The rule
For each model I take its real files — the original 16-bit weights, the FP8 or INT8 file if one exists, and the GGUF files from Q8_0 down to Q2_K — with their exact sizes from the Hugging Face file listing (read 2026-09-25).
For a GPU with V GB of VRAM, a file fits when:
file size + working memory + 0.8 GB reserve ≤ V
Files are tried from the best quality down. The first one that fits entirely is the recommendation. On GPUs with FP8 hardware (RTX 40, RTX 50, RX 9000) the FP8 file is tried before Q8_0. Elsewhere Q8_0 comes first. INT8 is used when it is the only 8-bit file.
"Fits entirely" is the safe, fastest case on every setup. On NVIDIA with an up-to-date ComfyUI, a native FP8 or INT8 file that does not quite fit can still be a good choice because of Dynamic VRAM — verdict pages point this out where it applies.
The verdicts
- Runs well a 16-bit or 8-bit file fits entirely in VRAM.
- Runs a 6-, 5- or 4-bit GGUF fits entirely.
- Tight only a 3- or 2-bit GGUF fits.
- Offload only nothing fits entirely; ComfyUI can stream part of the model from system RAM. When the overflow is under 2 GB the page says so. How much speed streaming costs depends on the ComfyUI version, the file format and the PCIe link: on NVIDIA with Dynamic VRAM (default since March 2026) native FP8/INT8 files are designed to stream well; on older versions, AMD and Intel it costs more. These are the numbers I am measuring next.
- Not practical even the 4-bit file is more than twice the GPU's VRAM: it may start with heavy offloading, but it is not a realistic everyday setup.
Working memory per model
On top of the weights, sampling needs memory for activations, the latent image and the VAE. These allowances are my estimates for each model's standard setting: about one megapixel for image models, a short 480p–720p clip for video models. Bigger images, longer clips and batches need more.
What to download, and system RAM
Each answer page lists the full set of files a standard ComfyUI workflow loads: the model file, the text encoder, the VAE and anything else the official workflow needs (CLIP Vision for image-to-video, audio encoders, audio VAEs). Sizes are read from Hugging Face like the model files. Where a text encoder comes in several precisions, the 8-bit one is listed (FP4 only on RTX 50), with the others as alternatives.
System RAM is estimated as the model file(s) + text encoder + VAE, which ComfyUI keeps in RAM and moves to the GPU as needed, plus about 6 GB for Windows, ComfyUI and a browser, rounded up to a common RAM size. calculated
The calculator
The working-memory allowances below are for one reference setting per model. The calculator keeps 30 % of the allowance fixed and scales the rest with pixels × frames × batch. At the reference setting it reproduces every verdict on the site exactly (this is checked automatically for all 98 GPUs and 49 models); away from it, it is an estimate.
Macs
On an Apple Silicon Mac the “VRAM” is the part of unified memory macOS lets the GPU use by default (about 74 % of 16 GB up to about 84 % of 128 GB on current macOS, measured by apps that print Metal’s recommendedMaxWorkingSetSize). FP8 and INT8 files are skipped on Macs, because Apple GPUs cannot compute FP8 and ComfyUI converts such files back to 16-bit. Details on the Mac page.
What is not counted
- The text encoder. ComfyUI runs it first and can move it out of VRAM before sampling, so it does not have to fit at the same time. Each page gives separate advice for it.
- LoRAs and ControlNets. They add memory. A LoRA is usually small; a ControlNet can add gigabytes.
- Speed. The verdicts are about memory. Speed depends on the GPU's compute and bandwidth, its power limit (on laptops), and whether you offload.
The three labels
calculated the rule above. measured timed on my own RTX 5060 Ti 16 GB by a script, with date and settings. reported someone else's result, always with a link.
Checking the data
File sizes were read from the Hugging Face API for every repository. On 2026-09-25 all 530 original file sizes were fetched a second time, independently, and compared byte for byte: zero mismatches. GPU memory sizes come from NVIDIA, AMD and Intel spec pages (Notebookcheck for some older laptop GPUs), and every GPU page links its source.
Who collected this
I built this site myself, but I did not type in 839 file sizes and 98 spec sheets by hand. That reading was done with the help of AI, and every figure links back to where it came from, so you can check it. If you find a mistake, please tell me — I would rather the site be right than look right.