FL Upscale Model

This is Fill-Nodes' version of the stock "Upscale Image (using Model)" node, with a few knobs bolted on that you actually end up wanting: batch processing, precision control, and an optional downscale after the model runs. You feed it a loaded upscale model and an image (or a stack of frames), and it runs the model over them.

First, set expectations, because "upscaler" hides three different jobs. FL_UpscaleModel does the first one: more pixels. It runs an ESRGAN-family model - the workhorse category - that adds resolution without inventing content. It cannot hallucinate detail into a soft image; that's a different lineage (SeedVR2, SUPIR, tile-ControlNet) and a different node. If your source is already sharp and you just need it bigger, this is exactly right and runs in a blink. If your source is blurry and you're hoping to recover detail, this node will just give you a bigger blurry image.

How it works

An ESRGAN model has a fixed native factor baked in - usually 4x. FL_UpscaleModel loads it, runs your image through, then optionally scales the result back down by downscale_by. That combination is the useful part: a 4x model plus downscale_by 0.5 lands you at an effective 2x, which is often what you actually want. The precision setting lets you trade a little quality for a lot less VRAM, and batching means you can push a whole sequence of video frames through in one pass.

The inputs that matter

• upscale_model - an UPSCALE_MODEL you load with a separate loader node (ComfyUI's built-in "Load Upscale Model"). Point that at a .pth in models/upscale_models - 4x-UltraSharp and RealESRGAN are the community defaults, with anime-specific variants like 4x-AnimeSharp for illustration.
• downscale_by (0.25–1.0) - scale the model's output back down to hit a target factor. Leave at 1 for the model's full 4x.
• precision - auto, 32, 16, or bfloat16. auto is fine most of the time; drop to 16 or bfloat16 if you're hitting out-of-memory.
• rescale_method - the interpolation used for the downscale step (lanczos, bicubic, area, bilinear, nearest-exact).
• batch_size (1–100) - how many images to process at once. Higher throughput on frame batches, more VRAM.

Output is a single IMAGE (or batch), ready to save or feed onward.

How to install it

Fill-Nodes is a large pack by filliptm; you install all of it to get any one node. ComfyUI Manager → search ComfyUI_Fill-Nodes → Install → restart, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes
pip install -r ComfyUI_Fill-Nodes/requirements.txt

The node itself needs no extra install, but it does need an actual upscale model on disk - grab one from OpenModelDB and drop the .pth in models/upscale_models.

Common issues & troubleshooting

"No upscale models found." The loader that feeds this node reads models/upscale_models. If it's empty, download a model there and refresh. This node consumes the model; it doesn't ship one.

The output is just bigger, not better. Expected. Pixel upscalers add resolution, not detail. If the source is soft, downscale it to roughly match its actual sharpness first and reach for a generative restorer instead - that's the standing advice for genuinely low-quality sources.

Out of memory on video. Big frame batches at full precision blow up fast. Drop batch_size, set precision to 16, and if you're upscaling video frame-by-frame, remember a per-frame image upscaler can make fine repeating textures shimmer across frames - a temporally-aware video upscaler is the correct tool there, not this one applied 24 times a second.