Gear
7 best GPUs for local AI image and video in November 2026

Key takeaways
VRAM sets what you can run. FLUX 3 Image Klein fits 8GB in its INT8 build, Wan 2.2's 5B model asks for 24GB in its own repository, and LTX-2.5 asks for 32GB.
Prices sit far above launch MSRP. On November 17, Tom's Hardware's tracker listed the RTX 5060 Ti 16GB at $799 against a $429 launch price, and the RTX 5090 at $6,549 against $1,999.
The RTX 5060 Ti 16GB has the lowest price per gigabyte of VRAM of the NVIDIA cards here, at about $50 a GB.
A used RTX 3090 is the cheapest 24GB card at about $1,663 on November 19, but that average is up 31% in 90 days.
NVIDIA's comparison page lists no RTX 50 Super cards, so every card below is a current or older release.
Which GPU is best for local AI image and video?
For most people, an NVIDIA card with 16GB of VRAM is the best GPU for local AI image and video, and at November 2026 prices the RTX 5060 Ti 16GB is the cheapest way to get one. A used RTX 3090 is the step up for video and LoRA training, and the RTX 5090 is the only consumer card with 32GB.
The market is the problem. Tom's Hardware's price tracker, last updated November 17, says the RTX 5080 and 5070 Ti are in poor supply at any reasonable price, and RTX 40-series cards are no longer in production. Every price below carries its date, and each one can move within days. If you're on a Mac, the trade-offs are different, so read our Mac vs. PC guide first.
How these cards were ranked
The ranking leans on five things. VRAM per dollar comes first, since a model that doesn't fit runs slowly or not at all. Software support is second, because ComfyUI's docs lead with NVIDIA and CUDA. Speed comes only from published figures. The last two are power and PSU needs, and fit for image generation, video, or training. Our VRAM guide covers what each model asks for in detail.
The seven options at a glance
Card | VRAM | Price and date | Best for |
|---|---|---|---|
RTX 5060 Ti 16GB | 16GB | $799 (Nov 17) | Image generation on a budget |
RTX 3090, used | 24GB | About $1,663 (Nov 19) | Video and LoRA training |
RTX 5070 Ti | 16GB | $1,400 (Nov 17) | Faster image work |
RTX 5090 | 32GB | $6,549 (Nov 17) | Large video models |
RTX 4090, used | 24GB | About $3,436 (Nov 19) | 24GB with more speed |
RTX 5080 | 16GB | $2,235 (Nov 17) | Speed when price is secondary |
RX 9070 XT or Arc B580 | 16GB or 12GB | $959 or $329 (Nov 17) | Tinkerers avoiding NVIDIA |
What the published speed numbers show
The best public speed data for these cards comes from model makers. The Kandinsky 6.0 repository times a 5-second text-to-video clip with audio on three consumer cards. On the Lite Distill SD setting, the RTX 5090 took 44.2 seconds, the 4090 took 68.0, and the 5060 Ti 16GB took 173.7. On Pro Full HD the three took 862.8, 1,096.7, and 3,571.4 seconds. Black Forest Labs' Klein post adds about 4 seconds for a 1MP image on a 5090 and under 9 seconds on an RTX 4070. No source in this piece publishes a 3090 time, so its speed isn't ranked.
1. RTX 5060 Ti 16GB: best value for 16GB
NVIDIA's 16GB version of the 5060 Ti is the cheapest way into 16GB of VRAM with CUDA support.
Strengths: It has 16GB of GDDR7, draws 180W, and NVIDIA lists 600W of system power for it. At $799 it works out to about $50 per GB of VRAM. Black Forest Labs says FLUX 3 Image Klein's INT8 build runs on RTX cards with as little as 8GB, so 16GB leaves room for LoRAs and bigger workflows. The 8GB version costs $499 but has half the room.
Limits: Its 448GB/s of memory bandwidth is the lowest of the cards here. In the Kandinsky 6.0 speed table, it took 173.7 seconds for a 5-second Lite Distill clip against 68.0 on an RTX 4090, though it ran the text encoder in NF4. The $799 price is 86% above its $429 launch price.
Best for: Image generation, Klein-sized models, and trying video before spending more.
2. RTX 3090, used: best 24GB for the money
The 3090 is a 2020 card that still has the memory current video and training tools want.
Strengths: ResalePrices put the used average at $1,663 on November 19, across 310 listings with a fair range of $1,586 to $1,714. That's about $69 per GB. Wan's repository says its 5B model command runs on at least 24GB, and Black Forest Labs' help center puts FLUX.2 [klein] 9B training at around 22GB, naming the 3090 and 4090 as examples. It's rated at 350W.
Limits: The 3090 average is up 30.7% over 90 days. It's an older generation, and ComfyUI says its NVFP4 speedup needs an RTX 50-series or Blackwell Pro card. Buy from a seller with a return policy.
Best for: Video and LoRA training on one card, as covered in our LoRA training guide.
3. RTX 5070 Ti: best for faster 16GB work
The 5070 Ti has the same 16GB as the 5060 Ti and a much wider memory bus.
