Battlezone-inspired AI short on a 12GB RTX 3080 Ti: full H3 + Pika Flex workflow shared
Successful_Fig8907 · reddit · 2026-11-04
A Reddit user made a Battlezone II–inspired sci-fi short on a 12GB RTX 3080 Ti (i9-11900K, 48GB RAM) and shared the full workflow while asking for advice.
- Workflow: ComfyUI + MiniMax H3 Extender Ref2VA with the H3 pruned INT8 ConvRot model and Pika H3 Flex Ref2VA adapter at strength 1.0; 6-step LCM/simple sampling, Pika sigmas, Comfy Kitchen attention. VRAM savings via the H3MemoryOptimization node: 4096-row chunks, native precision, automatic QKV/MLP/embedding memory modes, Kitchen V set to "Lower VRAM (slower)".
- Generation: 1MP (1344×736), 24fps, independent 5–8s clips; motion context between jobs disabled — continuity relies on references and written camera/action/opening/closing states. Dialogue uses H3 native audio with text-directed voices.
- Timing: 2h 32m total processing, 4:07 per clip on average; 3h 01m full batch window including gaps and discarded takes.
- References: 1–2 images per clip, 11 distinct images total; from clip 19 onward background photos were dropped in favor of text-described environments.
Open questions: physical coherence/object interactions, consistent positions and backgrounds across shots, and upscaling within 12GB VRAM — raising generation MP causes OOM, and VAE×2 hasn't delivered crisp HD detail. Specific VAE/decode or separate video-upscaler settings requested.
More from Multimodal
- Meituan launches LongCat-2.5: 1.6T-param native multimodal agentic model — thisdudelikesAI · 2026-11-04
- Redditor crafts superhero teaser 'The Stone Bearers' with Runway, Kling and Suno — Ulkaverse · 2026-11-04
- Why can't video models re-render the same take at higher resolution? An LTX 2.5 complaint — iJohnnyRobert · 2026-11-04
- AI filmmaking's hardest problem: Reddit practitioner seeks cutting edge for multi-shot character and spatial consistency — Due-Air-5626 · 2026-11-04
- AI filmmaking's hardest problem: Reddit practitioner seeks cutting edge for multi-shot character and spatial consistency — Due-Air-5626 · 2026-11-04
- 'Put this giant man on this tiny bicycle': AI video now out-entertains Pixar — HeyZoyaKhan · 2026-11-04