Check my GPU →

Stable Diffusion 3.5 Large VRAM requirements

Stability AI's 8B model. City96 only publishes the older GGUF types for it (Q8_0, Q5, Q4), not K-quants.

Released 2024-10Licence: Stability AI Community License (free <$1M revenue)Steps: 28–40Text encoder: CLIP-L + CLIP-G + T5-XXL (4.9 GB as FP8)
TypeImage
Parameters8.1B
Fits entirely from8 GB
8-bit or better from12 GB

01The files, and how much VRAM each needs

FileSizeNeededMin. VRAMQualitySource
16-bit16.5 GB18.8 GB19 GBthe original weightsstabilityai/stable-diffusion-3.5-large →
Q8_08.8 GB11.1 GB12 GBpractically identical to the originalcity96/stable-diffusion-3.5-large-gguf →
Q5_16.3 GB8.6 GB9 GBclose; small differences in fine detailcity96/stable-diffusion-3.5-large-gguf →
Q4_15.3 GB7.6 GB8 GBgood; some loss in fine detail and textcity96/stable-diffusion-3.5-large-gguf →

“Needed” = file + 1.5 GB working memory + 0.8 GB system reserve.

02By amount of VRAM

6GBOffload onlyQ4_18GBRunsQ4_110GBRunsQ5_112GBRuns wellQ8_016GBRuns wellQ8_020GBRuns well16-bit24GBRuns well16-bit32GBRuns well16-bit

03Best GPU for SD 3.5 Large

The cheapest cards (by launch price) that run it well, and every card sorted by memory: best GPU for SD 3.5 Large → Planning bigger images or longer clips? Open the calculator →

