# RTX 3060 12GB — Finally Got a Usable 25-Second Video at Native 1536×864 (ComfyUI + H3) — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/6012b99e9eb89399

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-10-28 11:47 ingested 2026-10-29 02:42

[original source ↗](https://www.reddit.com/r/comfyui/comments/1xhlut5/rtx_3060_12gb_finally_got_a_usable_25second_video/)

## retained text



```
After several rounds of testing, I finally managed to generate a usable 25-second AI wildlife video at native 1536×864 resolution on my RTX 3060 12GB.

In my previous experiments, I had already confirmed that my GPU could complete a 25-second generation at this resolution without running out of memory. However, the resulting videos had serious visual and motion issues and weren't really usable.

This time, I focused on getting an actual watchable result rather than just completing the generation.

**Hardware:**

* GPU: NVIDIA RTX 3060 12GB
* System RAM: 32GB

**Workflow and generation settings:**

* ComfyUI
* Model: MiniMax H3 Singularity V2
* DMAD 4-step
* LoRA strength: 1.0
* Resolution: 1536×864 (native generation, no upscaling)
* Frame rate: 24 FPS
* Duration: 25 seconds (600 frames)
* Final generation time: 2256.36 seconds (37 minutes, 36 seconds)

**Video concept:**

A short African wildlife documentary-style scene. Two wildlife observers spot a mother cheetah with her two cubs. The camera shifts toward the animals, the cubs play near their mother, and she eventually grooms both cubs.

I used an initial frame, character reference images, and a 3×3 storyboard to guide the sequence.

**What went wrong during testing:**

I ran three full generations while refining the prompts.

The earlier attempts had several problems:

* Inconsistent hand positions on one of the observers
* Unnatural swelling and twitching around the mother cheetah's abdomen
* Excessive ear twitching
* Camera continuity and lighting inconsistencies

One particularly interesting issue was the cheetah's breathing. Describing the abdominal movement too explicitly made the model exaggerate it into something completely unnatural.

Simplifying the instruction to natural, relaxed breathing produced a much better result.

By the third generation, the character consistency, animal movements, grooming actions, and camera transitions were finally good enough for a usable video.

**A note about the final result:**

The entire visual sequence was generated in a single pass at native resolution. There are no stitched clips, cuts added in editing, or upscaled frames.

I only replaced some unnatural audio effects in post-production.

For me, this is an important step forward. Previously, I knew the RTX 3060 12GB could technically finish this kind of generation. Now I've finally produced something that is actually watchable.

I'm sharing the result and settings for anyone experimenting with longer, higher-resolution video generation on older 12GB GPUs.

Feedback and comparisons with other setups are welcome!
```



document fetched 2026-10-29 00:43 · canonical https://reddit.com/r/comfyui/comments/1xhlut5/rtx_3060_12gb_finally_got_a_usable_25second_video
