
• upscale_by - feed it the scale factor you plan to use downstream (default 2). This drives the tile math only.
• model_type - pick your family: FLUX, SDXL, SD3, QWEN, or WAN. FLUX is default.
• batch_size - how many latents to stack, if you're generating a batch.

Outputs

Four outputs: latent (the empty latent, wire it straight into your KSampler), tile_width and tile_height (feed these into Ultimate SD Upscale, a tiled VAE decode, or similar after your first pass), and upscale_by passed through so a scale-expecting upscaler can grab it. The latent goes forward into sampling; the tile outputs sit in your back pocket for the upscale stage.

Installing it

Easiest via ComfyUI Manager - search for Bobs_Latent_Optimizer and hit install. Or the manual route:

cd ComfyUI/custom_nodes/
git clone https://github.com/BobsBlazed/Bobs_Latent_Optimizer.git

Restart ComfyUI. That's the whole install: this pack has zero Python dependencies and no model files - torch comes from ComfyUI itself, so there's nothing to fight over. Both nodes land in the latent/generate category.

Where people get burned

• Thinking upscale_by actually upscales. It won't - nothing comes out bigger than the base resolution you asked for. It only computes tiles for the upscaler you add later.
• Pushing mp_size too high. A 4 MP (2048×2048) FLUX latent is a big first pass on its own; the tile sizes only help the upscale stage, not the base generation. The node will warn in the log if a target can't be honored without distorting your aspect ratio.
• Not restarting after install. It's the classic. Clone, restart, then look under latent/generate.

The built-in EmptyLatentImage handles this job for one model at a time. If you hop between FLUX, SDXL, and WAN in one workflow folder, this node removes a whole category of "why is my latent 16 channels" confusion - and it's free to try.