Image models need far less GPU memory than the larger video models. By their makers' own figures, FLUX.2 [klein] 4B needs about 13 GB and SDXL works with 8 GB. Wan 2.2's T2V-A14B video model peaks at 59.8 GB for a 720p clip on one H100.
Settings change these figures a lot: resolution, clip length, precision and whether parts of the model wait in the computer's main memory. The tables below cover 12 open models at the settings their makers tested. For each one, they show the size of its files and the card sizes that fit, and the price boxes show GPUs you can rent.
Who this is for
- you make images or video with ComfyUI or Diffusers and want to know which card a model needs
- you are choosing between a 24 GB, 48 GB and 80 GB GPU to rent
- you want the makers' own memory figures, not forum guesses
Not for you if you run language models such as Llama or Qwen: our GPU memory guide for LLMs covers them.
Key numbers
- 8 GB of GPU memory is enough for SDXL, according to Stability AI Official page,
- 13 GB for FLUX.2 [klein] 4B, according to Black Forest Labs Official page,
- 59.8 GB peak for a 720p clip from Wan 2.2 T2V-A14B on one H100, in Wan's own table Official page,
- 112.8 GB the files of FLUX.2 [dev] at 16-bit, with its text encoder Our calculation,
What fills GPU memory when you make an image or a clip
A model that makes images or video has three main parts, and each takes its own share of VRAM, the GPU's memory:
- the text encoder, which turns your prompt into numbers once, at the start
- the image or video model itself, which works on the picture over many steps and is by far the largest part in most models
- the decoder, called a VAE, which turns the finished result into pixels at the end
- Text encoder (reads the prompt): connects to Image or video model (prompt as numbers).
- Image or video model (many steps on the picture): connects to VAE decoder (finished result).
- VAE decoder (turns the result into pixels)
On top of the files, the GPU needs working memory for the picture itself. It grows with the resolution and, for video, with the number of frames. That is why the makers' figures below are peaks for a stated setting, not fixed sizes.
Each file stores every parameter in a set number of bits. At 16-bit, a billion parameters take 2 GB; at 8-bit, 1 GB. Some makers publish files at 32-bit, which tools usually load at 16-bit, so we halve those sizes where the table says so.
Image models: SDXL and the Flux family
SDXL is the smallest of the image models here, and Stability AI says it works with 8 GB. The Flux models from Black Forest Labs range widely. FLUX.2 [klein] 4B fits in about 13 GB, its maker says, while the files of FLUX.2 [dev] come to 112.8 GB at 16-bit.
Black Forest Labs' figure for FLUX.2 [klein] 4B is below the 16.0 GB its files take. So its set-up cannot hold every part at 16-bit in GPU memory at once. File sums are an upper limit when parts can wait in main memory.
FLUX.2 [klein] 4B fits a 16 GB card by its maker's figure, while FLUX.2 [dev] needs 4-bit compression or offloading to fit even 24 GB.
| Model (maker) | Parameters | Files at 16-bit (GB) | Maker's memory figure | Card that fits, by our reading |
|---|---|---|---|---|
| SDXL base 1.0 (Stability AI) | 3.5 billion | 6.9, all parts in one file | works with 8 GB, says Stability AI | 8 GB |
| FLUX.1 [dev] (Black Forest Labs) | 12 billion | 33.9: model 23.8, text encoders 9.8, VAE 0.3 | none published; its example code offloads parts to the computer's main memory | 48 GB at 16-bit; 24 GB with the 8-bit model and the text encoders offloaded |
| FLUX.1 [schnell] (Black Forest Labs) | 12 billion | 33.9: model 23.8, text encoders 9.8, VAE 0.3 | none published | as FLUX.1 [dev] |
| FLUX.2 [dev] (Black Forest Labs) | 32 billion | 112.8: model 64.5, text encoder 48.0, VAE 0.3 | more than 80 GB without offloading; about 20 GB free at 4-bit (Hugging Face's guide) | 24 GB at 4-bit; at 16-bit, more than an 80 GB card |
| FLUX.2 [klein] 4B (Black Forest Labs) | 3.9 billion | 16.0: model 7.8, text encoder 8.0, VAE 0.2 | about 13 GB | 16 GB |
| FLUX.2 [klein] 9B (Black Forest Labs) | 9.1 billion | 34.7: model 18.2, text encoder 16.4, VAE 0.2 | none on the pages we could read | 48 GB at 16-bit |
| flux-3-image-klein (Black Forest Labs) | 5.6 billion in the image model | 20.3: model 11.1, text encoder 8.9, VAE 0.3 | peaks of 31 GB at 16-bit, 26 GB at 8-bit and 25 GB at 4-bit for a 1,024 x 1,024 image; 28 GB more with its prompt upsampler | 32 GB; 80 GB with the prompt upsampler |
The text encoder is easy to overlook. According to Hugging Face's guide, FLUX.2 [dev] uses a version of Mistral Small 3.1 as its encoder. Its files take 48.0 GB, three quarters the size of the image model.
