Image Upscale With Model Batched

If you've ever tried to run a whole batch of frames - or a big pile of images - through the native UpscaleImageWithModel node and watched ComfyUI fall over with an out-of-memory error, this is the fix. Same ESRGAN-style model upscaling, but it processes the batch in sub-batches so peak VRAM stays sane, with an optional downscale on the way out. It's a small, practical quality-of-life swap that turns "can't do it" into "done."

What it does differently

Model-based upscalers (the 4x-UltraSharp, RealESRGAN, SwinIR family of .pth files you load with an upscale-model loader) run a neural net over the image to add detail as they enlarge. That's memory-hungry, and the native node tries to process the entire batch at once. Feed it 60 video frames and it wants to hold all 60 upscaled results - and their intermediate tensors - in VRAM simultaneously. Boom.

This node's node-description says it plainly: "Same as ComfyUI native model upscaling node, but allows setting sub-batches for reduced VRAM usage. Optionally downscale the result with a ratio." So it chunks the batch, upscales each chunk, frees the memory, moves on. Slower wall-clock, but it actually finishes on a card that couldn't otherwise. For upscaling video frame sequences this is the one you want.

The inputs that matter

• upscale_model - the loaded upscale model (from an upscale-model loader), same as the native node.
• images - the batch to upscale.
• per_batch (default 16) - the key knob. How many images to process at a time. Lower it if you're still OOMing; raise it for a bit more speed if you've got headroom. This is the dial that trades memory for throughput.
• downscale_ratio (optional, default 1) - most upscale models are fixed 4x. If you wanted, say, 2x, you upscale 4x then set this below 1 to pull the result back down. A ratio of 1 means no downscale.
• downscale_method (default lanczos) - the resampling filter for that downscale; lanczos is the sharp, sensible default.
• precision (default float32) - drop to float16/bfloat16 to shave VRAM further if float32 is tight, at a small quality risk.

The output is the upscaled IMAGE batch.

How to install it

It's in kijai's KJNodes pack.

• ComfyUI Manager - search KJNodes for ComfyUI, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt, restart.

The node has no model of its own - you supply the upscale .pth. Grab those from the OpenModelDB catalog or wherever you get your ESRGAN models and drop them in models/upscale_models.

Common issues & troubleshooting

Still running out of memory. Lower per_batch - 8, 4, even 1 if you have to. At per_batch 1 it upscales one image at a time, which is the safest and slowest. Also try dropping precision to float16. If a single image at the model's native scale still won't fit, the image itself is too large for tiled-free upscaling and you need a tiled upscaler instead.

It's slow. That's the trade - sub-batching serializes work that the native node did all at once (when it fit). If you have the VRAM headroom, raise per_batch to claw some speed back.

Video frames flicker after upscaling. Model upscalers work per-frame with no temporal awareness, so tiny per-frame differences show up as shimmer. That's inherent to frame-by-frame upscaling, not this node specifically - dedicated video upscalers handle temporal consistency, at their own cost.

The output isn't the size you expected. Remember the model dictates the scale (usually 4x). If you wanted a different final size, that's what downscale_ratio is for - upscale at the model's native factor, then ratio back down to your target.