04By graphics card

GPUVRAMVerdictBest fileNeeded
Desktop graphics cards
RTX 2060 6 GB6 GBOffload onlyQ4_17.6 GB
RTX 3050 6 GB6 GBOffload onlyQ4_17.6 GB
RTX 2070 Super 8 GB8 GBRunsQ4_17.6 GB
RTX 2080 Super 8 GB8 GBRunsQ4_17.6 GB
RTX 3050 8 GB8 GBRunsQ4_17.6 GB
RTX 3060 8 GB8 GBRunsQ4_17.6 GB
RTX 3060 Ti 8 GB8 GBRunsQ4_17.6 GB
RTX 3070 8 GB8 GBRunsQ4_17.6 GB
RTX 3070 Ti 8 GB8 GBRunsQ4_17.6 GB
RTX 4060 8 GB8 GBRunsQ4_17.6 GB
RTX 4060 Ti 8 GB8 GBRunsQ4_17.6 GB
RTX 5050 8 GB8 GBRunsQ4_17.6 GB
RTX 5060 8 GB8 GBRunsQ4_17.6 GB
RTX 5060 Ti 8 GB8 GBRunsQ4_17.6 GB
RX 7600 8 GB8 GBRunsQ4_17.6 GB
RX 9050 8 GB8 GBRunsQ4_17.6 GB
RX 9060 XT 8 GB8 GBRunsQ4_17.6 GB
Arc B570 10 GB10 GBRunsQ5_18.6 GB
RTX 3080 10 GB10 GBRunsQ5_18.6 GB
RTX 2080 Ti 11 GB11 GBRunsQ5_18.6 GB
Arc B580 12 GB12 GBRuns wellQ8_011.1 GB
RTX 2060 12 GB12 GBRuns wellQ8_011.1 GB
RTX 3060 12 GB12 GBRuns wellQ8_011.1 GB
RTX 3080 12 GB12 GBRuns wellQ8_011.1 GB
RTX 3080 Ti 12 GB12 GBRuns wellQ8_011.1 GB
RTX 4070 12 GB12 GBRuns wellQ8_011.1 GB
RTX 4070 Super 12 GB12 GBRuns wellQ8_011.1 GB
RTX 4070 Ti 12 GB12 GBRuns wellQ8_011.1 GB
RTX 5070 12 GB12 GBRuns wellQ8_011.1 GB
RX 7700 XT 12 GB12 GBRuns wellQ8_011.1 GB
RX 9070 GRE 12 GB12 GBRuns wellQ8_011.1 GB
Arc A770 16 GB16 GBRuns wellQ8_011.1 GB
RTX 4060 Ti 16 GB16 GBRuns wellQ8_011.1 GB
RTX 4070 Ti Super 16 GB16 GBRuns wellQ8_011.1 GB
RTX 4080 16 GB16 GBRuns wellQ8_011.1 GB
RTX 4080 Super 16 GB16 GBRuns wellQ8_011.1 GB
RTX 5060 Ti 16 GB16 GBRuns wellQ8_011.1 GB
RTX 5070 Ti 16 GB16 GBRuns wellQ8_011.1 GB
RTX 5080 16 GB16 GBRuns wellQ8_011.1 GB
RX 7600 XT 16 GB16 GBRuns wellQ8_011.1 GB
RX 7800 XT 16 GB16 GBRuns wellQ8_011.1 GB
RX 7900 GRE 16 GB16 GBRuns wellQ8_011.1 GB
RX 9060 XT 16 GB16 GBRuns wellQ8_011.1 GB
RX 9070 16 GB16 GBRuns wellQ8_011.1 GB
RX 9070 XT 16 GB16 GBRuns wellQ8_011.1 GB
RX 7900 XT 20 GB20 GBRuns well16-bit18.8 GB
Arc Pro B60 24 GB24 GBRuns well16-bit18.8 GB
RTX 3090 24 GB24 GBRuns well16-bit18.8 GB
RTX 3090 Ti 24 GB24 GBRuns well16-bit18.8 GB
RTX 4090 24 GB24 GBRuns well16-bit18.8 GB
RX 7900 XTX 24 GB24 GBRuns well16-bit18.8 GB
Arc Pro B70 32 GB32 GBRuns well16-bit18.8 GB
RTX 5090 32 GB32 GBRuns well16-bit18.8 GB
Laptop GPUs
RTX 3050 Laptop 4 GB4 GBOffload onlyQ4_17.6 GB
RTX 3050 Ti Laptop 4 GB4 GBOffload onlyQ4_17.6 GB
RTX 2060 Laptop 6 GB6 GBOffload onlyQ4_17.6 GB
RTX 3050 Laptop 6 GB6 GBOffload onlyQ4_17.6 GB
RTX 3060 Laptop 6 GB6 GBOffload onlyQ4_17.6 GB
RTX 4050 Laptop 6 GB6 GBOffload onlyQ4_17.6 GB
RTX 2070 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 2070 Super Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 2080 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 2080 Super Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 3070 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 3070 Ti Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 3080 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 4060 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 4070 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 5050 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 5060 Laptop 8 GB8 GBRunsQ4_17.6 GB
RTX 5070 Laptop 8 GB8 GBRunsQ4_17.6 GB
RX 7600M 8 GB8 GBRunsQ4_17.6 GB
RX 7600M XT 8 GB8 GBRunsQ4_17.6 GB
RX 7600S 8 GB8 GBRunsQ4_17.6 GB
RX 7700S 8 GB8 GBRunsQ4_17.6 GB
RTX 4080 Laptop 12 GB12 GBRuns wellQ8_011.1 GB
RTX 5070 Laptop 12 GB12 GBRuns wellQ8_011.1 GB
RTX 5070 Ti Laptop 12 GB12 GBRuns wellQ8_011.1 GB
RX 7800M 12 GB12 GBRuns wellQ8_011.1 GB
RTX 3080 Laptop 16 GB16 GBRuns wellQ8_011.1 GB
RTX 3080 Ti Laptop 16 GB16 GBRuns wellQ8_011.1 GB
RTX 4090 Laptop 16 GB16 GBRuns wellQ8_011.1 GB
RTX 5080 Laptop 16 GB16 GBRuns wellQ8_011.1 GB
RX 7900M 16 GB16 GBRuns wellQ8_011.1 GB
RTX 5090 Laptop 24 GB24 GBRuns well16-bit18.8 GB
Unified memory
Radeon 8060S (Strix Halo) 96 GB96 GBRuns well16-bit18.8 GB
Radeon 8065S (Gorgon Halo) 160 GB160 GBRuns well16-bit18.8 GB
Apple Silicon Macs (by memory)
Mac 16 GB12.7 GBRuns wellQ8_011.1 GB
Mac 18 GB14.4 GBRuns wellQ8_011.1 GB
Mac 24 GB19.6 GBRuns well16-bit18.8 GB
Mac 32 GB26.8 GBRuns well16-bit18.8 GB
Mac 36 GB30.2 GBRuns well16-bit18.8 GB
Mac 48 GB40.2 GBRuns well16-bit18.8 GB
Mac 64 GB55.7 GBRuns well16-bit18.8 GB
Mac 96 GB85 GBRuns well16-bit18.8 GB
Mac 128 GB115.4 GBRuns well16-bit18.8 GB
Mac 192 GB175.4 GBRuns well16-bit18.8 GB
Mac 256 GB236.9 GBRuns well16-bit18.8 GB
Mac 512 GB498.1 GBRuns well16-bit18.8 GB