Hugging Face shows three ways to fit the model on smaller cards. A 4-bit version needs about 20 GB of free GPU memory. With the text encoder on a remote service as well, it needs 18 GB. With group offloading, it runs on as little as 8 GB, if the computer has about 32 GB of main memory.
Black Forest Labs also publishes 8-bit (FP8) and 4-bit (NVFP4) versions of several models, including FLUX.1 [dev]. At 8-bit, FLUX.1 [dev]'s image model shrinks to about 12 GB by our calculation.
Licences differ as much as sizes. FLUX.1 [schnell] and FLUX.2 [klein] 4B use the Apache 2.0 licence. FLUX.1 [dev], FLUX.2 [dev] and FLUX.2 [klein] 9B come under non-commercial licences from Black Forest Labs. The licence of flux-3-image-klein is its own. Read the terms before you use any of them commercially.
Video models: Wan, LTX and HunyuanVideo
Video models need more memory because they work on many frames at once. Wan 2.2's TI2V-5B is the exception. Wan measured a peak of 22.9 GB for a 720p clip on an RTX 4090. It also says the model makes a five-second 720p clip in under nine minutes on one consumer GPU.
Wan 2.2 TI2V-5B runs on a 24 GB card by Wan's own test. The 14-billion-parameter Wan models and HunyuanVideo at 720p need an 80 GB card.
| Model (maker) | Parameters | Files at 16-bit (GB) | Maker's memory figure | Card that fits, by our reading |
|---|---|---|---|---|
| Wan 2.2 TI2V-5B (Wan) | 5 billion | 22.8: model 10.0 (32-bit files halved), text encoder 11.4, VAE 1.4 | peak of 22.9 GB for a 720p clip on an RTX 4090, offloading on | 24 GB |
| Wan 2.2 T2V-A14B and I2V-A14B (Wan) | 27 billion: two experts of about 14 billion | 69.0: two experts of 28.6 (32-bit files halved), text encoder 11.4, VAE 0.5 | peaks of 41.3 GB at 480p and 59.8 GB at 720p on one H100, offloading on; one GPU needs at least 80 GB | 80 GB |
| LTX-2.5 (Lightricks), video with sound | 22 billion, by its file names | 70.1: model 42.0, text encoder 26.3, VAEs 1.8 | none published; low-memory options are 8-bit and offloading | 80 GB at 16-bit; 48 GB with its 8-bit files for ComfyUI (38.7 GB) |
| HunyuanVideo (Tencent) | more than 13 billion | 26.6: model 25.6, VAE 1.0; its text encoders come separately | peaks of 45 GB at 544 x 960 and 60 GB at 720 x 1,280, both 129 frames; 8-bit saves about 10 GB | 48 GB at 544 x 960; 80 GB at 720p |
| HunyuanVideo 1.5 (Tencent) | 8.3 billion | 16.7 for the model (32-bit files halved); its text encoders come separately | at least 14 GB with model offloading | 16 GB with offloading |
Wan 2.2 T2V-A14B and I2V-A14B are each two models of about 14 billion parameters. One handles the early steps of a clip, the other the later ones. Wan says this keeps memory nearly the same as one 14-billion model, because only one of the two works at each step. Its table shows the same peak on an A100 and on an H100, so the card's speed changes the time, not the memory.
