.ai

những mô hình AI mã nguồn mở tốt nhất

Tìm kiếm

VRAM requirements for every popular AI image and video model (2026): full precision, FP8 and GGUF

Q: Generation is painfully slow although nothing crashes

Automatic offloading to system RAM. Drop to the next precision (FP8 to Q6, Q6 to Q4) until Shared GPU memory stays near zero.

Q: FLUX.2 Dev "works" on 12 GB but needs 70 GB of RAM

FP8 weights swapped from system memory. Use Klein 4B/9B locally and FLUX.2 Dev on the cloud or a 32 GB card.

Chủ đề: Lỗi và cách khắc phục

Bởi P.I. Panda

Đăng ngày 2026-10-06

How much VRAM you need for SD 1.5, SDXL, SD 3.5, Z-Image Turbo, FLUX.1, FLUX.2 Klein and Dev, Qwen-Image, Wan 2.2, HunyuanVideo 1.5 and LTX-2, at full precision, FP8 and 4-bit GGUF, with the comfortable card and licence for each, what every GPU size from 8 to 48 GB can run, and the order of fixes when a model does not fit.

Trang này chưa được dịch nên đang hiển thị bằng tiếng Anh.

Tổng quan

Short answer: 8 GB of

VRAM

runs

SDXL

and the small fast models such as

and

. 12 GB runs

and

once they are quantised. 24 GB runs almost every image model plus

Wan 2.2

video in FP8. The newest large models,

FLUX.2 Dev

and

LTX-2

at full quality, want 32 GB or more.

The table below lists every popular open model with the memory it needs at full precision, at FP8 and as a 4-bit GGUF file, generating at about 1024x1024. Different testers count the text encoder differently, so treat every number as a range rather than a hard limit. Three things the table hides: SDXL is still the cheapest model to run well (the base fits in 8 GB, but

LoRAs

ControlNet

and

upscaling

add memory, which is why 12 GB is the comfortable number); FLUX.2 Klein 4B is the best fit for small cards right now (about 13 GB at full precision, roughly half in FP8, Apache 2.0); and FLUX.2 Dev is not a home model at full precision (one

ComfyUI

user ran FP8 on a 12 GB RTX 3060, using about 70 GB of system RAM to get there).

Precision changes everything. BF16/FP16 is full quality and full size, about 2 GB of VRAM per billion parameters. FP8 halves that, about 1 GB per billion, and most people cannot tell FP8 output from full precision. GGUF files (Q8, Q5, Q4 and lower) shrink models further: Q8 is close to full quality, Q4 trades some detail for a much smaller footprint, and below Q4 quality drops quickly. The model is also not the only thing in memory: text encoders (T5 for FLUX.1, a language model for Qwen-Image and

FLUX.2

), the VAE, LoRAs, ControlNet and the image itself all take space, and ComfyUI moves some of these to system RAM automatically, which is why two people with the same card report different numbers. Figures checked 6 October 2026 against model cards and community tests; updated monthly.

Tham khảo

Model

Full precision

FP8

4-bit / GGUF

Comfortable card

Stable Diffusion 1.5

0.9B

CreativeML OpenRAIL-M. Runs on almost anything.

SDXL and fine-tunes: Pony, Illustrious, NoobAI

3.5B

OpenRAIL++ (fine-tunes vary). 12 GB leaves room for LoRAs, ControlNet and upscaling in one workflow.

Stability Community License.

Apache 2.0. 8 steps; the fast model for small cards.

Apache 2.0, commercial use allowed. Best fit for 8 GB cards in 2026.

FLUX Non-Commercial License (outputs usable commercially).

Non-commercial licence. The fp8 T5 encoder saves about 4 GB on its own.

Full precision

~24 GB

more with the text encoder loaded

Apache 2.0. Strong text rendering; needs its Qwen2.5-VL encoder in FP8 on anything under 32 GB.

Non-commercial licence. Expect slow generations without a 32 GB card.

Full precision

~90 GB

with its text encoder

HunyuanImage 3.0

80B MoE

Tencent Community License. Datacenter only; use it through a hosted service.

Full precision

multi-GPU

at every precision

FP8

n/a

4-bit / GGUF

n/a

Comfortable card

datacenter

official, tested on an RTX 4090

Apache 2.0. Two 14B experts, one in VRAM at a time, which is why 24 GB cards handle it. Wan 2.5-3.0 are API only.

Video 720p, 1080p with its upscaler

Tencent Community License.

Full precision

14 GB

official, with offloading

80 GB for best quality

LTX-2 / LTX-2.3

19B / 22B

Video up to 4K

LTX-2 Community License. Synced audio.

Every row of this table, no VRAM required

BitVector Prism runs SDXL, Z-Image, FLUX.1, FLUX.2 and Qwen-Image at full precision on its own GPUs. Pick a model from this site, type a prompt, generate in the browser. Nothing to download.