Strengths: Its 896GB/s is double the 5060 Ti's, and it draws 300W with 750W of system power listed. As a Blackwell card it supports NVFP4, and Comfy Org ran its January NVFP4 benchmarks on a 5070 Ti, reporting about a 2x speedup over fp8 or bf16 with a CUDA 13.0 build of PyTorch.
Limits: At $1,400 it costs $601 more than the 5060 Ti for the same VRAM, and the price is 87% above its $749 launch price. A model that doesn't fit in 16GB still has to offload.
Best for: Image work where speed counts and 16GB is enough.
4. RTX 5090: best for large video models
The 5090 is the only consumer card here with 32GB, and the only one that meets LTX-2.5's stated minimum.
Strengths: It has 32GB of GDDR7 on a 512-bit bus, about 1,792GB/s. Black Forest Labs lists about 4 seconds for a 1MP Klein image in INT8 through ComfyUI. In the Kandinsky table it finished the Lite Distill SD clip in 44.2 seconds and the Pro Full HD clip in 862.8, against 68.0 and 1,096.7 on a 4090. Our open-weight video roundup covers which models want the memory.
Limits: Tom's lists $6,549, 3.3 times the $1,999 MSRP. It's rated 575W and NVIDIA lists a 1,000W PSU.
Best for: LTX-class video, bigger training runs, and work where waiting costs money.
5. RTX 4090, used: best for 24GB with more speed
The 4090 is out of production, so every card is used or leftover stock.
Strengths: It has 24GB and finished Kandinsky's Pro Full HD clip in 1,096.7 seconds, against 3,571.4 on the 5060 Ti. The Wan and Klein 9B memory figures cited for the 3090 name the 4090 as well.
Limits: ResalePrices put the average at $3,436 on November 19, up 37.4% in 90 days, which is more than twice the 3090 for the same 24GB. It draws 450W, and NVIDIA lists 850W of system power. It lacks NVFP4.
Best for: People who want 24GB, can't wait on a 3090, and found one at a fair price.
6. RTX 5080: best for speed on a 16GB budget
The 5080 trades the 5090's memory for a lower price, though not by much in this market.
Strengths: Its 960GB/s bus is more than double the 5060 Ti's, and it supports NVFP4. It draws 360W with 850W of system power listed.
Limits: Tom's lists $2,235, 2.2 times the $999 MSRP and 2.8 times the 5060 Ti's price, for the same 16GB. That makes it the weakest value here.
Best for: Image work where speed matters and the budget is flexible.
7. RX 9070 XT and Arc B580: best outside NVIDIA
Both cards run ComfyUI, and both sit outside the CUDA path the tools are built around.
Strengths: The RX 9070 XT has 16GB of GDDR6 at 640GB/s, draws 304W, and AMD lists a 750W PSU. The Arc B580 has 12GB at 456GB/s, draws 190W, and Intel suggests 600W. At $329 it's the cheapest VRAM here at about $27 a GB. ComfyUI's docs list AMD on Windows through ROCm, including RDNA 4, and Intel through torch.xpu, and Comfy Org announced official Windows ROCm support in January.
Limits: The RX 9070 XT sells at $959 against a $599 MSRP. NVFP4 is NVIDIA-only, and ComfyUI's system requirements list NVIDIA with CUDA 13.0 first.
Best for: Tinkerers comfortable when a node or model assumes CUDA.
Which card fits which job
For image generation only, the 5060 Ti 16GB covers Klein, SDXL, and similar models at the lowest NVIDIA cost per GB.
For video, a 24GB card such as the used 3090 meets Wan's 5B command, and the 5090 meets LTX-2.5's 32GB.
For LoRA training, 22GB covers FLUX.2 [klein] 9B per Black Forest Labs, while klein 4B trains on 12GB.
For the lowest spend, the Arc B580 at $329 has the cheapest VRAM, with the most software friction.
Before you buy
Check the power supply first. The 5090 needs a 1,000W unit by NVIDIA's guidance and the 5080 or 4090 need 850W. Local models also fill drives fast, so budget for storage using our storage guide. New to the software? Start with ComfyUI on Windows. For FLUX 3 specifically, the FLUX 3 explainer covers the lineup.
Local AI GPU FAQ
What is the best GPU for local AI image generation?
The RTX 5060 Ti 16GB is the lowest-cost NVIDIA card with 16GB, and Klein runs on as little as 8GB in INT8. If you can afford it, a used 3090's 24GB gives more room for video and training.
How much VRAM do I need for AI video?
It depends on the model. ComfyUI says Wan 2.2's 5B model should fit 8GB with offloading, Wan's own command asks for 24GB, and LTX's docs list a 32GB minimum for LTX-2.5.
Is a used RTX 3090 still worth it?
For 24GB, it's the cheapest route, at about $1,663 on November 19. The average has risen 30.7% in 90 days, so check sold listings first.
Are RTX 50 Super cards coming?
NVIDIA's comparison page lists none as of November 19, 2026, so any Super refresh is still unannounced there.
Can I use an AMD or Intel GPU?
Yes. ComfyUI's docs list AMD on Windows through ROCm and Intel Arc through torch.xpu, though NVFP4 stays NVIDIA-only.