LTX-2.5, from Lightricks, makes video and sound together. Its page offers 8-bit and offloading options for smaller cards but gives no memory figure. The table therefore shows only our file sums. Its 4-bit files are marked for NVIDIA's Blackwell cards.
Tencent released HunyuanVideo in December 2024. HunyuanVideo 1.5, from November 2025, is smaller, at 8.3 billion parameters, and Tencent says it runs with 14 GB when offloading is on.
What changes the answer: resolution, clip length, precision and offloading
The makers' own figures show how much each setting moves the peak:
- resolution: the same Wan 2.2 T2V-A14B model peaks at 41.3 GB for a 480p clip and 59.8 GB for 720p
- frame size: HunyuanVideo needs 45 GB at 544 x 960 and 60 GB at 720 x 1,280, for the same 129 frames
- precision: 8-bit weights save about 10 GB on HunyuanVideo, and flux-3-image-klein drops from 31 GB at 16-bit to 26 GB at 8-bit
- extras: the optional prompt upsampler of flux-3-image-klein needs about 28 GB more
Compressing the image model saves only that model's share. At 8-bit, the model file of flux-3-image-klein is 4.8 GB smaller than at 16-bit. Its measured peak falls by about the same, from 31 GB to 26 GB. The text encoder and the working memory stay the same, so the whole peak shrinks far less than the model file.
Offloading: less memory, more time
Offloading keeps parts of the model in the computer's main memory and moves them to the GPU only when needed. It lets a model run on a smaller card, but each run takes longer. Tencent suggests turning it off when the card has enough memory. Wan says the same for its TI2V-5B model on cards with 80 GB or more.
Wan's peaks above were measured with offloading on, so they already include that saving. Without it, the same settings need more memory.
Which GPU sizes fit: from 8 GB to 80 GB and more
This table turns the figures above into card sizes. Where a maker gives a peak, we pick the smallest common card above it. Where only file sizes exist, we keep a tenth of the card free for working memory, as in our LLM memory guide. Those entries are marked as our reading.
A 24 GB card runs most of the image models and Wan 2.2's 5-billion model; 720p video from the larger models needs 80 GB.
| GPU memory | What fits |
|---|---|
| 8 to 12 GB | SDXL (Stability AI's figure); FLUX.2 [dev] with group offloading, if the computer has about 32 GB of main memory (Hugging Face's guide) |
| 16 GB | FLUX.2 [klein] 4B (about 13 GB); HunyuanVideo 1.5 with offloading (at least 14 GB) |
| 24 GB, such as the L4 and RTX 4090 | Wan 2.2 TI2V-5B at 720p (22.9 GB peak); FLUX.2 [dev] at 4-bit (about 20 GB free); FLUX.1 with its 8-bit model and the text encoders offloaded (our reading) |
| 32 GB, such as the RTX 5090 | flux-3-image-klein at 16-bit (31 GB peak) or 8-bit (26 GB) |
| 48 GB, such as the L40S | Wan 2.2 T2V-A14B at 480p (41.3 GB peak); HunyuanVideo at 544 x 960 (45 GB); FLUX.1 and FLUX.2 [klein] 9B at 16-bit, and LTX-2.5 with its 8-bit ComfyUI files (our reading) |
| 80 GB, such as the H100 and A100 80 GB | Wan 2.2 T2V-A14B and I2V-A14B at 720p (59.8 GB peak); HunyuanVideo at 720 x 1,280 (60 GB); LTX-2.5 at 16-bit (70.1 GB of files, our reading) |
| More than 80 GB | FLUX.2 [dev] at 16-bit without offloading (Hugging Face's guide) |
A card at the edge of its memory works for the tested setting, but little else. Wan's 22.9 GB peak on a 24 GB card leaves almost no room for a higher resolution, a longer clip or extra workflow nodes.
