Qwen-Image VRAM requirements
Alibaba's 20B text-to-image model, known for rendering long text and posters well. Big: the 8-bit file alone is over 20 GB.
TypeImage
Parameters20B
Fits entirely from10 GB
8-bit or better from24 GB
01Measured on my RTX 5060 Ti 16 GB
Q4_K_M · 1024×1024
123.3s / image
qwen-image-Q4_K_M.gguf
Peak VRAM15.7 GB
Run range122.9–124.1 s
02The files, and how much VRAM each needs
| File | Size | Needed | Min. VRAM | Quality | Source |
|---|---|---|---|---|---|
| 16-bit | 40.9 GB | 43.7 GB | 44 GB | the original weights | Comfy-Org/Qwen-Image_ComfyUI → |
| Q8_0 | 21.8 GB | 24.6 GB | 25 GB | practically identical to the original | city96/Qwen-Image-gguf → |
| FP8 | 20.4 GB | 23.2 GB | 24 GB | practically identical to the original | Comfy-Org/Qwen-Image_ComfyUI → |
| Q6_K | 16.8 GB | 19.6 GB | 20 GB | very close to the original | city96/Qwen-Image-gguf → |
| Q5_K_M | 14.9 GB | 17.7 GB | 18 GB | close; small differences in fine detail | city96/Qwen-Image-gguf → |
| Q4_K_M | 13.1 GB | 15.9 GB | 16 GB | good; some loss in fine detail and text | city96/Qwen-Image-gguf → |
| Q3_K_M | 9.7 GB | 12.5 GB | 13 GB | noticeable loss of detail | city96/Qwen-Image-gguf → |
| Q2_K | 7.1 GB | 9.9 GB | 10 GB | heavy loss; a last resort | city96/Qwen-Image-gguf → |
“Needed” = file + 2 GB working memory + 0.8 GB system reserve.
Free ComfyUI workflow, tested on my RTX 5060 Ti: GGUF Q4_K_M workflow (.json). Drag the file into ComfyUI. All workflows →
03By amount of VRAM
04Best GPU for Qwen-Image
The cheapest cards (by launch price) that run it well, and every card sorted by memory: best GPU for Qwen-Image → Planning bigger images or longer clips? Open the calculator →
05By graphics card
| GPU | VRAM | Verdict | Best file | Needed |
|---|---|---|---|---|
| Desktop graphics cards | ||||
| RTX 2060 6 GB | 6 GB | Not practical | Q2_K | 9.9 GB |
| RTX 3050 6 GB | 6 GB | Not practical | Q2_K | 9.9 GB |
| RTX 2070 Super 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 2080 Super 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3050 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3060 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3060 Ti 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3070 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3070 Ti 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 4060 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 4060 Ti 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 5050 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 5060 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 5060 Ti 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 7600 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 9050 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 9060 XT 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| Arc B570 10 GB | 10 GB | Tight | Q2_K | 9.9 GB |
| RTX 3080 10 GB | 10 GB | Tight | Q2_K | 9.9 GB |
| RTX 2080 Ti 11 GB | 11 GB | Tight | Q2_K | 9.9 GB |
| Arc B580 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 2060 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 3060 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 3080 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 3080 Ti 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 4070 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 4070 Super 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 4070 Ti 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 5070 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RX 7700 XT 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RX 9070 GRE 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| Arc A770 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 4060 Ti 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 4070 Ti Super 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 4080 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 4080 Super 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 5060 Ti 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 5070 Ti 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 5080 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 7600 XT 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 7800 XT 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 7900 GRE 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 9060 XT 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 9070 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 9070 XT 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 7900 XT 20 GB | 20 GB | Runs | Q6_K | 19.6 GB |
| Arc Pro B60 24 GB | 24 GB | Runs well | FP8 | 23.2 GB |
| RTX 3090 24 GB | 24 GB | Runs well | FP8 | 23.2 GB |
| RTX 3090 Ti 24 GB | 24 GB | Runs well | FP8 | 23.2 GB |
| RTX 4090 24 GB | 24 GB | Runs well | FP8 | 23.2 GB |
| RX 7900 XTX 24 GB | 24 GB | Runs well | FP8 | 23.2 GB |
| Arc Pro B70 32 GB | 32 GB | Runs well | Q8_0 | 24.6 GB |
| RTX 5090 32 GB | 32 GB | Runs well | FP8 | 23.2 GB |
| Laptop GPUs | ||||
