Short answer: a 6GB graphics card can run useful AI chat and modest still pictures. It is a poor fit for the better picture models and for real AI video. I know because I tried it on my own Windows PC.
I am Blake. I have fixed computers for years. I am building a local AI named Kellie on my desktop (BAMPC11), which has a 6GB RTX 2060 and 32GB of system RAM. From Nov 3 to Nov 8, 2026, I used Kellie's real logs, not lab scores, to see what this card can and cannot do. This page is that report.
I do not list prices. They change. As an Amazon Associate I earn from qualifying purchases. If you want a second opinion on what your own PC can run, use the free Check My PC tool.
What I tried
Kellie uses ComfyUI to make pictures, Ollama for text chat (I was using qwen3:8b), and a voice tool for spoken lines. I wanted still pictures that looked right, then real AI video that stays consistent from frame to frame.
I started with SDXL, the better picture model. When that failed, I fell back to the older, smaller SD1.5 model (Realistic Vision V5.1) in fast mode. I also let pictures and chat share the same card, the way a normal desk PC would.
What broke (from the logs)
Here is what the logs showed that week. These are not guesses. They are the lines Kellie wrote down.
- SDXL came out corrupted on the 6GB card. I switched to SD1.5 (Realistic Vision V5.1) in fast mode, which works.
- ComfyUI image jobs hit the 540-second (9 minute) limit twice on Nov 5. Once the job was still on the face and hand detail pass. Once it was still on the base pass. It never finished those two jobs.
- Ollama timed out the same day while pictures were running. I had asked
qwen3:8bfor a JSON reply. The log said:
Ollama took too long
The picture self-checks failed again and again. Small models kept drawing the wrong thing, even after several tries:
scene 2 FAILED after 4 tries: unrequested person
scene 1 FAILED after 5 tries: sunshine is not the main subject
laptop is the main subject
car is
That third line came up when I asked for "high-performance." The fourth came up when I asked for "high-speed." The model grabbed a thing it already knew (a laptop, a car) instead of the idea I meant.
Two other problems showed up in the same week, and they are worth knowing about even though they are not about picture quality:
- Voice lines clipped at full scale. I had to scale the audio down so the spoken lines did not distort.
- On Nov 8 the desktop app crashed with a Windows fatal exception while a file dialog was open and the monitor was asleep. I cannot prove those two things caused the crash. I just no longer leave that window open when I walk away.
The result: I moved video making off this local card to a cloud video helper, until I get a 16GB card.
Why this happens (think of VRAM as a workbench)
The memory on a graphics card is called VRAM. Think of it as a workbench. The AI model has to sit on that bench while it works, along with the picture it is building. If the model does not fit, the computer shoves pieces back and forth into regular system RAM. That path is much slower. It can also leave you with a corrupted picture, or a job that is still running when the clock runs out.
A few rules of thumb, stated carefully, because I did not measure these as lab benchmarks:
- SDXL, the better picture model, comfortably wants about 8GB or more.
- Real AI video models that keep frames consistent typically want 12 to 24GB, because they have to keep many frames on the bench at once.
- 16GB is the practical sweet spot for a single card. A card like the RTX 5060 Ti 16GB sits in that range. I have a full guide on that card planned for Nov 13. Until then, see the best budget GPU for Ollama.
My 6GB RTX 2060 is under those marks. That is why SDXL broke, why ComfyUI hit the 9 minute wall, and why video had to leave this machine.
What still works on 6GB
The card is not useless. It is just limited. This is what still works for me:
- SD1.5 still pictures in fast mode (Realistic Vision V5.1). The pictures are weaker than SDXL, and prompts need to be concrete, but the model actually finishes.
- Text chat with 7B to 8B models in Ollama. That is everyday writing, summaries and questions on your own PC. For the memory math, see how much RAM and VRAM Ollama needs and the best Ollama models for 16GB RAM.
- Voice, if you scale the audio down so it does not clip at full volume.
One more limit: run one heavy job at a time. When I ran pictures and Ollama together, the chat model timed out. If you have even less to work with, like an old PC with 8GB of RAM and no big graphics card, start with a small text model. The LFM3 in Ollama guide covers that setup.
Tips if you have a 6GB card
- Use SD1.5, not SDXL, for still pictures on 6GB. On my card, SDXL came out corrupted.
- Do one heavy job at a time. Unload the chat model before you start pictures, then bring chat back when the pictures are done.
- Ask for concrete objects, not abstract words. "A silver laptop on a wooden desk" works better than "high-performance." "A red sports car on a wet road" works better than "high-speed." "Bright sunlight on a kitchen table" works better than "sunshine" as a vague idea. That is the lesson from the failed self-checks.
- Scale voice audio down so spoken lines do not clip at full volume.
- Do not leave a file dialog open if the monitor might go to sleep. That is the cautious habit I took from the Nov 8 crash, not a proven fix.
- Do not expect real AI video with consistent frames on 6GB. I tried, then I moved that work off the card.
When to upgrade
Upgrade when you want the better picture models, or video that stays consistent from the first frame to the last. 16GB is the practical sweet spot. I am looking at an RTX 5060 Ti 16GB for that reason.
A 12GB card, like the RTX 3060, is a cheaper step up from 6GB and is a solid card for 8B to 14B chat. It sits at the low end of what video models typically want, so I would not count on it for the same video work I am saving the 16GB card for. A used 24GB card has more room, but it needs a strong power supply (750W or more is the usual advice). I have not run these same tests on a 16GB or 24GB card yet. That is the point of the upgrade.
For buying help, see the best budget GPU for Ollama and the budget AI PC under $1,000 guide. If you are still installing the basics, use the Windows Ollama install guide. More guides live on the Local AI hub.
I'm saving for a 16GB card to keep testing. If this helped, ko-fi.com/bamtx.
FAQ
Can a 6GB graphics card run AI picture models?
Yes, with limits. On my 6GB RTX 2060, SDXL came out corrupted. I switched to SD1.5 (Realistic Vision V5.1) in fast mode, which works. The better picture models comfortably want about 8GB or more.
Can a 6GB card run AI video?
Not well enough for me. Real AI video models that keep frames consistent typically want 12 to 24GB. I moved video making off my local card to a cloud helper until I get a 16GB card.
Why did ComfyUI time out on my PC?
My image jobs hit the 540-second (9 minute) limit twice on Nov 5, 2026: once still on the face and hand detail pass, and once still on the base pass. When a model does not fit in the card's memory, pieces swap into regular RAM, which is much slower.
Why do AI pictures come out with the wrong subject?
On my tests, the self-checks failed repeatedly. After 4 tries, one log said "scene 2 FAILED after 4 tries: unrequested person." After 5 tries, another said "scene 1 FAILED after 5 tries: sunshine is not the main subject." When I asked for high-performance, the laptop was the main subject. When I asked for high-speed, the car was. Small models follow abstract words poorly. Ask for concrete objects.
What still works on a 6GB card?
SD1.5 still pictures in fast mode, text chat with 7B to 8B models, and voice if you scale the audio down so it does not clip at full volume. Run one heavy job at a time. When I ran pictures and Ollama together, the chat model timed out.
What card should I upgrade to for AI pictures and video?
16GB is the practical sweet spot, for example an RTX 5060 Ti 16GB. A 12GB RTX 3060 is a cheaper step up from 6GB. I have not run these same tests on a 16GB card yet. That is why I am saving for one.