GPUs you can rent, by memory size
The boxes below list GPUs from our price records, with 24 GB, 32 GB, 48 GB or 80 GB of memory. By the hour suits image and video work that comes in bursts; by the month suits a card that works every day.
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| UpCloud GPU Server, 1 x NVIDIA L4, 8 cores, 64 GB 24 GB GPU | €0.58/hour (about €390/month; no charge after 28 days of use)30-day month: the last 2 days are free.31-day month: the last 3 days are free.February: no free days (one in a leap year).Started mid-month: it may never reach the cap that month. on demand, billed by the starting hour; monthly cost capped at 28 days of use checked 25 Oct 2026 | No offer | Visit |
| UpCloud GPU Server, 1 x NVIDIA L40S, 8 cores, 64 GB 48 GB GPU | €1.11/hour (about €746/month; no charge after 28 days of use)30-day month: the last 2 days are free.31-day month: the last 3 days are free.February: no free days (one in a leap year).Started mid-month: it may never reach the cap that month. on demand, billed by the starting hour; monthly cost capped at 28 days of use checked 25 Oct 2026 | No offer | Visit |
| DigitalOcean GPU Droplet, 1 x NVIDIA L40S 48 GB GPU | $1.57/hour (about $1,146/month) on demand, billed per second · VAT not stated checked 25 Oct 2026 | New accounts get a $5 credit that lasts 90 days and covers every product except SaaS add-ons checked 29 Oct 2026 | Visit |
| UpCloud GPU Server, 1 x NVIDIA H100, 12 cores, 240 GB 80 GB GPU | €2.17/hour (about €1,458/month; no charge after 28 days of use)30-day month: the last 2 days are free.31-day month: the last 3 days are free.February: no free days (one in a leap year).Started mid-month: it may never reach the cap that month. on demand, billed by the starting hour; monthly cost capped at 28 days of use checked 25 Oct 2026 | No offer | Visit |
| DigitalOcean GPU Droplet, 1 x NVIDIA H100 80 GB GPU | $4.41/hour (about $3,219/month) on demand, billed per second · VAT not stated checked 25 Oct 2026 | New accounts get a $5 credit that lasts 90 days and covers every product except SaaS add-ons checked 29 Oct 2026 | Visit |
Monthly costs for hourly prices assume the server runs all month (730 hours), or less where a provider caps the monthly bill.
UpCloud prices are in euros and DigitalOcean prices in US dollars; we do not convert them.
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| Unihost AZM-04, NVIDIA RTX 4090 24 GB GPU | €359.90/month monthly, paid in advance; minimum one month checked 25 Oct 2026 | No offer | Visit |
| GPU Mart Enterprise Dedicated GPU Server - RTX 4090 24 GB GPU | $409/month monthly rate of a 24-month term, paid in advance · VAT not stated checked 25 Oct 2026 | A free 24-hour trial of a VPS, dedicated or GPU server for new clients, on request checked 29 Oct 2026 | Visit |
| Liquid Web GPU server, NVIDIA L4 24GB 24 GB GPU | $584.25/month monthly; 25% off the first month, then $779 · VAT not stated checked 25 Oct 2026 | No offer | Visit |
| GPU Mart Enterprise Dedicated GPU Server - RTX 5090 32 GB GPU | $479/month monthly rate of a 24-month term, paid in advance · VAT not stated checked 25 Oct 2026 | A free 24-hour trial of a VPS, dedicated or GPU server for new clients, on request checked 29 Oct 2026 | Visit |
| Liquid Web GPU server, NVIDIA L40S 48GB 48 GB GPU | $1,049.25/month monthly; 25% off the first month, then $1,399 · VAT not stated checked 25 Oct 2026 | No offer | Visit |
| GPU Mart Enterprise Dedicated GPU Server - A100(80GB) 80 GB GPU | $1,559/month monthly rate of a 24-month term, paid in advance · VAT not stated checked 25 Oct 2026 | A free 24-hour trial of a VPS, dedicated or GPU server for new clients, on request checked 29 Oct 2026 | Visit |
Read the term under each price: some need a 24-month term, and Liquid Web's first month is discounted.