| RTX 3050 Laptop 4 GB | 4 GB | Not practical | Q2_K | 9.9 GB |
| RTX 3050 Ti Laptop 4 GB | 4 GB | Not practical | Q2_K | 9.9 GB |
| RTX 2060 Laptop 6 GB | 6 GB | Not practical | Q2_K | 9.9 GB |
| RTX 3050 Laptop 6 GB | 6 GB | Not practical | Q2_K | 9.9 GB |
| RTX 3060 Laptop 6 GB | 6 GB | Not practical | Q2_K | 9.9 GB |
| RTX 4050 Laptop 6 GB | 6 GB | Not practical | Q2_K | 9.9 GB |
| RTX 2070 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 2070 Super Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 2080 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 2080 Super Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3070 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3070 Ti Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 3080 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 4060 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 4070 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 5050 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 5060 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 5070 Laptop 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 7600M 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 7600M XT 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 7600S 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RX 7700S 8 GB | 8 GB | Offload only | Q2_K | 9.9 GB |
| RTX 4080 Laptop 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 5070 Laptop 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 5070 Ti Laptop 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RX 7800M 12 GB | 12 GB | Tight | Q2_K | 9.9 GB |
| RTX 3080 Laptop 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 3080 Ti Laptop 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 4090 Laptop 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 5080 Laptop 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RX 7900M 16 GB | 16 GB | Runs | Q4_K_M | 15.9 GB |
| RTX 5090 Laptop 24 GB | 24 GB | Runs well | FP8 | 23.2 GB |
| Unified memory | ||||
| Radeon 8060S (Strix Halo) 96 GB | 96 GB | Runs well | 16-bit | 43.7 GB |
| Radeon 8065S (Gorgon Halo) 160 GB | 160 GB | Runs well | 16-bit | 43.7 GB |
| Apple Silicon Macs (by memory) | ||||
| Mac 16 GB | 12.7 GB | Tight | Q3_K_M | 12.5 GB |
| Mac 18 GB | 14.4 GB | Tight | Q3_K_M | 12.5 GB |
| Mac 24 GB | 19.6 GB | Runs | Q5_K_M | 17.7 GB |
| Mac 32 GB | 26.8 GB | Runs well | Q8_0 | 24.6 GB |
| Mac 36 GB | 30.2 GB | Runs well | Q8_0 | 24.6 GB |
| Mac 48 GB | 40.2 GB | Runs well | Q8_0 | 24.6 GB |
| Mac 64 GB | 55.7 GB | Runs well | 16-bit | 43.7 GB |
| Mac 96 GB | 85 GB | Runs well | 16-bit | 43.7 GB |
| Mac 128 GB | 115.4 GB | Runs well | 16-bit | 43.7 GB |
| Mac 192 GB | 175.4 GB | Runs well | 16-bit | 43.7 GB |
| Mac 256 GB | 236.9 GB | Runs well | 16-bit | 43.7 GB |
| Mac 512 GB | 498.1 GB | Runs well | 16-bit | 43.7 GB |
06Text encoder, VAE and other files
| File | Folder | Size | When | |
|---|---|---|---|---|
| Qwen2.5-VL 7B FP8 qwen_2.5_vl_7b_fp8_scaled.safetensors | models/text_encoders | 9.4 GB | text encoder · smaller, recommended | Download → |
| Qwen2.5-VL 7B BF16 qwen_2.5_vl_7b.safetensors | models/text_encoders | 16.6 GB | text encoder · alternative | Download → |
| Qwen-Image VAE qwen_image_vae.safetensors | models/vae | 0.3 GB | required | Download → |
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
| File | Folder | Loader node |
|---|---|---|
| Diffusion model (.safetensors: 16-bit, FP8, INT8) | ComfyUI/models/diffusion_models | Load Diffusion Model |
| GGUF file (.gguf) | ComfyUI/models/unet | Unet Loader (GGUF) - from the ComfyUI-GGUF node pack |
| Text encoder | ComfyUI/models/text_encoders | Load CLIP / DualCLIPLoader (or the GGUF versions) |
| VAE | ComfyUI/models/vae | Load VAE |
08Measured and reported results
| Label | GPU | Setup | Result | Peak VRAM | Date | Source |
|---|---|---|---|---|---|---|
| measured | RTX 5060 Ti 16 GB | qwen-image-Q4_K_M.gguf · 1024x1024 · 20 steps 20 steps, CFG 2.5, euler/simple, ModelSamplingAuraFlow 3.1. Text encoder Qwen2.5-VL 7B FP8 (scaled), Qwen-Image VAE. 3 timed runs after 1 warm-up (138.7 s incl. loading). VRAM was | 123.32 s / image | 15.7 GB | 2026-10-12 | |
| reported | Radeon 8060S (Strix Halo) 96 GB | Qwen-Image-2512 BF16 + 4-step Lightning LoRA · 1328x1328 · 4 steps “"workflow": "Qwen-Image-2512-BF16-4-Step-LoRA.json", ... "duration_seconds": 75.37661480903625” kyuz0 Strix Halo ComfyUI toolbox benchmark (Ryzen AI Max, ROCm); cold run incl. model load, flags --disable-mmap --gpu-only --disable-smart-memory --cache-none; resolution from ben | 75.38 s / image | — | 2026-02-13 | raw.githubusercontent.com → |
| reported | RTX 3060 12 GB | 20 steps “Qwen-ImageがRTX 3060（12GB）で動くと聞いて、早速ComfyUI版をお試し。確かに問題なく動いて、20stepでちょうど5分。” 'ちょうど5分' = exactly 5 minutes (converted to 300 s); ComfyUI version, file/resolution not stated; quote taken from search index (x.com not fetchable); date from tweet ID | 300 s / image | — | 2025-08-05 | x.com → |
Other people's numbers are copied as published, with a link. Send yours.
