Check my GPU →

Z-Image (base) VRAM requirements

The undistilled base of Z-Image Turbo: same 6B size, but 28–50 steps with CFG. Better for LoRA training and negative prompts.

Released 2026-01Licence: Apache-2.0Steps: 28–50Text encoder: Qwen3 4B (5.6 GB as FP8)
TypeImage
Parameters6.15B
Fits entirely from7 GB
8-bit or better from9 GB

01Measured on my RTX 5060 Ti 16 GB

16-bit · 1024×1024
78.0s / image
z_image_bf16.safetensors
Peak VRAM15.5 GB
Run range78.0–78.0 s

02The files, and how much VRAM each needs

FileSizeNeededMin. VRAMQualitySource
16-bit12.3 GB14.3 GB15 GBthe original weightsComfy-Org/z_image →
Q8_07.2 GB9.2 GB10 GBpractically identical to the originalunsloth/Z-Image-GGUF →
INT86.2 GB8.2 GB9 GBpractically identical to the originalComfy-Org/z_image →
Q6_K6.1 GB8.1 GB9 GBvery close to the originalunsloth/Z-Image-GGUF →
Q5_K_M5.6 GB7.6 GB8 GBclose; small differences in fine detailunsloth/Z-Image-GGUF →
Q4_K_M5.1 GB7.1 GB8 GBgood; some loss in fine detail and textunsloth/Z-Image-GGUF →
Q3_K_M4.6 GB6.6 GB7 GBnoticeable loss of detailunsloth/Z-Image-GGUF →
Q2_K4.0 GB6.0 GB7 GBheavy loss; a last resortunsloth/Z-Image-GGUF →

“Needed” = file + 1.2 GB working memory + 0.8 GB system reserve.

Free ComfyUI workflow, tested on my RTX 5060 Ti: 16-bit workflow (.json). Drag the file into ComfyUI. All workflows →

03By amount of VRAM

6GBOffload onlyQ5_K_M8GBRunsQ5_K_M10GBRuns wellQ8_012GBRuns wellQ8_016GBRuns well16-bit20GBRuns well16-bit24GBRuns well16-bit32GBRuns well16-bit