If you would rather not choose a card at all, Comfy Cloud runs ComfyUI for you. Its page says it runs on Blackwell RTX 6000 Pro cards with 96 GB of memory, more than any card in the boxes above. Our guide to running ComfyUI in the cloud compares it with renting a GPU.
Mistakes to avoid
Counting only the image or video model's file
Instead: Add the text encoder and the VAE: FLUX.2 [dev]'s text encoder alone is 48.0 GB.
Reading a 32-bit file size as the memory you need
Instead: Halve 32-bit files, such as Wan 2.2's, when the tool loads them at 16-bit.
Comparing a peak measured with offloading to one without it
Instead: Check the setting behind each figure: Wan's peaks were measured with offloading on.
Expecting 4-bit files to run on every card
Instead: Check the format: Lightricks marks its 4-bit files for NVIDIA's Blackwell cards.
Renting a card that only just fits
Instead: Leave room for higher resolutions, longer clips and extra workflow nodes.
Other questions
Can I run Flux with 8 GB or 12 GB of VRAM?
FLUX.2 [dev] can, with group offloading, if the computer also has about 32 GB of main memory. Hugging Face's guide says so, and offloading makes each image slower. FLUX.2 [klein] 4B needs about 13 GB, so 16 GB is the comfortable size for it.
How much VRAM does Wan 2.2 need?
The 5-billion-parameter TI2V-5B peaked at 22.9 GB for a 720p clip in Wan's test on an RTX 4090. The two 14-billion models peaked at 41.3 GB at 480p and 59.8 GB at 720p on one H100. Wan says a single GPU needs at least 80 GB for them.
What GPU do I need for ComfyUI?
ComfyUI loads the models in your workflow, so the models decide. Find the largest model you plan to use in the tables above and pick the card size from the GPU table.
Is 24 GB enough for video?
For Wan 2.2 TI2V-5B at 720p and HunyuanVideo 1.5 with offloading, yes, by their makers' figures. The 14-billion Wan models, LTX-2.5 at 16-bit and HunyuanVideo at 720p need much more.
Download the table
The figures in the tables above are also available as one data file, without prices, for your own sums:
This table lists 12 open image and video models with their parameters, file sizes and their makers' GPU memory figures. AI Hosting Index read the files and figures on the makers' own pages on 26 October 2026.
| Model | Type | Maker | Parameters (billions) | Model files as released (GB) | Precision of those files | Text encoder files (GB) | Maker's memory figure (GB) | What that figure is | Card that fits, our reading | Hugging Face repository |
|---|---|---|---|---|---|---|---|---|---|---|
| SDXL base 1.0 | image | Stability AI | 3.5 | 6.94 | 16-bit | included in one file | 8 | stated as enough for consumer GPUs | 8 GB | stabilityai/stable-diffusion-xl-base-1.0 |
| FLUX.1 [dev] | image | Black Forest Labs | 11.9 | 23.80 | 16-bit | 9.77 | none published | 48 GB at 16-bit | black-forest-labs/FLUX.1-dev | |
| FLUX.1 [schnell] | image | Black Forest Labs | 11.9 | 23.78 | 16-bit | 9.77 | none published | 48 GB at 16-bit | black-forest-labs/FLUX.1-schnell | |
| FLUX.2 [dev] | image | Black Forest Labs | 32.2 | 64.45 | 16-bit | 48.02 | 20 | free memory needed at 4-bit (Hugging Face's guide); more than 80 without offloading | 24 GB at 4-bit | black-forest-labs/FLUX.2-dev |
| FLUX.2 [klein] 4B | image | Black Forest Labs | 3.9 | 7.75 | 16-bit | 8.04 | 13 | stated | 16 GB | black-forest-labs/FLUX.2-klein-4B |