09Training a LoRA for Qwen-Image
| Trainer | VRAM | Type | Settings and quote | Source |
|---|---|---|---|---|
| musubi-tuner | 12 GB | stated minimum | 1024x1024, batch 1, bf16 mixed precision, gradient checkpointing, xformers, --fp8_base --fp8_scaled + --blocks_to_swap 45; 64GB RAM recommended. Table: none 42GB, fp8 30GB, +swap16 24GB “+ --blocks_to_swap 45|12GB” | github.com → |
| OneTrainer | 16 GB | example run | Official preset: 512px, batch 2, transformer fp8, TE fp8, layer offload fraction 0.5 (24GB preset uses 0.1) “#qwen LoRA 16GB.json” | github.com → |
| ai-toolkit | 24 GB | example run | Official example config: 3-bit (uint3) base with accuracy recovery adapter, fp8 TE, cached text embeddings, low_vram, rank 16, batch 1, gradient checkpointing “# 3bit is required for 24GB” | github.com → |
| diffusion-pipe | 24 GB | stated minimum | Example config: fp8 transformer, blocks_to_swap 8, rank 32, activation checkpointing, 640px dataset suggested, expandable_segments “You will need block swapping. See the [example 24GB VRAM config]” | github.com → |
| SimpleTuner | 24 GB | stated minimum | int2-quanto or nf4-bnb base, batch 1, gradient checkpointing, LoRA rank 1-8, 512-768px start; 40GB+ strongly recommended “A 24GB GPU is the absolute minimum, and even then you'll need extensive quantization and careful configuration.” | github.com → |
reported From each trainer's own docs, 2026-09-25. Why the figures differ →
10Every file tracked for Qwen-Image
| File | Type | Size | Repo |
|---|---|---|---|
| qwen_image_bf16.safetensors (original) | BF16 | 40.86 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen_image_2512_bf16.safetensors (2512) | BF16 | 40.86 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen_image_fp8_hq.safetensors (original) | FP8 | 22.74 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen_image_fp8mixed.safetensors (original) | FP8 | 20.53 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen_image_2512_fp8_e4m3fn.safetensors (2512) | FP8 | 20.43 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen_image_fp8_e4m3fn.safetensors (original) | FP8 | 20.43 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen_image_nvfp4.safetensors (original) | NVFP4 | 19.77 GB | Comfy-Org/Qwen-Image_ComfyUI → |
| qwen-image-BF16.gguf | BF16 | 40.87 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q8_0.gguf | Q8_0 | 21.76 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q6_K.gguf | Q6_K | 16.82 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q5_1.gguf | Q5_1 | 15.39 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q5_K_M.gguf | Q5_K_M | 14.93 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q5_0.gguf | Q5_0 | 14.40 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q5_K_S.gguf | Q5_K_S | 14.12 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q4_K_M.gguf | Q4_K_M | 13.07 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q4_1.gguf | Q4_1 | 12.84 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q4_K_S.gguf | Q4_K_S | 12.14 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q4_0.gguf | Q4_0 | 11.85 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q3_K_M.gguf | Q3_K_M | 9.68 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q3_K_S.gguf | Q3_K_S | 8.95 GB | city96/Qwen-Image-gguf → |
| qwen-image-Q2_K.gguf | Q2_K | 7.06 GB | city96/Qwen-Image-gguf → |
MMDiT 20B. Text encoder Qwen2.5-VL-7B (see qwen2.5-vl-7b entry). Qwen-Image-2512 is an updated checkpoint of the same architecture (same sizes). nvfp4 needs RTX 50-series for speedup.