← sources

20 steps, 720p, RTX 3090: A narrative video experiment with Wan2GP

source: reddit (reddit:StableDiffusion)published 2026-10-14 02:58ingested 2026-10-15 02:40

retained text

I've been experimenting with Wan2GP to see if I could make something closer to an actual TV episode instead of the usual collection of unrelated AI clips.

The result is **The Prompting Zone**, a black-and-white sci-fi parody inspired by the original Twilight Zone. The first episode is called *The Last Human*, about an office worker who starts noticing something strange about his coworkers.

For generation I used **MiniMax H3 Ref2VA Pruned 20B** through **Wan2GP**, at **720p, 20 steps**, running on a single **RTX 3090 24GB**. Each clip took approximately **30 minutes** to generate, so experimenting with different prompts wasn't exactly cheap in terms of time.

I started with a complete script, including all the dialogue, and divided it into scenes of around 5–10 seconds.

For each scene I generated a reference image containing four camera compositions arranged in a 2×2 grid. The idea was to give the video model a rough storyboard instead of relying entirely on text prompts.

I then fed those references into MiniMax with prompts describing the camera movements, character actions and exact dialogue.

It worked reasonably well, but there were plenty of issues. Sometimes the model interpreted the reference as an actual four-panel video instead of four sequential shots. Other times it invented movements, changed the background or decided a character should do something completely unrelated to the scene.

Character consistency was probably the biggest headache. Even with references, faces tend to drift between generations. I had to repeat descriptions of the characters, clothing and environment in nearly every prompt.

Computer screens were also problematic. The model struggled with readable text, so I eventually replaced most of the onscreen information with portraits and simple icons. Much more reliable than trying to generate legible UI elements.

The black-and-white style helped make everything look a little more consistent, although it obviously doesn't solve continuity problems.

Overall, I'm pretty happy with the result. There's still a lot of manual work involved, especially when a generation goes wrong after waiting half an hour.

I'd like to automate more of the process eventually, from script to storyboard, video prompts and final assembly.

Anyone else working on narrative videos with Wan2GP? Particularly interested in how you're handling character consistency and dialogue across multiple clips.

document fetched 2026-10-15 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1x5gcki/20_steps_720p_rtx_3090_a_narrative_video