04Best GPU for Z-Image

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

05By graphics card

GPUVRAMVerdictBest fileNeeded
Desktop graphics cards
RTX 2060 6 GB6 GBOffload onlyQ5_K_M7.6 GB
RTX 3050 6 GB6 GBOffload onlyQ5_K_M7.6 GB
RTX 2070 Super 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 2080 Super 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3050 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3060 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3060 Ti 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3070 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3070 Ti 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 4060 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 4060 Ti 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 5050 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 5060 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 5060 Ti 8 GB8 GBRunsQ5_K_M7.6 GB
RX 7600 8 GB8 GBRunsQ5_K_M7.6 GB
RX 9050 8 GB8 GBRunsQ5_K_M7.6 GB
RX 9060 XT 8 GB8 GBRunsQ5_K_M7.6 GB
Arc B570 10 GB10 GBRuns wellQ8_09.2 GB
RTX 3080 10 GB10 GBRuns wellQ8_09.2 GB
RTX 2080 Ti 11 GB11 GBRuns wellQ8_09.2 GB
Arc B580 12 GB12 GBRuns wellQ8_09.2 GB
RTX 2060 12 GB12 GBRuns wellQ8_09.2 GB
RTX 3060 12 GB12 GBRuns wellQ8_09.2 GB
RTX 3080 12 GB12 GBRuns wellQ8_09.2 GB
RTX 3080 Ti 12 GB12 GBRuns wellQ8_09.2 GB
RTX 4070 12 GB12 GBRuns wellINT88.2 GB
RTX 4070 Super 12 GB12 GBRuns wellINT88.2 GB
RTX 4070 Ti 12 GB12 GBRuns wellINT88.2 GB
RTX 5070 12 GB12 GBRuns wellINT88.2 GB
RX 7700 XT 12 GB12 GBRuns wellQ8_09.2 GB
RX 9070 GRE 12 GB12 GBRuns wellINT88.2 GB
Arc A770 16 GB16 GBRuns well16-bit14.3 GB
RTX 4060 Ti 16 GB16 GBRuns well16-bit14.3 GB
RTX 4070 Ti Super 16 GB16 GBRuns well16-bit14.3 GB
RTX 4080 16 GB16 GBRuns well16-bit14.3 GB
RTX 4080 Super 16 GB16 GBRuns well16-bit14.3 GB
RTX 5060 Ti 16 GB16 GBRuns well16-bit14.3 GB
RTX 5070 Ti 16 GB16 GBRuns well16-bit14.3 GB
RTX 5080 16 GB16 GBRuns well16-bit14.3 GB
RX 7600 XT 16 GB16 GBRuns well16-bit14.3 GB
RX 7800 XT 16 GB16 GBRuns well16-bit14.3 GB
RX 7900 GRE 16 GB16 GBRuns well16-bit14.3 GB
RX 9060 XT 16 GB16 GBRuns well16-bit14.3 GB
RX 9070 16 GB16 GBRuns well16-bit14.3 GB
RX 9070 XT 16 GB16 GBRuns well16-bit14.3 GB
RX 7900 XT 20 GB20 GBRuns well16-bit14.3 GB
Arc Pro B60 24 GB24 GBRuns well16-bit14.3 GB
RTX 3090 24 GB24 GBRuns well16-bit14.3 GB
RTX 3090 Ti 24 GB24 GBRuns well16-bit14.3 GB
RTX 4090 24 GB24 GBRuns well16-bit14.3 GB
RX 7900 XTX 24 GB24 GBRuns well16-bit14.3 GB
Arc Pro B70 32 GB32 GBRuns well16-bit14.3 GB
RTX 5090 32 GB32 GBRuns well16-bit14.3 GB
Laptop GPUs
RTX 3050 Laptop 4 GB4 GBOffload onlyQ4_K_M7.1 GB
RTX 3050 Ti Laptop 4 GB4 GBOffload onlyQ4_K_M7.1 GB
RTX 2060 Laptop 6 GB6 GBOffload onlyQ5_K_M7.6 GB
RTX 3050 Laptop 6 GB6 GBOffload onlyQ5_K_M7.6 GB
RTX 3060 Laptop 6 GB6 GBOffload onlyQ5_K_M7.6 GB
RTX 4050 Laptop 6 GB6 GBOffload onlyQ5_K_M7.6 GB
RTX 2070 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 2070 Super Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 2080 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 2080 Super Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3070 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3070 Ti Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 3080 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 4060 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 4070 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 5050 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 5060 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 5070 Laptop 8 GB8 GBRunsQ5_K_M7.6 GB
RX 7600M 8 GB8 GBRunsQ5_K_M7.6 GB
RX 7600M XT 8 GB8 GBRunsQ5_K_M7.6 GB
RX 7600S 8 GB8 GBRunsQ5_K_M7.6 GB
RX 7700S 8 GB8 GBRunsQ5_K_M7.6 GB
RTX 4080 Laptop 12 GB12 GBRuns wellINT88.2 GB
RTX 5070 Laptop 12 GB12 GBRuns wellINT88.2 GB
RTX 5070 Ti Laptop 12 GB12 GBRuns wellINT88.2 GB
RX 7800M 12 GB12 GBRuns wellQ8_09.2 GB
RTX 3080 Laptop 16 GB16 GBRuns well16-bit14.3 GB
RTX 3080 Ti Laptop 16 GB16 GBRuns well16-bit14.3 GB
RTX 4090 Laptop 16 GB16 GBRuns well16-bit14.3 GB
RTX 5080 Laptop 16 GB16 GBRuns well16-bit14.3 GB
RX 7900M 16 GB16 GBRuns well16-bit14.3 GB
RTX 5090 Laptop 24 GB24 GBRuns well16-bit14.3 GB
Unified memory
Radeon 8060S (Strix Halo) 96 GB96 GBRuns well16-bit14.3 GB
Radeon 8065S (Gorgon Halo) 160 GB160 GBRuns well16-bit14.3 GB
Apple Silicon Macs (by memory)
Mac 16 GB12.7 GBRuns wellQ8_09.2 GB
Mac 18 GB14.4 GBRuns well16-bit14.3 GB
Mac 24 GB19.6 GBRuns well16-bit14.3 GB
Mac 32 GB26.8 GBRuns well16-bit14.3 GB
Mac 36 GB30.2 GBRuns well16-bit14.3 GB
Mac 48 GB40.2 GBRuns well16-bit14.3 GB
Mac 64 GB55.7 GBRuns well16-bit14.3 GB
Mac 96 GB85 GBRuns well16-bit14.3 GB
Mac 128 GB115.4 GBRuns well16-bit14.3 GB
Mac 192 GB175.4 GBRuns well16-bit14.3 GB
Mac 256 GB236.9 GBRuns well16-bit14.3 GB
Mac 512 GB498.1 GBRuns well16-bit14.3 GB

