What's Feasible for a Local Video Generation with a 24GB 4090 GPU?
retained text
I’m trying to Run Comfy UI locally. I’m new to Comfy & Gen-AI, and Reddit, but experienced in traditional 3D Animation. How much AI video (length & resolution) can I realistically expect to generate locally on a 24GB 4090 GPU? (my full specs are below). I’m also getting lots of GPU crashes: screens go black, but my PC fans, lights, and Spotify keep running. Is this normal? I still have a lot to learn, and I mostly work with existing workflow templates, but I keep hearing reports of people running 15 seconds of Minimax H3 1280x720 video with my hardware setup. I’m not even coming close. With MiniMax H3 the most I got was: ◦ 4 seconds of t2i video at 0.2 megapixels (608 x 342) ◦ 2 seconds of t2i video at 0.6 megapixels (1024 x 576) ◦ Anything resolution or length beyond that, or adding a reference image = GPU Black Screen Crash With Wan 2.2: (5B model) 5 seconds / 124 frames at 1280x704, but my GPU crashed after 7 or so tries. It seems like my GPU & PC crash way more than it should, and my workflow is really fragile. Adding anything more than one reference image becomes the proverbial straw that breaks the camel’s back. I don’t dare try LoRAs, reference videos, audio or other additional inputs. The crashes seemed like a hardware issue, but I just brought my PC back from the shop and they said the GPU and hardware checked out - even after stress testing it for several days. Turned out to be a bad monitor. Still, constantly pushing my GPU past its limits has me worried about hardware damage. Any advice on what I can realistically expect? or what improvements are realistically possible with my setup? Will becoming a ComfyUI node master improve things? Or should I just accept that I’ll have to generate via Comfy UI cloud? Thanks in advance. More details if you're interested… ========== Full PC specs Thermal Take Build Geoforce RTX 4090 24GB 64GB DDR5 Ram Intel Core i9-14900K PSU Toughpower GF3 1000W Gold - TT Premium Edition ========= WHAT I’VE TRIED SO FAR: PHASE 1 First efforts: default t2i image generation ComfyUI workflow — I believe flux.dev. Got several successful local video generations using the Wan 2.2 5B parameter model: ◦ Both t2v and i2v short videos, at their default 124 frames. • After a few crashes, and I figured out the right parameters, I got Wan 2.2 generating video from a source image at 640 & 1280p. • However, even once it was running again, I noticed Wan would crash after about 7 successful i2v generations. I only changed the start frame image and prompt; the workflow's output length and resolution were the same. • Estimated GPU crashes this phase: 5-7• Somewhere around here I installed the MSI afterburner utility to power limit my PC• I also added some code to the .bat file, via AI instruction, to ease the GPU burden. PHASE 2 Tried local AI video again, this time with MiniMax H3. Video generation barely worked: ◦ 4 seconds of video at 0.2 megapixels (608 x 342) — OK ◦ 2 seconds of video at 0.6 megapixels (1024 x 576) — OK ◦ Anything beyond that = CRASH ◦ Adding a reference image also crashed it. • Estimated GPU crashes this phase: about 10 PHASE 3 Frustrated with MiniMax, I turned to image generation. Models tested: ◦ Flux.2 \[Klein\] 4B: Image Edit — GPU crash when I added a reference image, so I abandoned this workflow and model. ◦ Flux.1 Dev fp8: Text to Image — success. ◦ Flux.2 \[Klein\] 4B Distilled — success. I was able to load reference images and generate around 300 1024x1024 images without any issues. The image quality wasn’t great, but this worked for my brainstorming/idea generation purposes. PHASE 4 • Tried Qwen Image 2.1: Text to Image and generated about 20 images successfully. • Encouraged, I tried the qwen\_image\_2\_1\_image\_edit workflow. It told me I needed a missing model, and I think I made the mistake of adding a 32GB model. Got another GPU crash, and brought it to the shop.
document fetched 2026-10-24 10:42 · canonical https://reddit.com/r/comfyui/comments/1xdkghu/whats_feasible_for_a_local_video_generation_with