Dùng thử BitVector

Từng bước

Find your VRAM (Task Manager > Performance > GPU, or nvidia-smi) and match it to a tier below; then pick models whose "comfortable card" column is at or below it.
6-8 GB (RTX 3060 Ti, 4060, 2070):
SD 1.5
and SDXL including
Pony
and Illustrious with a LoRA or two; Z-Image Turbo and FLUX.2 Klein 4B in FP8 or GGUF; FLUX.1 in GGUF Q4, slowly;
Wan 2.1
1.3B at 480p and Wan 2.2 5B in GGUF with offloading.
12 GB (RTX 3060 12GB, 4070, 5070): everything above, faster, with room for ControlNet and upscaling;
FLUX.1 Dev
and Schnell in FP8 or GGUF; Qwen-Image in GGUF Q4;
FLUX.2 Klein 9B
in GGUF; Wan 2.2 A14B in GGUF Q4 with offloading at low resolution and short clips.
16 GB (RTX 4060 Ti 16GB, 4070 Ti Super, 5070 Ti, 5080): FLUX.1 Dev in FP8 with headroom; Z-Image Turbo and Klein 4B at full precision; Qwen-Image in GGUF Q4-Q5;
HunyuanVideo
1.5 with offloading; Wan 2.2 A14B in GGUF Q5 at 720p (community reports).
24 GB (RTX 3090, 4090): almost every image model except FLUX.2 Dev at full precision; Qwen-Image in FP8 or Q8; Wan 2.2 A14B in FP8 and Wan 2.2 5B at full precision; LTX-2 in FP4; most LoRA training for SDXL and Flux.
32 GB (RTX 5090): FLUX.2 Dev in FP8; LTX-2 and
LTX-2.3
in FP8; Wan 2.2 A14B at 720p without running
out of memory
partway through a clip.
48 GB and up (RTX 6000 Ada, A100, H100): full-precision Qwen-Image, LTX-2 and FLUX.2 Dev with room to spare; long or high-resolution video, batch generation and serious training.
Download the precision for your tier from the model page on this site (it lists each variant with its file size), load text encoders in FP8, and keep 32 GB of system RAM (64 GB for the largest models) for offloading.

Ví dụ

FLUX.1 Dev on a 12 GB card (ComfyUI)

Load Diffusion Model: flux1-dev-fp8-e4m3fn (11.9 GB) or UnetLoaderGGUF flux1-dev-Q6_K (9.8 GB) | DualCLIPLoader: clip_l + t5xxl_fp8_e4m3fn_scaled | Load VAE: ae

1024x1024, 20 steps, euler, simple, guidance 3.5 | VAE Decode (Tiled) if the last step fails

Qwen-Image on 16 GB

UnetLoaderGGUF: qwen_image_Q4_K_M.gguf (~12.5 GB) | CLIPLoader: qwen_2.5_vl_7b_fp8_scaled, device cpu | Load VAE: qwen_image_vae

1328x1328 native; start at 1024x1024, 20 steps, cfg 2.5

The same models with no card at all

/render <flux-dev> a studio photo of a vintage camera on a walnut desk, soft window light /size:1024x1024 (PirateDiffusion, Telegram)

Mẹo

When you see CUDA out of memory, do these in order: switch to an FP8 or GGUF version of the same model (the biggest single saving); lower the resolution or clip length (memory grows with pixel count, 1536 to 1024 px cuts a lot); use tiled VAE decoding; load the text encoder in FP8 or keep it on the CPU; close other GPU apps (browsers and games hold VRAM too).
Offloading works but you pay in speed and system RAM: a model can run on a card "too small" for it, several times slower, and you may need 32-64 GB or more of system RAM.
Licences in one line: Z-Image Turbo, FLUX.2 Klein 4B, FLUX.1 Schnell, Qwen-Image and
Wan
2.1/2.2 are Apache 2.0 (commercial use allowed); FLUX.1 Dev, FLUX.2 Dev and Klein 9B use Black Forest Labs' non-commercial licence, which forbids using the model in a commercial product or service but lets you use the generated images commercially. Licences change; the model page links the current one.
Buying a 32 GB card to run FLUX.2 Dev or LTX-2 for a weekend project rarely pays off; if the model you want sits in the 24 GB+ rows, renting by the hour or a flat-rate cloud plan is usually cheaper and faster than offloading on a small card.
Per-billion-parameter rule: 2 GB at BF16, 1 GB at FP8, about 0.6 GB at Q4, plus 2-5 GB for the encoder, VAE and latent.
For video, resolution and clip length matter as much as the model; the companion guide on
VRAM for AI video
breaks that down.

Check the licence column and the model card before using a model in a product.

The 24 GB+ rows, from your phone

PirateDiffusion hosts FLUX.2 Dev, Qwen-Image, Wan 2.2 14B and LTX-2 on server hardware and runs them from Telegram commands. Unlimited generation on a fixed monthly price; no card, no offloading, no 70 GB of RAM.

Dùng thử PirateDiffusion

Câu hỏi

Can I run Flux on 8 GB of VRAM?

Yes, with a quantised file. FLUX.1 in GGUF Q4 needs about 7 GB, and FLUX.2 Klein 4B in FP8 about 8 GB. Expect slower generations than on a 12 or 16 GB card. Full-precision FLUX.1 needs about 24 GB.

Is 12 GB enough for SDXL?

Yes. SDXL itself fits in 8 GB; 12 GB gives you room for LoRAs, ControlNet and upscaling in the same workflow.

How much VRAM does Wan 2.2 need?

It depends on the version. The 5B model runs on 8 GB as a GGUF file and comfortably on 24 GB. The A14B model officially needs 80 GB without offloading, but its FP8 version runs well on 24 GB and GGUF versions run on 12-16 GB with offloading.

Does system RAM matter?

Yes, once you start offloading. 32 GB is a sensible minimum for Flux and Wan, and the largest models (FLUX.2 Dev on a 12 GB card, for example) can use 64 GB or more.

Which models can I use commercially?

Z-Image Turbo, FLUX.2 Klein 4B, FLUX.1 Schnell, Qwen-Image and Wan 2.1/2.2 (Apache 2.0). FLUX.1 Dev, FLUX.2 Dev and Klein 9B are non-commercial for the model itself, though the images you generate may be used commercially. Always check the model card.