Fit Image And Resize (FS)

This is the pack's img2img workhorse, and it's the node the README gets visibly more excited about: "Now this is where things get interesting." Feed it an image you already have in the graph plus a VAE, and it fits the image to a max size, resamples it, and VAE-encodes the result straight to a latent - one node instead of three or four. That's a real time-saver if you're building an img2img chain and don't want a Resize → VAEEncode pair cluttering the graph every time.

Where it fits in a workflow

The typical case: you generated an image, or loaded one from disk with the plain ComfyUI Load Image node, and now you want to run it back through a KSampler for a second pass - restyling, refining, or just varying it. Normally that's a resize node feeding a VAE Encode node feeding your sampler. This node collapses that into one box: image and VAE go in, a properly-sized latent (plus the resized image itself, and its dimensions) comes out, ready to plug into a KSampler with denoise set below 1.0.

The README is specific about that setting, and it's worth repeating because it's the single most useful tip in the whole pack: set denoise around 0.5 as your starting point. Lower keeps you close to the original image; higher lets the sampler deviate more; go too high and you'll get something only loosely related to what you fed in.

How it works

Same fitting math the whole pack shares - longest side scaled to max_size, aspect ratio preserved, result rounded to a multiple of 8 so the output lines up with SD's latent grid - followed by a resample using whichever filter you pick, then a VAE encode of the resized pixels into a latent batch.

The inputs and outputs that matter

• image (IMAGE) - the source image, already in your graph (this node doesn't load files itself - that's what "Load Image And Resize To Fit" is for).
• vae (VAE) - needs to match the checkpoint you'll sample with. A mismatched VAE won't error loudly; it'll just decode to garbage or washed-out color later, so keep the VAE tied to the same model family as your sampler.
• max_size (INT, default 768, step 8) - target size for the longer side.
• resampling - lanczos, nearest, bilinear, or bicubic. Lanczos is the safe general-purpose default; nearest is only useful for pixel art where you don't want any blending.
• upscale - false (default, shrink-only) or true (also scales up undersized images).
• batch_size (INT, default 1, max 64) - how many copies of the latent to produce. Bump this if you want several sampler variations from the same source image in one run.
• add_noise (FLOAT, default 0, 0–1) - adds noise directly to the image before encoding, which the README recommends as a way to push further iteration variety without cranking denoise so high you lose the source entirely.

Outputs: Latent (feed straight to your KSampler), Image (the resized pixels, not VAE-encoded - useful for previewing or compositing), Fit Width, Fit Height, and Aspect Ratio.

Installing it

ComfyUI Manager: search "Comfy Fit Size" or "fitsize." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/bronkula/comfyui-fitsize.git

Restart ComfyUI. No model downloads and nothing heavy to install - it's a small, self-contained pack. You'll find this node under Fitsize → Image.

Common issues

The VAE match matters more here than in most nodes, because this one does the encode for you silently - if you wire in a VAE from a different model family than the checkpoint you're about to sample with, you won't get an error, you'll get bad output (washed-out color or noise) further down the graph, which is a much harder thing to debug than a red node. Keep the VAE and checkpoint paired.

The other thing to know: VAE encode/decode is inherently lossy, so if this node sits in a loop you're running repeatedly (encode, sample, decode, feed back in, encode again), image quality erodes a little each pass - that's a property of the VAE round-trip itself, not a bug in this node. For a one-shot img2img pass it's a non-issue; for iterative refinement chains, it's worth knowing why quality creeps down over many cycles.

And the same denoise guidance from the README applies regardless of what's feeding this node: start around 0.5 in your KSampler and adjust from there rather than guessing at the extremes.