Graphics card memory
Why things do not fit, and what helps: lighter files, offloading, smaller jobs.
In plain words
The graphics card has its own memory (VRAM), usually 8 to 24 GB at home. The model's weights must be there while it works, together with the latent and the intermediate results of each step. Your computer's ordinary memory (RAM) is larger but slower to reach. When the work does not fit in VRAM, ComfyUI stops with an "out of memory" error (in its log: OutOfMemoryError or "Allocation on device").
Rough sizes: an SDXL checkpoint is about 7 GB in fp16; Flux.1 dev about 24 GB in fp16 and 12 GB in fp8; each Wan 2.2 14B model about 14 GB in fp8; the T5-XXL text encoder about 10 GB in fp16 and 5 GB in fp8.
What uses memory
- Model weights: the biggest part. Several models in one graph are loaded in turn when they do not all fit.
- Size of the job: width × height × frames × batch. Video at 720p and 81 frames needs far more than a picture.
- Decoding: VAE Decode of a large picture or a long clip briefly needs a lot.
What helps, in order
- Make the job smaller: lower size,
length,batch_size. Generate small, upscale later. - Lighter files: fp8 versions of models and text encoders (half of fp16, small quality loss);
weight_dtypefp8 on Load Diffusion Model; GGUF quantized models (Q8, Q6, Q5, Q4: smaller, slower, lower quality as the number drops) through the ComfyUI-GGUF pack and its GGUF loader nodes. - Offloading: ComfyUI moves models between VRAM and RAM by itself. Launch options change how hard it tries:
--lowvram(split the model, keep parts in RAM),--novram(keep almost nothing in VRAM),--reserve-vram(leave room for other programs). WanVideoWrapper's BlockSwap keeps a number of the model's blocks in RAM. All of these are slower. - Tiled VAE: VAE Decode (Tiled) for large pictures and long clips.
- Smaller models: Wan 2.2 5B instead of 14B; Z-Image Turbo or SDXL instead of Flux on small cards.
- Free memory between runs: ComfyUI's unload models / free cache commands; close other programs that use the card.
If even RAM runs out (the computer swaps or ComfyUI is killed), the model files are too large for the machine: use fp8 or GGUF files.
Typical mistakes
- fp16 files on an 8 to 12 GB card when fp8 versions exist.
- A first test at full size and length.
- Two copies of the same large model loaded under different nodes.
How the app helps
The app does not measure your card. Lessons name fp8 files and tell you to lower size or length when memory runs out. The lessons When the graph does not fit the graphics card, When a video does not fit in memory and A quantised model: GGUF are about memory; the WanVideoWrapper lesson explains BlockSwap and fp8 quantization.