| FLUX.2 [klein] 9B | image | Black Forest Labs | 9.1 | 18.16 | 16-bit | 16.38 | none on the pages read | 48 GB at 16-bit | black-forest-labs/FLUX.2-klein-9B | |
| flux-3-image-klein | image | Black Forest Labs | 5.6 | 11.10 | 16-bit | 8.88 | 31 | peak, 1024 x 1024, 16-bit, measured on a B200 | 32 GB | black-forest-labs/flux-3-image-klein |
| Wan 2.2 TI2V-5B | video | Wan | 5 | 20.00 | 32-bit | 11.36 | 22.9 | peak, 720p clip, RTX 4090, offloading on | 24 GB | Wan-AI/Wan2.2-TI2V-5B |
| Wan 2.2 T2V-A14B | video | Wan | 27 (two experts) | 114.30 | 32-bit | 11.36 | 59.8 | peak, 720p clip, one H100, offloading on (41.3 at 480p) | 80 GB | Wan-AI/Wan2.2-T2V-A14B |
| LTX-2.5 | video with sound | Lightricks | 22 | 42.02 | 16-bit | 26.26 | none published | 80 GB at 16-bit | Lightricks/LTX-2.5 | |
| HunyuanVideo | video | Tencent | 13 | 25.64 | 16-bit | separate repository | 60 | peak, 720 x 1280, 129 frames (45 at 544 x 960) | 80 GB | tencent/HunyuanVideo |
| HunyuanVideo 1.5 | video | Tencent | 8.3 | 33.31 | 32-bit | separate repository | 14 | minimum with model offloading | 16 GB | tencent/HunyuanVideo-1.5 |
Download this table as CSV (2 KB)
Licence: all rights reserved.
How to cite this table: AI Hosting Index, GPU memory of 12 image and video models, https://aihostingindex.com/guides/image-video-model-vram/, last updated 26 October 2026.
Sources: each model's Hugging Face repository and its maker's page or README, read 26 Oct 2026. The method is under "How we worked out these figures" below.
How we worked out these figures
File sizes come from the file lists of each model's official Hugging Face repository, read on 26 October 2026. We count 1 GB as a billion bytes. Where a maker publishes 32-bit files, we halve them for the 16-bit column. Memory figures come from each maker's own page or README, with the setting it states.
Prices in the boxes come from our weekly price records, each with the date we checked it; how we check prices explains the method. We plan to read these pages again every three months, and sooner when a maker releases a new version.
Sources and test notes
- Stability AI: Announcing SDXL 1.0, read
- SDXL base 1.0 on Hugging Face, read
- FLUX.1 [dev] on Hugging Face, read
- FLUX.1 [schnell] on Hugging Face, read
- FLUX.2 [dev] on Hugging Face, read
- Hugging Face blog: FLUX.2 in Diffusers (25 Nov 2025), read
- FLUX.2 [klein] 4B on Hugging Face, read
- FLUX.2 [klein] 9B on Hugging Face, read
- flux-3-image-klein on Hugging Face, read
- Black Forest Labs models on Hugging Face (FP8 and NVFP4 versions), read
- Wan 2.2 README on GitHub (with its computational efficiency table), read
- Wan 2.2 TI2V-5B on Hugging Face, read
- Wan 2.2 T2V-A14B on Hugging Face, read
- Wan 2.2 I2V-A14B on Hugging Face, read
- LTX-2.5 on Hugging Face, read
- HunyuanVideo README on GitHub, read
- HunyuanVideo on Hugging Face, read
- HunyuanVideo 1.5 README on GitHub, read
- HunyuanVideo 1.5 on Hugging Face, read
- Comfy Cloud, read
We ran no test for this page. Every memory figure is either the maker's own, with its stated setting, or our sum of the files in the maker's Hugging Face repository. We read both on 26 October 2026. Wan's peak memory figures come from the computational efficiency table in its README, which is published as an image.
Prices in the boxes come from our weekly price records, each with its own date.