How much VRAM does FLUX.1 need?
According to the Hugging Face Diffusers Flux docs, "Flux is a very large model and requires ~50GB of RAM/VRAM to load all the modeling components." That covers the transformer, both text encoders (CLIP and T5) and the VAE at full precision.
Nobody runs it like that on a gaming card. The same docs list these ways to cut memory:
| Method (Diffusers docs) | What the docs say |
|---|---|
| Sequential CPU offload + VAE slicing/tiling | Example commented "to run on low vram GPUs (i.e. between 4 and 32 GB VRAM)". Slow, because weights stream from RAM. |
| FP8 transformer + FP8 T5 (optimum-quanto) + model CPU offload | "run Flux with less than 16GB of VRAM" |
| 8-bit bitsandbytes transformer and T5 | Supported; the docs give no exact VRAM figure |
Group offloading (apply_group_offloading) |
Lowers VRAM by offloading groups of layers; no exact figure given |
We only list figures the docs actually state. Real use depends on resolution, steps and your Diffusers version.
ComfyUI
ComfyUI's README says it can "run even the biggest open source models on as low as 4GB vram + 8GB ram" using its asynchronous weight streaming. ComfyUI doesn't publish a per-model Flux VRAM table, so we don't give one. Expect low-VRAM setups to be slower than running the model fully on the GPU.
Practical take
- Under 16 GB: use an FP8 or 8-bit setup with CPU offload, as in the docs above.
- 4–8 GB: sequential offload works, but it's slow.
- Want it fast at full precision? Renting a 48 GB or 80 GB card by the hour is often cheaper than buying one. Compare rates on our GPU prices page (we may earn from some links, disclosure).
Related
Sources
- Hugging Face Diffusers, Flux pipeline docs, "Optimize" (checked 2026-11-01)
- ComfyUI README (checked 2026-11-01)