06Text encoder, VAE and other files

FileFolderSizeWhen
Qwen3 4B FP8
qwen_3_4b_fp8_mixed.safetensors
models/text_encoders5.6 GBtext encoder · smaller, recommendedDownload →
Qwen3 4B BF16
qwen_3_4b.safetensors
models/text_encoders8.0 GBtext encoder · alternativeDownload →
FLUX.1 VAE (ae)
ae.safetensors
models/vae0.3 GBrequiredDownload →

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

07Where 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

08Measured and reported results

LabelGPUSetupResultPeak VRAMDateSource
measuredRTX 5060 Ti 16 GBz_image_bf16.safetensors · 1024x1024 · 30 steps
Base (non-Turbo) model: 30 steps, CFG 4, res_multistep/simple, ModelSamplingAuraFlow 3. Text encoder Qwen3 4B FP8 (mixed), FLUX VAE (ae). 3 timed runs after 1 warm-up (82.3 s incl.
77.97 s / image15.5 GB2026-10-12

09Training a LoRA for Z-Image

TrainerVRAMTypeSettings and quoteSource
OneTrainer8 GBexample runOfficial preset for Tongyi-MAI/Z-Image: 512px, batch 2, transformer fp8, TE fp8, layer offload fraction 0.6, torch.compile; De-Turbo 8GB preset also exists
“#z-image LoRA 8GB.json”
github.com →
SimpleTuner10 GBstated minimumSame guide ('Z-Image [base / turbo]'): rank-16 LoRA with NF4 base ~10-12 GB, 512px, batch 1
“~10–12G VRAM when quantising to NF4 + bf16 base/LoRA weights”
github.com →

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

10Every file tracked for Z-Image

FileTypeSizeRepo
z_image_bf16.safetensorsBF1612.31 GBComfy-Org/z_image →
z_image_int8_convrot.safetensorsINT86.20 GBComfy-Org/z_image →
z-image-BF16.ggufBF1612.31 GBunsloth/Z-Image-GGUF →
z-image-F16.ggufF1612.31 GBunsloth/Z-Image-GGUF →
z-image-Q8_0.ggufQ8_07.22 GBunsloth/Z-Image-GGUF →
z-image-Q6_K.ggufQ6_K6.10 GBunsloth/Z-Image-GGUF →
z-image-Q5_K_M.ggufQ5_K_M5.58 GBunsloth/Z-Image-GGUF →
z-image-Q5_1.ggufQ5_15.53 GBunsloth/Z-Image-GGUF →
z-image-Q5_K_S.ggufQ5_K_S5.29 GBunsloth/Z-Image-GGUF →
z-image-Q5_0.ggufQ5_05.26 GBunsloth/Z-Image-GGUF →
z-image-Q4_K_M.ggufQ4_K_M5.07 GBunsloth/Z-Image-GGUF →
z-image-Q4_1.ggufQ4_14.85 GBunsloth/Z-Image-GGUF →
z-image-Q4_K_S.ggufQ4_K_S4.79 GBunsloth/Z-Image-GGUF →
z-image-Q3_K_L.ggufQ3_K_L4.60 GBunsloth/Z-Image-GGUF →
z-image-Q4_0.ggufQ4_04.59 GBunsloth/Z-Image-GGUF →
z-image-Q3_K_M.ggufQ3_K_M4.56 GBunsloth/Z-Image-GGUF →
z-image-Q3_K_S.ggufQ3_K_S4.36 GBunsloth/Z-Image-GGUF →
z-image-Q2_K.ggufQ2_K4.01 GBunsloth/Z-Image-GGUF →

Non-distilled base of Z-Image Turbo (Tongyi-MAI/Z-Image, same 6B S3-DiT, same file sizes) - needs ~28-50 steps with CFG, better for fine-tuning/LoRA training and negative prompts. Same text encoder and VAE as Z-Image Turbo.