05Text encoder, VAE and other files

FileFolderSizeWhen
CLIP-L
clip_l.safetensors
models/text_encoders0.2 GBrequiredDownload →
CLIP-G
clip_g.safetensors
models/text_encoders1.4 GBrequiredDownload →
T5-XXL FP8
t5xxl_fp8_e4m3fn.safetensors
models/text_encoders4.9 GBtext encoder · smaller, recommendedDownload →
T5-XXL FP16
t5xxl_fp16.safetensors
models/text_encoders9.8 GBtext encoder · alternativeDownload →
SD 3.5 VAE (gated repo: log in and accept the licence first)
diffusion_pytorch_model.safetensors
models/vae0.2 GBrequiredDownload →

The files the official ComfyUI workflows load next to the model. Sizes read from Hugging Face (2026-10-01).

06Where the files go in ComfyUI

FileFolderLoader node
Diffusion model (.safetensors: 16-bit, FP8, INT8)ComfyUI/models/diffusion_modelsLoad Diffusion Model
GGUF file (.gguf)ComfyUI/models/unetUnet Loader (GGUF) - from the ComfyUI-GGUF node pack
Text encoderComfyUI/models/text_encodersLoad CLIP / DualCLIPLoader (or the GGUF versions)
VAEComfyUI/models/vaeLoad VAE

07Measured and reported results

LabelGPUSetupResultPeak VRAMDateSource
reportedArc B580 12 GBfp8 · 1024x1024 · 20 steps
“SD 3.5 large fp8 (20step): GOOD (26s)”
Intel Arc B580 12GB, PyTorch 2.8 XPU in Docker (YanWenKun); single 1024x1024 image, pre-warmed inference only (model load excluded)
26 s / image—2025-08-03github.com →

Other people's numbers are copied as published, with a link. Send yours.

08Training a LoRA for SD 3.5 Large

TrainerVRAMTypeSettings and quoteSource
SimpleTuner13.9 GBexample runint2-quanto base, rank-1 LoRA (config comment in SD3 guide, which targets SD 3.5 Large)
“int2-quanto is pretty extreme and gets the whole rank-1 LoRA down to about 13.9GB VRAM.”
github.com →
ai-toolkit24 GBexample runOfficial example config: rank 16, 1024px, batch 1, gradient checkpointing, 8-bit quantized base
“train_lora_sd35_large_24gb.yaml”
github.com →

reported From each trainer's own docs, 2026-09-25. Why the figures differ →

09Every file tracked for SD 3.5 Large

FileTypeSizeRepo
sd3.5_large_fp8_scaled.safetensorsFP814.93 GBComfy-Org/stable-diffusion-3.5-fp8 →
all-in-one checkpoint: includes text encoder(s) and/or VAE, not diffusion-model-only (clip_l, clip_g, T5-XXL fp8 bundled)
sd3.5_large-F16.ggufF1616.29 GBcity96/stable-diffusion-3.5-large-gguf →
sd3.5_large-Q8_0.ggufQ8_08.78 GBcity96/stable-diffusion-3.5-large-gguf →
sd3.5_large-Q5_1.ggufQ5_16.27 GBcity96/stable-diffusion-3.5-large-gguf →
sd3.5_large-Q5_0.ggufQ5_05.77 GBcity96/stable-diffusion-3.5-large-gguf →
sd3.5_large-Q4_1.ggufQ4_15.27 GBcity96/stable-diffusion-3.5-large-gguf →
sd3.5_large-Q4_0.ggufQ4_04.77 GBcity96/stable-diffusion-3.5-large-gguf →
sd3.5_large.safetensors16-bit16.46 GBstabilityai/stable-diffusion-3.5-large →

city96 GGUF repo only has legacy quants (Q4_0..Q8_0), no K-quants. Text encoders: clip_l + clip_g + T5-XXL.