All models

CogVideoX / VIDEO

CogVideoX 5B VRAM requirements

Explore memory tradeoffs with sourced model data and transparent estimates.

FP8 estimate

15.6 GiB

at default settings

FP16 estimate

15.6 GiB

at default settings

Suggested GPU

1 × NVIDIA T4 16 GiB

Model files & facts

Mediavideo
Default resolution720 × 480
Default frames49
Temporal compression4
NotesWeights are BF16 (same size as FP16). The model card reports ~5 GB with diffusers sequential CPU offload, which trades speed for memory.

Precision & offloading comparison

Source default resolution and frame count · batch 1 · FP16 text encoders · safety margin included.

BackboneText encodersTotal
FP16GPU15.59 GiB
FP16CPU15.59 GiB

ComfyUI file placement

Example component locations; install matching loader nodes. Brace patterns in source filenames denote all shards. Preserve the source layout when a loader requires it.

models/diffusion_models/transformer/diffusion_pytorch_model-0000{1,2}-of-00002.safetensors
models/text_encoders/text_encoder/model-0000{1,2}-of-00002.safetensors
models/vae/diffusion_pytorch_model.safetensors

Recommended GPUs by memory fit

FP16 backbone at source default resolution, FP16 text encoders on GPU. Smallest nominal memory first; green results only. Check backend and precision support separately.

Sources & verification

Hugging Face model card

Exact file sources are linked in the component table above. ComfyUI memory planner constant

Last verified: 2026-11-18

Try your configuration

Your workload

Custom diffusion model parameters

Uses the ComfyUI activation and decode profile for CogVideoX 5B. Leave fields empty to use catalog files. Placeholder counts approximate half-precision file sizes.

Your hardware

Output dimensions

Images or videos per batch. Serving concurrency and throughput require separate measurements.

Resolution and peak VRAM
Sampling and timing

One step is one scheduler iteration. Enter timings from the same device, batch, resolution and guidance settings. Fixed overhead includes encoding, decoding and other work outside sampling.

Advanced
Decode at full size
Off — text encoders stay on the GPU
Off — ComfyUI swaps models between stages; the peak stage counts

Performance and memory results

64.9%Memory use

Memory estimate

15.6 GiB needed

Fits8.4 GiB to spare24.0 GiB available

The estimated workload fits with the safety margin included.

CogVideoX 5B

RTX 4090 24 GiB

Model specifications
Memory breakdown · sampling stage (peak)GiB

Vertical line = usable hardware memory

  • Diffusion model (FP16)10.38 GiB
  • Sampling activations2.74 GiB
  • ComfyUI inference reserve0.80 GiB
  • Safety margin1.67 GiB

Per-stage peak

ComfyUI moves idle models to system RAM between stages; the largest stage sets the requirement. Totals include the safety margin.

  • Text encode10.83 GiB

    Text encoders (FP16) 8.87 + ComfyUI inference reserve 0.80

    Subtotal 9.67 + safety margin (12%) 1.16 ≈ 10.83 GiB

  • SamplingPeak15.59 GiB

    Diffusion model (FP16) 10.38 + Sampling activations 2.74 + ComfyUI inference reserve 0.80

    Subtotal 13.92 + safety margin (12%) 1.67 ≈ 15.59 GiB

  • VAE decode11.89 GiB

    VAE 0.80 + VAE decode activations 9.01 + ComfyUI inference reserve 0.80

    Subtotal 10.62 + safety margin (12%) 1.27 ≈ 11.89 GiB

Generation performance

Denoising steps
50
Conditioning branches per step
1
Latent patch tokens per sample
17,550
Estimated backbone work per step
354.478 TFLOP
Step time at single-device BF16 peak
2,145.75 ms

DiT work estimates cover linear layers and visual self-attention. Enter measured step time and fixed overhead under Sampling and timing.

Compare GPU capacity

Frequently asked questions

How much VRAM does CogVideoX 5B need?

Our default estimate at the source default resolution, batch 1, FP16 text encoders on GPU and FP16 backbone is 15.6 GiB including the safety margin. These are estimates, not measured guarantees.

Can I offload text encoders?

Yes. The CPU option removes text encoder weights from the GPU estimate, trading GPU memory for system RAM and transfer time. Each component still needs to fit on its assigned device.

Does a green result guarantee compatibility?

No. Green means the estimated memory including safety headroom fits. Backend support, precision support, kernels and runtime allocation can differ. Confirm compatibility and measure your actual workflow.

Where do these estimates come from?

Architecture and file sizes are traced to model sources. The precision table identifies measured GGUF rates, block storage and nominal bit widths. The source data for CogVideoX 5B was last verified on 2026-11-18. Memory estimates use GiB.