How it reads a workflow
A workflow names every model file it loads in its loader nodes. The checker reads those names from the .json, or from the copy ComfyUI writes into every PNG it saves, and then:
- Sizes each file exactly. For the 101 files of the image and video models this site covers, and the 375 files ComfyUI's 383 official templates download, it has the byte counts Hugging Face reports. A file it does not know gets a box for the size from your own disk, and is left out of the answer until you fill it in.
- Works out the stages. Text encoders first, then any prompt enhancer, then the diffusion model with its VAE, LoRAs and ControlNets. Two-expert models such as Wan 2.2 hold one expert at a time, with that expert's LoRA. Depth, segmentation and background-removal models run in their own step. The largest stage is what has to fit: under 85% of the card's usable memory loads fully, up to 95% loads tightly.
- Finds a swap. If the workflow does not load on your card, it looks for the highest-precision files of the same model that would: the diffusion model first, then the text encoder, never another model's file.
- Lists custom nodes. Every node type not built into ComfyUI 0.41.0, with its node pack where the workflow records one.
Or start from a model: pick one, and the checker loads the model files of ComfyUI's own template for it and answers the same way. Beside the memory answer, it says whether Comfy Cloud runs the workflow, the cheapest cloud GPU by the hour that holds it, and what one image or clip costs, where a provider prints a price or we measured one ourselves. Each figure carries the date it was read or measured.
Working memory is not estimated, here or anywhere on this site: it grows with resolution, frame count and batch, depends on the attention kernel, and nobody publishes it per model.
When ComfyUI runs out of memory
- The files load, then generation fails: lower the resolution or the frame count. Working memory is what grew.
- The largest stage does not fit the card: swap to a smaller file of the same model, FP8 or INT8 instead of BF16. The checker names the file and links to it.
- Nothing smaller fits: ComfyUI streams what does not fit from system RAM. Its README says it "can run even the biggest open source models on as low as 4GB vram + 8GB ram". Slower, not impossible.
Where each file goes
A "missing model" message usually means a file is in the wrong folder. Each kind of model has its own, under ComfyUI/models/:
| Kind of file | Folder |
|---|---|
| Diffusion model | models/diffusion_models |
| Checkpoint, all in one | models/checkpoints |
| Text encoder or prompt enhancer | models/text_encoders |
| VAE | models/vae |
| LoRA | models/loras |
| Vision encoder | models/clip_vision |
| Upscaler | models/upscale_models |
| ControlNet | models/controlnet |
| Model patch | models/model_patches |
For a model's every file at every precision, and which one loads on which card, see the image and video models or the ComfyUI VRAM checker. If your card falls short, the ComfyUI cloud guide compares every way to rent one, by price.